{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.12.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[],"dockerImageVersionId":28755,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# ============================================================\n# RSNA KNEE ABNORMALITY DETECTION\n# PHASE 1: DATASET INTELLIGENCE / DATA AUDIT\n# ============================================================\n\nimport os\nimport re\nimport gc\nimport json\nimport time\nimport math\nimport random\nimport hashlib\nimport warnings\nfrom pathlib import Path\nfrom collections import Counter, defaultdict\n\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\nwarnings.filterwarnings(\"ignore\")\n\nSEED = 42\nrandom.seed(SEED)\nnp.random.seed(SEED)\n\n# ------------------------------------------------------------\n# Audit configuration\n# ------------------------------------------------------------\n\nN_STUDY_AUDIT = 60\nN_DICOM_AUDIT = 80\nN_VISUAL_SERIES = 12\nN_VISUAL_SLICES = 9\n\n# Do not increase these blindly.\n# Phase 1 is deliberately sampling-based.\n\nOUTPUT_DIR = Path(\"/kaggle/working/rsna_knee_audit\")\nVISUAL_DIR = OUTPUT_DIR / \"visual_audit\"\n\nOUTPUT_DIR.mkdir(parents=True, exist_ok=True)\nVISUAL_DIR.mkdir(parents=True, exist_ok=True)\n\nprint(\"Output directory:\", OUTPUT_DIR)\nprint(\"Seed:\", SEED)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:37:50.700344Z","iopub.execute_input":"2026-08-11T07:37:50.701682Z","iopub.status.idle":"2026-08-11T07:37:50.712182Z","shell.execute_reply.started":"2026-08-11T07:37:50.701627Z","shell.execute_reply":"2026-08-11T07:37:50.710964Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# KAGGLE ENVIRONMENT CHECK\n# ============================================================\n\nprint(\"Current working directory:\")\nprint(Path.cwd())\n\nprint(\"\\nKaggle input exists:\", Path(\"/kaggle/input\").exists())\nprint(\"Kaggle working exists:\", Path(\"/kaggle/working\").exists())\n\nif Path(\"/kaggle/input\").exists():\n input_dirs = [p for p in Path(\"/kaggle/input\").iterdir() if p.is_dir()]\n \n print(\"\\nDatasets mounted under /kaggle/input:\")\n for p in input_dirs:\n print(\" -\", p)\n\nprint(\"\\nPython version:\")\nimport sys\nprint(sys.version)\n\nprint(\"\\nNumPy:\", np.__version__)\nprint(\"Pandas:\", pd.__version__)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:37:50.714097Z","iopub.execute_input":"2026-08-11T07:37:50.714543Z","iopub.status.idle":"2026-08-11T07:37:50.743642Z","shell.execute_reply.started":"2026-08-11T07:37:50.714512Z","shell.execute_reply":"2026-08-11T07:37:50.742443Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# DATASET PATH DISCOVERY\n# ============================================================\n\nKAGGLE_INPUT = Path(\"/kaggle/input\")\n\nEXPECTED_FILES = [\n \"train.csv\",\n \"train_series.csv\",\n \"test.csv\",\n \"test_series.csv\",\n \"sample_submission.csv\",\n]\n\ndef find_file(filename):\n matches = list(KAGGLE_INPUT.rglob(filename))\n \n if not matches:\n return None\n \n # Prefer paths where several expected competition files\n # are located nearby.\n scored = []\n \n for path in matches:\n parent = path.parent\n \n score = 0\n \n for expected in EXPECTED_FILES:\n if (parent / expected).exists():\n score += 1\n \n if (parent / \"train_series\").exists():\n score += 3\n \n if (parent / \"test_series\").exists():\n score += 3\n \n scored.append((score, path))\n \n scored.sort(key=lambda x: x[0], reverse=True)\n return scored[0][1]\n\n\nDATA_PATHS = {}\n\nfor filename in EXPECTED_FILES:\n path = find_file(filename)\n DATA_PATHS[filename] = path\n\nprint(\"Discovered competition files:\\n\")\n\nfor name, path in DATA_PATHS.items():\n print(f\"{name:25s} -> {path}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:37:50.778997Z","iopub.execute_input":"2026-08-11T07:37:50.779583Z","iopub.status.idle":"2026-08-11T07:51:06.60133Z","shell.execute_reply.started":"2026-08-11T07:37:50.779543Z","shell.execute_reply":"2026-08-11T07:51:06.6Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# DISCOVER TRAIN/TEST SERIES DIRECTORIES\n# ============================================================\n\ndef find_directory(dirname):\n matches = [p for p in KAGGLE_INPUT.rglob(dirname) if p.is_dir()]\n \n if not matches:\n return None\n \n # Prefer the directory closest to the discovered CSV files.\n train_csv = DATA_PATHS.get(\"train.csv\")\n \n if train_csv:\n train_parent = train_csv.parent\n \n same_parent = [\n p for p in matches\n if p.parent == train_parent\n ]\n \n if same_parent:\n return same_parent[0]\n \n return matches[0]\n\n\nTRAIN_SERIES_DIR = find_directory(\"train_series\")\nTEST_SERIES_DIR = find_directory(\"test_series\")\n\nprint(\"TRAIN_SERIES_DIR:\")\nprint(TRAIN_SERIES_DIR)\n\nprint(\"\\nTEST_SERIES_DIR:\")\nprint(TEST_SERIES_DIR)\n\nassert TRAIN_SERIES_DIR is not None, \"train_series directory was not found.\"\nassert TEST_SERIES_DIR is not None, \"test_series directory was not found.\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:51:06.604281Z","iopub.execute_input":"2026-08-11T07:51:06.604714Z","iopub.status.idle":"2026-08-11T07:54:32.79123Z","shell.execute_reply.started":"2026-08-11T07:51:06.604682Z","shell.execute_reply":"2026-08-11T07:54:32.790122Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# LOAD CSV FILES\n# ============================================================\n\ndef load_csv(name):\n path = DATA_PATHS[name]\n \n if path is None:\n raise FileNotFoundError(f\"{name} was not found.\")\n \n print(f\"Loading {name} from:\")\n print(path)\n \n df = pd.read_csv(path)\n \n print(\"Shape:\", df.shape)\n \n return df\n\n\ntrain = load_csv(\"train.csv\")\ntrain_series = load_csv(\"train_series.csv\")\ntest = load_csv(\"test.csv\")\ntest_series = load_csv(\"test_series.csv\")\nsample_submission = load_csv(\"sample_submission.csv\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:32.792394Z","iopub.execute_input":"2026-08-11T07:54:32.792675Z","iopub.status.idle":"2026-08-11T07:54:33.028552Z","shell.execute_reply.started":"2026-08-11T07:54:32.792645Z","shell.execute_reply":"2026-08-11T07:54:33.027258Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# BASIC DATASET STRUCTURE\n# ============================================================\n\nprint(\"TRAIN\")\nprint(train.head())\nprint(\"\\nColumns:\")\nprint(train.columns.tolist())\n\nprint(\"\\nTRAIN SERIES\")\nprint(train_series.head())\nprint(\"\\nColumns:\")\nprint(train_series.columns.tolist())\n\nprint(\"\\nTEST\")\nprint(test.head())\nprint(\"\\nColumns:\")\nprint(test.columns.tolist())\n\nprint(\"\\nTEST SERIES\")\nprint(test_series.head())\n\nprint(\"\\nSAMPLE SUBMISSION\")\nprint(sample_submission.head())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:33.032069Z","iopub.execute_input":"2026-08-11T07:54:33.032983Z","iopub.status.idle":"2026-08-11T07:54:33.064258Z","shell.execute_reply.started":"2026-08-11T07:54:33.032939Z","shell.execute_reply":"2026-08-11T07:54:33.063062Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# TARGET DISCOVERY\n# ============================================================\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\",\n]\n\nmissing_targets = [t for t in TARGETS if t not in train.columns]\n\nif missing_targets:\n print(\"WARNING: Expected target columns missing:\")\n print(missing_targets)\n\nTARGETS_PRESENT = [t for t in TARGETS if t in train.columns]\n\nprint(\"\\nTargets found:\")\nfor i, target in enumerate(TARGETS_PRESENT, 1):\n print(f\"{i:2d}. {target}\")\n\nassert len(TARGETS_PRESENT) == 12, (\n f\"Expected 12 targets, found {len(TARGETS_PRESENT)}\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:33.06679Z","iopub.execute_input":"2026-08-11T07:54:33.067615Z","iopub.status.idle":"2026-08-11T07:54:33.076623Z","shell.execute_reply.started":"2026-08-11T07:54:33.06758Z","shell.execute_reply":"2026-08-11T07:54:33.075693Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# TRAIN CSV AUDIT\n# ============================================================\n\nsummary_rows = []\n\ndef add_summary(category, metric, value, notes=\"\"):\n summary_rows.append({\n \"category\": category,\n \"metric\": metric,\n \"value\": value,\n \"notes\": notes\n })\n\n\nadd_summary(\"train\", \"rows\", len(train))\nadd_summary(\"train\", \"columns\", len(train.columns))\nadd_summary(\n \"train\",\n \"unique_studies\",\n train[\"StudyInstanceUID\"].nunique()\n)\nadd_summary(\n \"train\",\n \"duplicate_study_rows\",\n int(train[\"StudyInstanceUID\"].duplicated().sum())\n)\n\nif \"PatientSex\" in train.columns:\n add_summary(\n \"train\",\n \"missing_patient_sex\",\n int(train[\"PatientSex\"].isna().sum())\n )\n\nif \"Report\" in train.columns:\n add_summary(\n \"train\",\n \"reports_present\",\n int(train[\"Report\"].notna().sum())\n )\n \n add_summary(\n \"train\",\n \"reports_missing\",\n int(train[\"Report\"].isna().sum())\n )\n\n\nprint(\"Train rows:\", len(train))\nprint(\"Unique studies:\", train[\"StudyInstanceUID\"].nunique())\nprint(\"Duplicate StudyInstanceUID rows:\",\n train[\"StudyInstanceUID\"].duplicated().sum())\n\nprint(\"\\nMissing values:\")\nprint(train.isna().sum().sort_values(ascending=False))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:33.078006Z","iopub.execute_input":"2026-08-11T07:54:33.078389Z","iopub.status.idle":"2026-08-11T07:54:33.112273Z","shell.execute_reply.started":"2026-08-11T07:54:33.078362Z","shell.execute_reply":"2026-08-11T07:54:33.111265Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# LABEL DISTRIBUTION\n# ============================================================\n\nlabel_rows = []\n\nfor target in TARGETS_PRESENT:\n \n series = pd.to_numeric(train[target], errors=\"coerce\")\n \n positive = int((series == 1).sum())\n negative = int((series == 0).sum())\n missing = int(series.isna().sum())\n total_available = positive + negative\n \n prevalence = (\n positive / total_available\n if total_available > 0\n else np.nan\n )\n \n ratio = (\n negative / positive\n if positive > 0\n else np.inf\n )\n \n label_rows.append({\n \"target\": target,\n \"positive\": positive,\n \"negative\": negative,\n \"missing\": missing,\n \"available_labels\": total_available,\n \"prevalence\": prevalence,\n \"negative_positive_ratio\": ratio\n })\n \n add_summary(\n \"labels\",\n f\"{target}_positive\",\n positive\n )\n \n add_summary(\n \"labels\",\n f\"{target}_missing\",\n missing\n )\n\n\nlabel_distribution = pd.DataFrame(label_rows)\n\ndisplay(label_distribution)\n\nlabel_distribution.to_csv(\n OUTPUT_DIR / \"label_distribution.csv\",\n index=False\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:33.113629Z","iopub.execute_input":"2026-08-11T07:54:33.114104Z","iopub.status.idle":"2026-08-11T07:54:33.15051Z","shell.execute_reply.started":"2026-08-11T07:54:33.11407Z","shell.execute_reply":"2026-08-11T07:54:33.149277Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 11: EXACT LABELED STUDY AUDIT\n# ============================================================\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\",\n]\n\nlabel_mask = train[TARGETS].notna().all(axis=1)\n\nlabeled_train = train.loc[label_mask].copy()\n\nprint(\"=\" * 70)\nprint(\"EXPLICITLY LABELED STUDY AUDIT\")\nprint(\"=\" * 70)\n\nprint(\"\\nTotal train studies:\", len(train))\nprint(\"Fully labeled studies:\", len(labeled_train))\n\nprint(\"\\nLabel values:\")\nfor target in TARGETS:\n print(\n f\"{target:20s} \"\n f\"unique={sorted(labeled_train[target].dropna().unique().tolist())}\"\n )\n\nprint(\"\\nLabeled study IDs:\")\nprint(labeled_train[\"StudyInstanceUID\"].head(10).tolist())\n\n# Save the small labeled dataset.\nlabeled_train.to_csv(\n OUTPUT_DIR / \"explicitly_labeled_studies.csv\",\n index=False\n)\n\nprint(\n \"\\nSaved:\",\n OUTPUT_DIR / \"explicitly_labeled_studies.csv\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:33.152098Z","iopub.execute_input":"2026-08-11T07:54:33.152519Z","iopub.status.idle":"2026-08-11T07:54:33.17933Z","shell.execute_reply.started":"2026-08-11T07:54:33.152482Z","shell.execute_reply":"2026-08-11T07:54:33.177556Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 12: LABELED TARGET RELATIONSHIPS\n# ============================================================\n\nlabeled_targets = labeled_train[TARGETS].astype(int)\n\nprint(\"=\" * 70)\nprint(\"TARGET CORRELATION\")\nprint(\"=\" * 70)\n\ncorr = labeled_targets.corr()\n\ndisplay(corr.round(3))\n\ncorr.to_csv(\n OUTPUT_DIR / \"labeled_target_correlations.csv\"\n)\n\n# ------------------------------------------------------------\n# Positive abnormality combinations\n# ------------------------------------------------------------\n\nfrom collections import Counter\n\ncombination_counter = Counter()\n\nfor _, row in labeled_targets.iterrows():\n \n positive_targets = tuple(\n target\n for target in TARGETS\n if row[target] == 1\n )\n \n combination_counter[positive_targets] += 1\n\n\ncombination_rows = []\n\nfor combination, count in combination_counter.most_common():\n \n combination_rows.append({\n \"combination\": \" | \".join(combination),\n \"count\": count\n })\n\n\ncombinations_df = pd.DataFrame(combination_rows)\n\nprint(\"\\nMost common abnormality combinations:\")\ndisplay(combinations_df.head(20))\n\ncombinations_df.to_csv(\n OUTPUT_DIR / \"label_combinations.csv\",\n index=False\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:33.182597Z","iopub.execute_input":"2026-08-11T07:54:33.18309Z","iopub.status.idle":"2026-08-11T07:54:33.245086Z","shell.execute_reply.started":"2026-08-11T07:54:33.183059Z","shell.execute_reply":"2026-08-11T07:54:33.243726Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 13: INSPECT LABELED REPORTS\n# ============================================================\n\nprint(\"=\" * 70)\nprint(\"SAMPLE OF LABELED REPORTS\")\nprint(\"=\" * 70)\n\ndisplay_columns = [\n \"StudyInstanceUID\",\n \"Report\"\n] + TARGETS\n\n# Show all 58 because the dataset is tiny.\nfor i, (_, row) in enumerate(\n labeled_train[display_columns].iterrows()\n):\n \n print(\"\\n\" + \"=\" * 100)\n print(f\"LABELED STUDY {i + 1} / {len(labeled_train)}\")\n print(\"=\" * 100)\n \n print(\"StudyInstanceUID:\")\n print(row[\"StudyInstanceUID\"])\n \n print(\"\\nLabels:\")\n \n positive_labels = [\n target\n for target in TARGETS\n if row[target] == 1\n ]\n \n negative_labels = [\n target\n for target in TARGETS\n if row[target] == 0\n ]\n \n print(\"Positive:\", positive_labels)\n print(\"Negative:\", negative_labels)\n \n print(\"\\nREPORT:\")\n print(row[\"Report\"])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:33.248382Z","iopub.execute_input":"2026-08-11T07:54:33.24875Z","iopub.status.idle":"2026-08-11T07:54:33.284271Z","shell.execute_reply.started":"2026-08-11T07:54:33.24872Z","shell.execute_reply":"2026-08-11T07:54:33.282969Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 14: REPORT SCRIPT / LANGUAGE AUDIT\n# ============================================================\n\ndef script_statistics(text):\n \n text = str(text)\n \n if not text.strip():\n return {\n \"ascii_letters\": 0,\n \"latin_extended\": 0,\n \"cyrillic\": 0,\n \"greek\": 0,\n \"other_letters\": 0,\n \"total_letters\": 0\n }\n \n ascii_letters = 0\n latin_extended = 0\n cyrillic = 0\n greek = 0\n other_letters = 0\n \n for char in text:\n \n code = ord(char)\n \n if (\"A\" <= char <= \"Z\") or (\"a\" <= char <= \"z\"):\n ascii_letters += 1\n \n elif 0x00C0 <= code <= 0x024F:\n latin_extended += 1\n \n elif 0x0400 <= code <= 0x04FF:\n cyrillic += 1\n \n elif 0x0370 <= code <= 0x03FF:\n greek += 1\n \n elif char.isalpha():\n other_letters += 1\n \n \n total_letters = (\n ascii_letters\n + latin_extended\n + cyrillic\n + greek\n + other_letters\n )\n \n return {\n \"ascii_letters\": ascii_letters,\n \"latin_extended\": latin_extended,\n \"cyrillic\": cyrillic,\n \"greek\": greek,\n \"other_letters\": other_letters,\n \"total_letters\": total_letters\n }\n\n\nlanguage_rows = []\n\nfor _, row in train.iterrows():\n \n stats = script_statistics(row[\"Report\"])\n \n total = stats[\"total_letters\"]\n \n if total == 0:\n dominant = \"UNKNOWN\"\n \n elif (\n stats[\"cyrillic\"] / total > 0.5\n ):\n dominant = \"CYRILLIC\"\n \n elif (\n stats[\"greek\"] / total > 0.5\n ):\n dominant = \"GREEK\"\n \n elif (\n (\n stats[\"ascii_letters\"]\n + stats[\"latin_extended\"]\n ) / total > 0.8\n ):\n dominant = \"LATIN\"\n \n else:\n dominant = \"MIXED/OTHER\"\n \n \n language_rows.append({\n \"StudyInstanceUID\": row[\"StudyInstanceUID\"],\n \"dominant_script\": dominant,\n \"ascii_letters\": stats[\"ascii_letters\"],\n \"latin_extended\": stats[\"latin_extended\"],\n \"cyrillic\": stats[\"cyrillic\"],\n \"greek\": stats[\"greek\"],\n \"other_letters\": stats[\"other_letters\"],\n \"total_letters\": total,\n \"report_words\": len(str(row[\"Report\"]).split())\n })\n\n\nreport_script_audit = pd.DataFrame(language_rows)\n\nprint(\"Dominant report script:\")\ndisplay(\n report_script_audit[\"dominant_script\"]\n .value_counts()\n)\n\nprint(\"\\nReport length:\")\ndisplay(\n report_script_audit[\"report_words\"].describe()\n)\n\nreport_script_audit.to_csv(\n OUTPUT_DIR / \"report_script_audit.csv\",\n index=False\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:33.285715Z","iopub.execute_input":"2026-08-11T07:54:33.286066Z","iopub.status.idle":"2026-08-11T07:54:34.64945Z","shell.execute_reply.started":"2026-08-11T07:54:33.28604Z","shell.execute_reply":"2026-08-11T07:54:34.64848Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 15: TEST DATASET STRUCTURE AUDIT\n# ============================================================\n\nprint(\"=\" * 70)\nprint(\"TEST DATASET AUDIT\")\nprint(\"=\" * 70)\n\nprint(\"\\nTEST CSV\")\nprint(\"Rows:\", len(test))\nprint(\"Unique studies:\", test[\"StudyInstanceUID\"].nunique())\n\nprint(\"\\nTEST SERIES CSV\")\nprint(\"Rows:\", len(test_series))\nprint(\n \"Unique studies:\",\n test_series[\"StudyInstanceUID\"].nunique()\n)\nprint(\n \"Unique series:\",\n test_series[\"SeriesInstanceUID\"].nunique()\n)\n\nprint(\"\\nSAMPLE SUBMISSION\")\nprint(\"Rows:\", len(sample_submission))\nprint(\n \"Unique StudyInstanceUID:\",\n sample_submission[\"StudyInstanceUID\"].nunique()\n)\n\nprint(\"\\nTest series per study:\")\n\ntest_series_per_study = (\n test_series\n .groupby(\"StudyInstanceUID\")\n [\"SeriesInstanceUID\"]\n .nunique()\n)\n\ndisplay(test_series_per_study)\n\nprint(\"\\nTest series metadata:\")\ndisplay(test_series)\n\nprint(\"\\nSample submission columns:\")\nprint(sample_submission.columns.tolist())\n\n# ------------------------------------------------------------\n# Check that sample submission IDs exactly match test IDs\n# ------------------------------------------------------------\n\ntest_ids = set(\n test[\"StudyInstanceUID\"].astype(str)\n)\n\nsubmission_ids = set(\n sample_submission[\"StudyInstanceUID\"].astype(str)\n)\n\nprint(\"\\nTest/submission ID comparison:\")\nprint(\"IDs in test but not submission:\",\n test_ids - submission_ids)\n\nprint(\"IDs in submission but not test:\",\n submission_ids - test_ids)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:34.650726Z","iopub.execute_input":"2026-08-11T07:54:34.65119Z","iopub.status.idle":"2026-08-11T07:54:34.678128Z","shell.execute_reply.started":"2026-08-11T07:54:34.651141Z","shell.execute_reply":"2026-08-11T07:54:34.677055Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 16: DICOM DIRECTORY STRUCTURE AUDIT\n# ============================================================\n\nfrom pathlib import Path\nfrom collections import Counter\nimport os\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"DICOM DIRECTORY STRUCTURE AUDIT\")\nprint(\"=\" * 70)\n\n# Use paths discovered earlier\nprint(\"\\nTrain series directory:\")\nprint(TRAIN_SERIES_DIR)\n\nprint(\"\\nTest series directory:\")\nprint(TEST_SERIES_DIR)\n\n# Basic existence checks\nprint(\"\\nDirectory existence:\")\nprint(\"Train series exists:\", TRAIN_SERIES_DIR.exists())\nprint(\"Test series exists:\", TEST_SERIES_DIR.exists())\n\n# ------------------------------------------------------------\n# Study directories\n# ------------------------------------------------------------\n\ntrain_studies = [\n p for p in TRAIN_SERIES_DIR.iterdir()\n if p.is_dir()\n]\n\ntest_studies = [\n p for p in TEST_SERIES_DIR.iterdir()\n if p.is_dir()\n]\n\nprint(\"\\nTrain study directories:\", len(train_studies))\nprint(\"Test study directories:\", len(test_studies))\n\n# ------------------------------------------------------------\n# Series directory counts\n# ------------------------------------------------------------\n\ntrain_series_dirs = []\ntest_series_dirs = []\n\nfor study_dir in train_studies:\n train_series_dirs.extend(\n [p for p in study_dir.iterdir() if p.is_dir()]\n )\n\nfor study_dir in test_studies:\n test_series_dirs.extend(\n [p for p in study_dir.iterdir() if p.is_dir()]\n )\n\nprint(\"\\nTrain series directories:\", len(train_series_dirs))\nprint(\"Test series directories:\", len(test_series_dirs))\n\n# ------------------------------------------------------------\n# Compare directory counts with CSV metadata\n# ------------------------------------------------------------\n\nprint(\"\\nCSV vs filesystem comparison:\")\n\nprint(\n \"train_series.csv unique series:\",\n train_series[\"SeriesInstanceUID\"].nunique()\n)\n\nprint(\n \"Filesystem train series:\",\n len(train_series_dirs)\n)\n\nprint(\n \"test_series.csv unique series:\",\n test_series[\"SeriesInstanceUID\"].nunique()\n)\n\nprint(\n \"Filesystem test series:\",\n len(test_series_dirs)\n)\n\n# ------------------------------------------------------------\n# Sample directory structures\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"SAMPLE TRAIN DIRECTORY STRUCTURE\")\nprint(\"=\" * 70)\n\nfor study_dir in train_studies[:3]:\n\n print(\"\\nStudy:\", study_dir.name)\n\n series_dirs = [\n p for p in study_dir.iterdir()\n if p.is_dir()\n ]\n\n for series_dir in series_dirs[:5]:\n\n print(\n \" Series:\",\n series_dir.name\n )\n\n files = [\n p for p in series_dir.iterdir()\n if p.is_file()\n ]\n\n print(\n \" Files:\",\n len(files)\n )\n\n if files:\n print(\n \" Example file:\",\n files[0].name\n )\n\n# ------------------------------------------------------------\n# Sample test directory structures\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"SAMPLE TEST DIRECTORY STRUCTURE\")\nprint(\"=\" * 70)\n\nfor study_dir in test_studies[:3]:\n\n print(\"\\nStudy:\", study_dir.name)\n\n series_dirs = [\n p for p in study_dir.iterdir()\n if p.is_dir()\n ]\n\n for series_dir in series_dirs:\n\n files = [\n p for p in series_dir.iterdir()\n if p.is_file()\n ]\n\n print(\n \" Series:\",\n series_dir.name,\n \"| Files:\",\n len(files)\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:34.679526Z","iopub.execute_input":"2026-08-11T07:54:34.679847Z","iopub.status.idle":"2026-08-11T07:54:48.044552Z","shell.execute_reply.started":"2026-08-11T07:54:34.679811Z","shell.execute_reply":"2026-08-11T07:54:48.043428Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 17: SLICE COUNT AUDIT\n# ============================================================\n\nfrom collections import Counter\nimport numpy as np\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"SLICE COUNT AUDIT\")\nprint(\"=\" * 70)\n\n\ndef count_files_in_series(series_dir):\n \"\"\"\n Count files in one series directory without reading\n DICOM pixel data.\n \"\"\"\n return sum(\n 1\n for p in series_dir.iterdir()\n if p.is_file()\n )\n\n\n# ------------------------------------------------------------\n# TRAIN SERIES\n# ------------------------------------------------------------\n\nprint(\"\\nCounting train-series files...\")\n\ntrain_slice_rows = []\n\nfor i, series_dir in enumerate(train_series_dirs):\n\n count = count_files_in_series(series_dir)\n\n train_slice_rows.append({\n \"StudyInstanceUID\": series_dir.parent.name,\n \"SeriesInstanceUID\": series_dir.name,\n \"slice_count\": count\n })\n\n if (i + 1) % 5000 == 0:\n print(\n f\"Processed {i + 1:,} / \"\n f\"{len(train_series_dirs):,} train series\"\n )\n\n\ntrain_slice_counts = pd.DataFrame(train_slice_rows)\n\n\n# ------------------------------------------------------------\n# TEST SERIES\n# ------------------------------------------------------------\n\nprint(\"\\nCounting test-series files...\")\n\ntest_slice_rows = []\n\nfor series_dir in test_series_dirs:\n\n count = count_files_in_series(series_dir)\n\n test_slice_rows.append({\n \"StudyInstanceUID\": series_dir.parent.name,\n \"SeriesInstanceUID\": series_dir.name,\n \"slice_count\": count\n })\n\n\ntest_slice_counts = pd.DataFrame(test_slice_rows)\n\n\n# ------------------------------------------------------------\n# BASIC TRAIN STATISTICS\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TRAIN SLICE COUNT STATISTICS\")\nprint(\"=\" * 70)\n\ndisplay(\n train_slice_counts[\"slice_count\"].describe(\n percentiles=[\n 0.01,\n 0.05,\n 0.10,\n 0.25,\n 0.50,\n 0.75,\n 0.90,\n 0.95,\n 0.99\n ]\n )\n)\n\n\n# ------------------------------------------------------------\n# TEST STATISTICS\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TEST SLICE COUNT STATISTICS\")\nprint(\"=\" * 70)\n\ndisplay(\n test_slice_counts[\"slice_count\"].describe()\n)\n\n\n# ------------------------------------------------------------\n# SLICE COUNT FREQUENCY\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"MOST COMMON TRAIN SLICE COUNTS\")\nprint(\"=\" * 70)\n\ndisplay(\n train_slice_counts[\"slice_count\"]\n .value_counts()\n .sort_index()\n .head(50)\n)\n\n\n# ------------------------------------------------------------\n# EXTREME SERIES\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"SMALLEST TRAIN SERIES\")\nprint(\"=\" * 70)\n\ndisplay(\n train_slice_counts\n .sort_values(\"slice_count\")\n .head(20)\n)\n\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"LARGEST TRAIN SERIES\")\nprint(\"=\" * 70)\n\ndisplay(\n train_slice_counts\n .sort_values(\"slice_count\", ascending=False)\n .head(20)\n)\n\n\n# ------------------------------------------------------------\n# Merge with series metadata\n# ------------------------------------------------------------\n\ntrain_slice_audit = train_slice_counts.merge(\n train_series,\n on=[\"StudyInstanceUID\", \"SeriesInstanceUID\"],\n how=\"left\",\n validate=\"one_to_one\"\n)\n\ntest_slice_audit = test_slice_counts.merge(\n test_series,\n on=[\"StudyInstanceUID\", \"SeriesInstanceUID\"],\n how=\"left\",\n validate=\"one_to_one\"\n)\n\n\n# ------------------------------------------------------------\n# Slice counts by anatomical plane\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"SLICE COUNT BY ANATOMICAL PLANE\")\nprint(\"=\" * 70)\n\nplane_stats = (\n train_slice_audit\n .groupby(\"Anatomical_Plane\")[\"slice_count\"]\n .agg(\n [\"count\", \"min\", \"median\", \"mean\", \"max\"]\n )\n .sort_values(\"count\", ascending=False)\n)\n\ndisplay(plane_stats)\n\n\n# ------------------------------------------------------------\n# Slice counts by sequence properties\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"SLICE COUNT BY SEQUENCE PROPERTIES\")\nprint(\"=\" * 70)\n\nsequence_stats = (\n train_slice_audit\n .groupby(\n [\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\"\n ]\n )[\"slice_count\"]\n .agg(\n [\"count\", \"min\", \"median\", \"mean\", \"max\"]\n )\n .sort_values(\"count\", ascending=False)\n)\n\ndisplay(sequence_stats)\n\n\n# ------------------------------------------------------------\n# Save audit files\n# ------------------------------------------------------------\n\ntrain_slice_audit.to_csv(\n OUTPUT_DIR / \"train_slice_audit.csv\",\n index=False\n)\n\ntest_slice_audit.to_csv(\n OUTPUT_DIR / \"test_slice_audit.csv\",\n index=False\n)\n\nplane_stats.to_csv(\n OUTPUT_DIR / \"slice_count_by_plane.csv\"\n)\n\nsequence_stats.to_csv(\n OUTPUT_DIR / \"slice_count_by_sequence.csv\"\n)\n\n\nprint(\"\\nSaved:\")\nprint(\n OUTPUT_DIR / \"train_slice_audit.csv\"\n)\nprint(\n OUTPUT_DIR / \"test_slice_audit.csv\"\n)\nprint(\n OUTPUT_DIR / \"slice_count_by_plane.csv\"\n)\nprint(\n OUTPUT_DIR / \"slice_count_by_sequence.csv\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T07:54:48.046057Z","iopub.execute_input":"2026-08-11T07:54:48.046431Z","iopub.status.idle":"2026-08-11T08:12:20.196519Z","shell.execute_reply.started":"2026-08-11T07:54:48.046392Z","shell.execute_reply":"2026-08-11T08:12:20.19514Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 18: REPRESENTATIVE DICOM METADATA AUDIT\n# ============================================================\n\nimport random\nimport numpy as np\nimport pandas as pd\nimport pydicom\nfrom pathlib import Path\n\n\nprint(\"=\" * 70)\nprint(\"REPRESENTATIVE DICOM METADATA AUDIT\")\nprint(\"=\" * 70)\n\n\n# ------------------------------------------------------------\n# 1. Verify pydicom\n# ------------------------------------------------------------\n\nprint(\"\\nPydicom version:\")\nprint(pydicom.__version__)\n\n\n# ------------------------------------------------------------\n# 2. Helper function\n# ------------------------------------------------------------\n\ndef safe_get(ds, keyword, default=None):\n \"\"\"\n Safely retrieve a DICOM attribute.\n \"\"\"\n try:\n value = getattr(ds, keyword, default)\n\n if value is None:\n return default\n\n return value\n\n except Exception:\n return default\n\n\ndef normalize_dicom_value(value):\n \"\"\"\n Convert DICOM values into CSV-friendly representations.\n \"\"\"\n\n if value is None:\n return None\n\n if isinstance(value, (list, tuple)):\n return \"|\".join(str(x) for x in value)\n\n return str(value)\n\n\n# ------------------------------------------------------------\n# 3. Select representative series\n# ------------------------------------------------------------\n\nmetadata_source = train_slice_audit.copy()\n\n# Keep only rows with usable series metadata.\nmetadata_source = metadata_source[\n metadata_source[\"SeriesInstanceUID\"].notna()\n].copy()\n\n\nselected_series = []\n\n\n# ------------------------------------------------------------\n# A. One representative series from each plane\n# ------------------------------------------------------------\n\nfor plane in sorted(\n metadata_source[\"Anatomical_Plane\"]\n .dropna()\n .astype(str)\n .unique()\n):\n\n candidates = metadata_source[\n metadata_source[\"Anatomical_Plane\"].astype(str) == plane\n ]\n\n if len(candidates) > 0:\n\n selected_series.append(\n candidates.sample(\n n=1,\n random_state=42\n ).iloc[0]\n )\n\n\n# ------------------------------------------------------------\n# B. One representative series from each\n# sequence-property combination\n# ------------------------------------------------------------\n\nsequence_columns = [\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\"\n]\n\nfor _, group in metadata_source.groupby(\n sequence_columns,\n dropna=False\n):\n\n if len(group) == 0:\n continue\n\n selected_series.append(\n group.sample(\n n=1,\n random_state=42\n ).iloc[0]\n )\n\n\n# ------------------------------------------------------------\n# C. Very small series\n# ------------------------------------------------------------\n\nsmallest = (\n metadata_source\n .sort_values(\"slice_count\")\n .head(2)\n)\n\nfor _, row in smallest.iterrows():\n selected_series.append(row)\n\n\n# ------------------------------------------------------------\n# D. Very large series\n# ------------------------------------------------------------\n\nlargest = (\n metadata_source\n .sort_values(\n \"slice_count\",\n ascending=False\n )\n .head(3)\n)\n\nfor _, row in largest.iterrows():\n selected_series.append(row)\n\n\n# ------------------------------------------------------------\n# E. A few labeled studies if available\n# ------------------------------------------------------------\n\nlabeled_ids = set(\n labeled_train[\"StudyInstanceUID\"].astype(str)\n)\n\nlabeled_series = metadata_source[\n metadata_source[\"StudyInstanceUID\"]\n .astype(str)\n .isin(labeled_ids)\n]\n\nif len(labeled_series) > 0:\n\n selected_series.append(\n labeled_series.sample(\n n=1,\n random_state=42\n ).iloc[0]\n )\n\n\n# ------------------------------------------------------------\n# Remove duplicate series\n# ------------------------------------------------------------\n\nselected_df = pd.DataFrame(selected_series)\n\nselected_df = (\n selected_df\n .drop_duplicates(\n subset=[\"SeriesInstanceUID\"]\n )\n .reset_index(drop=True)\n)\n\n\nprint(\"\\nSelected representative series:\")\nprint(len(selected_df))\n\ndisplay(\n selected_df[\n [\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n \"slice_count\",\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\"\n ]\n ]\n)\n\n\n# ------------------------------------------------------------\n# 4. Read representative DICOM files\n# ------------------------------------------------------------\n\nmetadata_rows = []\nfailed_files = []\n\n\nfor _, series_row in selected_df.iterrows():\n\n study_uid = str(\n series_row[\"StudyInstanceUID\"]\n )\n\n series_uid = str(\n series_row[\"SeriesInstanceUID\"]\n )\n\n series_dir = (\n TRAIN_SERIES_DIR\n / study_uid\n / series_uid\n )\n\n files = sorted(\n [\n p\n for p in series_dir.iterdir()\n if p.is_file()\n ]\n )\n\n if not files:\n print(\n \"\\nWARNING: No files found:\",\n series_dir\n )\n continue\n\n\n # Select first, middle and last files.\n positions = sorted(\n set(\n [\n 0,\n len(files) // 2,\n len(files) - 1\n ]\n )\n )\n\n\n for position in positions:\n\n filepath = files[position]\n\n try:\n\n # IMPORTANT:\n # stop_before_pixels=True means pixel arrays\n # are NOT loaded.\n ds = pydicom.dcmread(\n filepath,\n stop_before_pixels=True,\n force=True\n )\n\n\n row = {\n \"StudyInstanceUID\":\n study_uid,\n\n \"SeriesInstanceUID\":\n series_uid,\n\n \"FileName\":\n filepath.name,\n\n \"FilePosition\":\n position,\n\n \"DirectorySliceCount\":\n len(files),\n\n \"Rows\":\n safe_get(ds, \"Rows\"),\n\n \"Columns\":\n safe_get(ds, \"Columns\"),\n\n \"PixelSpacing\":\n normalize_dicom_value(\n safe_get(ds, \"PixelSpacing\")\n ),\n\n \"SliceThickness\":\n safe_get(ds, \"SliceThickness\"),\n\n \"SpacingBetweenSlices\":\n safe_get(\n ds,\n \"SpacingBetweenSlices\"\n ),\n\n \"ImageOrientationPatient\":\n normalize_dicom_value(\n safe_get(\n ds,\n \"ImageOrientationPatient\"\n )\n ),\n\n \"ImagePositionPatient\":\n normalize_dicom_value(\n safe_get(\n ds,\n \"ImagePositionPatient\"\n )\n ),\n\n \"InstanceNumber\":\n safe_get(ds, \"InstanceNumber\"),\n\n \"PhotometricInterpretation\":\n safe_get(\n ds,\n \"PhotometricInterpretation\"\n ),\n\n \"BitsAllocated\":\n safe_get(\n ds,\n \"BitsAllocated\"\n ),\n\n \"BitsStored\":\n safe_get(\n ds,\n \"BitsStored\"\n ),\n\n \"HighBit\":\n safe_get(ds, \"HighBit\"),\n\n \"PixelRepresentation\":\n safe_get(\n ds,\n \"PixelRepresentation\"\n ),\n\n \"RescaleSlope\":\n safe_get(\n ds,\n \"RescaleSlope\"\n ),\n\n \"RescaleIntercept\":\n safe_get(\n ds,\n \"RescaleIntercept\"\n ),\n\n \"TransferSyntaxUID\":\n normalize_dicom_value(\n safe_get(\n ds,\n \"file_meta\"\n )\n ),\n\n \"SOPInstanceUID\":\n safe_get(\n ds,\n \"SOPInstanceUID\"\n ),\n\n \"DICOMStudyInstanceUID\":\n safe_get(\n ds,\n \"StudyInstanceUID\"\n ),\n\n \"DICOMSeriesInstanceUID\":\n safe_get(\n ds,\n \"SeriesInstanceUID\"\n ),\n\n \"DICOMModality\":\n safe_get(ds, \"Modality\"),\n\n \"Manufacturer\":\n safe_get(ds, \"Manufacturer\"),\n\n \"ManufacturerModelName\":\n safe_get(\n ds,\n \"ManufacturerModelName\"\n ),\n\n \"SeriesDescription\":\n safe_get(\n ds,\n \"SeriesDescription\"\n )\n }\n\n metadata_rows.append(row)\n\n\n except Exception as e:\n\n failed_files.append({\n \"StudyInstanceUID\":\n study_uid,\n\n \"SeriesInstanceUID\":\n series_uid,\n\n \"FileName\":\n filepath.name,\n\n \"Error\":\n repr(e)\n })\n\n\n# ------------------------------------------------------------\n# 5. Create metadata DataFrame\n# ------------------------------------------------------------\n\ndicom_metadata_audit = pd.DataFrame(\n metadata_rows\n)\n\ndicom_failures = pd.DataFrame(\n failed_files\n)\n\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"DICOM METADATA RESULTS\")\nprint(\"=\" * 70)\n\nprint(\n \"\\nDICOM files successfully inspected:\",\n len(dicom_metadata_audit)\n)\n\nprint(\n \"DICOM files that failed:\",\n len(dicom_failures)\n)\n\n\n# ------------------------------------------------------------\n# 6. Display metadata\n# ------------------------------------------------------------\n\ndisplay(\n dicom_metadata_audit\n)\n\n\n# ------------------------------------------------------------\n# 7. Missing metadata analysis\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"MISSING METADATA\")\nprint(\"=\" * 70)\n\nif len(dicom_metadata_audit) > 0:\n\n metadata_missingness = (\n dicom_metadata_audit\n .isna()\n .sum()\n .sort_values(\n ascending=False\n )\n )\n\n display(\n metadata_missingness[\n metadata_missingness > 0\n ]\n )\n\n\n# ------------------------------------------------------------\n# 8. Common dimensions\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"IMAGE DIMENSIONS\")\nprint(\"=\" * 70)\n\nif len(dicom_metadata_audit) > 0:\n\n dimension_counts = (\n dicom_metadata_audit[\n [\"Rows\", \"Columns\"]\n ]\n .value_counts()\n .reset_index(\n name=\"count\"\n )\n )\n\n display(\n dimension_counts.head(20)\n )\n\n\n# ------------------------------------------------------------\n# 9. Photometric interpretation\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PHOTOMETRIC INTERPRETATION\")\nprint(\"=\" * 70)\n\ndisplay(\n dicom_metadata_audit[\n \"PhotometricInterpretation\"\n ].value_counts(\n dropna=False\n )\n)\n\n\n# ------------------------------------------------------------\n# 10. Transfer syntax\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TRANSFER SYNTAX\")\nprint(\"=\" * 70)\n\ndisplay(\n dicom_metadata_audit[\n \"TransferSyntaxUID\"\n ].value_counts(\n dropna=False\n )\n)\n\n\n# ------------------------------------------------------------\n# 11. Modality\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"MODALITY\")\nprint(\"=\" * 70)\n\ndisplay(\n dicom_metadata_audit[\n \"DICOMModality\"\n ].value_counts(\n dropna=False\n )\n)\n\n\n# ------------------------------------------------------------\n# 12. Save results\n# ------------------------------------------------------------\n\ndicom_metadata_audit.to_csv(\n OUTPUT_DIR / \"dicom_metadata_audit.csv\",\n index=False\n)\n\ndicom_failures.to_csv(\n OUTPUT_DIR / \"dicom_metadata_failures.csv\",\n index=False\n)\n\nprint(\"\\nSaved:\")\nprint(\n OUTPUT_DIR / \"dicom_metadata_audit.csv\"\n)\nprint(\n OUTPUT_DIR / \"dicom_metadata_failures.csv\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:12:20.19816Z","iopub.execute_input":"2026-08-11T08:12:20.198456Z","iopub.status.idle":"2026-08-11T08:12:21.965854Z","shell.execute_reply.started":"2026-08-11T08:12:20.198432Z","shell.execute_reply":"2026-08-11T08:12:21.964628Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 19: VISUAL DICOM + SLICE ORDER AUDIT\n# ============================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport pydicom\n\nfrom pathlib import Path\n\nprint(\"=\" * 70)\nprint(\"VISUAL DICOM + SLICE ORDER AUDIT\")\nprint(\"=\" * 70)\n\nTRAIN_SERIES_DIR = Path(\n \"/kaggle/input/competitions/rsna-knee-abnormality-detection/train_series\"\n)\n\nAUDIT_DIR = Path(\"/kaggle/working/rsna_knee_audit\")\nAUDIT_DIR.mkdir(parents=True, exist_ok=True)\n\n\n# ------------------------------------------------------------\n# 1. Helper: read a series\n# ------------------------------------------------------------\n\ndef read_dicom_series(series_path):\n records = []\n\n for file_path in sorted(series_path.glob(\"*.dcm\")):\n try:\n ds = pydicom.dcmread(str(file_path), force=False)\n\n if not hasattr(ds, \"PixelData\"):\n continue\n\n position = getattr(ds, \"ImagePositionPatient\", None)\n orientation = getattr(ds, \"ImageOrientationPatient\", None)\n\n if position is not None:\n position = np.asarray(position, dtype=np.float64)\n else:\n position = None\n\n if orientation is not None:\n orientation = np.asarray(orientation, dtype=np.float64)\n else:\n orientation = None\n\n records.append({\n \"path\": file_path,\n \"ds\": ds,\n \"position\": position,\n \"orientation\": orientation,\n \"instance_number\": getattr(ds, \"InstanceNumber\", None)\n })\n\n except Exception:\n continue\n\n if not records:\n raise ValueError(f\"No readable DICOM images found: {series_path}\")\n\n # --------------------------------------------------------\n # Determine spatial sorting direction\n # --------------------------------------------------------\n\n valid_positions = [\n r[\"position\"] for r in records\n if r[\"position\"] is not None\n ]\n\n orientation = None\n\n for r in records:\n if r[\"orientation\"] is not None:\n orientation = r[\"orientation\"]\n break\n\n if len(valid_positions) == len(records) and orientation is not None:\n row_cosines = orientation[:3]\n col_cosines = orientation[3:]\n normal = np.cross(row_cosines, col_cosines)\n\n for r in records:\n r[\"sort_position\"] = float(\n np.dot(r[\"position\"], normal)\n )\n\n records.sort(key=lambda x: x[\"sort_position\"])\n\n sort_method = \"ImagePositionPatient + orientation normal\"\n\n elif all(r[\"position\"] is not None for r in records):\n for r in records:\n r[\"sort_position\"] = float(\n r[\"position\"][2]\n )\n\n records.sort(key=lambda x: x[\"sort_position\"])\n\n sort_method = \"ImagePositionPatient Z\"\n\n elif all(r[\"instance_number\"] is not None for r in records):\n records.sort(key=lambda x: x[\"instance_number\"])\n\n sort_method = \"InstanceNumber\"\n\n else:\n records.sort(key=lambda x: str(x[\"path\"]))\n\n sort_method = \"Filename fallback\"\n\n images = []\n\n for r in records:\n try:\n pixel_array = r[\"ds\"].pixel_array.astype(np.float32)\n images.append(pixel_array)\n except Exception:\n continue\n\n return records, images, sort_method\n\n\n# ------------------------------------------------------------\n# 2. Find representative series\n# ------------------------------------------------------------\n\nprint(\"\\nSearching for representative series...\")\n\nseries_dirs = [\n p for p in TRAIN_SERIES_DIR.rglob(\"*\")\n if p.is_dir() and any(p.glob(\"*.dcm\"))\n]\n\nprint(f\"Series directories discovered: {len(series_dirs)}\")\n\n\n# ------------------------------------------------------------\n# 3. Gather basic information without reading every pixel\n# ------------------------------------------------------------\n\nrepresentatives = {\n \"sagittal_normal\": None,\n \"coronal_normal\": None,\n \"axial_normal\": None,\n \"small_11\": None,\n \"large_320\": None\n}\n\nfor series_path in series_dirs:\n\n files = list(series_path.glob(\"*.dcm\"))\n count = len(files)\n\n if count == 0:\n continue\n\n try:\n ds = pydicom.dcmread(\n str(files[0]),\n stop_before_pixels=True\n )\n except Exception:\n continue\n\n plane = str(getattr(ds, \"AnatomicalPlane\", \"\")).strip()\n\n if plane == \"\":\n description = str(\n getattr(ds, \"SeriesDescription\", \"\")\n ).upper()\n\n if \"SAG\" in description:\n plane = \"Sagittal\"\n elif \"COR\" in description:\n plane = \"Coronal\"\n elif \"TRA\" in description or \"AX\" in description:\n plane = \"Axial\"\n\n if count == 11 and representatives[\"small_11\"] is None:\n representatives[\"small_11\"] = series_path\n\n if count == 320 and representatives[\"large_320\"] is None:\n representatives[\"large_320\"] = series_path\n\n if 20 <= count <= 40:\n\n if (\n plane == \"Sagittal\"\n and representatives[\"sagittal_normal\"] is None\n ):\n representatives[\"sagittal_normal\"] = series_path\n\n elif (\n plane == \"Coronal\"\n and representatives[\"coronal_normal\"] is None\n ):\n representatives[\"coronal_normal\"] = series_path\n\n elif (\n plane == \"Axial\"\n and representatives[\"axial_normal\"] is None\n ):\n representatives[\"axial_normal\"] = series_path\n\n if all(v is not None for v in representatives.values()):\n break\n\n\nprint(\"\\nSelected representatives:\")\n\nfor name, path in representatives.items():\n print(f\"{name}: {path}\")\n\n\n# ------------------------------------------------------------\n# 4. Analyze and visualize each representative\n# ------------------------------------------------------------\n\naudit_rows = []\n\nfor name, series_path in representatives.items():\n\n if series_path is None:\n print(f\"\\n{name}: NOT FOUND\")\n continue\n\n print(\"\\n\" + \"=\" * 70)\n print(f\"SERIES: {name}\")\n print(\"=\" * 70)\n\n records, images, sort_method = read_dicom_series(series_path)\n\n if not images:\n print(\"No pixel arrays could be decoded.\")\n continue\n\n first_ds = records[0][\"ds\"]\n\n volume = np.stack(images)\n\n print(f\"Directory: {series_path}\")\n print(f\"Slice count: {len(images)}\")\n print(f\"Volume shape: {volume.shape}\")\n print(f\"Dtype: {volume.dtype}\")\n print(f\"Min: {volume.min():.3f}\")\n print(f\"Max: {volume.max():.3f}\")\n print(f\"Mean: {volume.mean():.3f}\")\n print(f\"Std: {volume.std():.3f}\")\n print(f\"Sort method: {sort_method}\")\n\n print(\n \"Pixel spacing:\",\n getattr(first_ds, \"PixelSpacing\", \"MISSING\")\n )\n\n print(\n \"Slice thickness:\",\n getattr(first_ds, \"SliceThickness\", \"MISSING\")\n )\n\n print(\n \"Spacing between slices:\",\n getattr(first_ds, \"SpacingBetweenSlices\", \"MISSING\")\n )\n\n print(\n \"Image orientation:\",\n getattr(first_ds, \"ImageOrientationPatient\", \"MISSING\")\n )\n\n print(\n \"Image position available:\",\n all(\n r[\"position\"] is not None\n for r in records\n )\n )\n\n print(\n \"Photometric interpretation:\",\n getattr(\n first_ds,\n \"PhotometricInterpretation\",\n \"MISSING\"\n )\n )\n\n print(\n \"Series description:\",\n getattr(\n first_ds,\n \"SeriesDescription\",\n \"MISSING\"\n )\n )\n\n # --------------------------------------------------------\n # Check spatial positions\n # --------------------------------------------------------\n\n positions = [\n r[\"sort_position\"]\n for r in records\n if \"sort_position\" in r\n ]\n\n duplicate_positions = 0\n\n if positions:\n rounded_positions = np.round(\n np.asarray(positions),\n decimals=5\n )\n\n duplicate_positions = (\n len(rounded_positions)\n - len(np.unique(rounded_positions))\n )\n\n print(\n \"Duplicate spatial positions:\",\n duplicate_positions\n )\n\n # --------------------------------------------------------\n # Save audit row\n # --------------------------------------------------------\n\n audit_rows.append({\n \"representative\": name,\n \"series_path\": str(series_path),\n \"slice_count\": len(images),\n \"rows\": volume.shape[1],\n \"columns\": volume.shape[2],\n \"dtype\": str(volume.dtype),\n \"min\": float(volume.min()),\n \"max\": float(volume.max()),\n \"mean\": float(volume.mean()),\n \"std\": float(volume.std()),\n \"sort_method\": sort_method,\n \"pixel_spacing\": str(\n getattr(first_ds, \"PixelSpacing\", \"MISSING\")\n ),\n \"slice_thickness\": str(\n getattr(first_ds, \"SliceThickness\", \"MISSING\")\n ),\n \"spacing_between_slices\": str(\n getattr(\n first_ds,\n \"SpacingBetweenSlices\",\n \"MISSING\"\n )\n ),\n \"photometric\": str(\n getattr(\n first_ds,\n \"PhotometricInterpretation\",\n \"MISSING\"\n )\n ),\n \"duplicate_spatial_positions\": duplicate_positions\n })\n\n # --------------------------------------------------------\n # Normalize only for visualization\n # --------------------------------------------------------\n\n display_volume = volume.copy()\n\n low = np.percentile(display_volume, 1)\n high = np.percentile(display_volume, 99)\n\n if high > low:\n display_volume = np.clip(\n display_volume,\n low,\n high\n )\n\n display_volume = (\n display_volume - low\n ) / (high - low)\n\n # --------------------------------------------------------\n # First / middle / last\n # --------------------------------------------------------\n\n indices = [\n 0,\n len(images) // 2,\n len(images) - 1\n ]\n\n fig, axes = plt.subplots(\n 1,\n 3,\n figsize=(15, 5)\n )\n\n for ax, idx in zip(axes, indices):\n\n ax.imshow(\n display_volume[idx],\n cmap=\"gray\"\n )\n\n ax.set_title(\n f\"Slice {idx + 1}/{len(images)}\"\n )\n\n ax.axis(\"off\")\n\n fig.suptitle(\n f\"{name}\\n{series_path.name}\",\n fontsize=12\n )\n\n plt.tight_layout()\n\n output_path = (\n AUDIT_DIR /\n f\"cell19_{name}_first_middle_last.png\"\n )\n\n plt.savefig(\n output_path,\n dpi=150,\n bbox_inches=\"tight\"\n )\n\n plt.show()\n\n # --------------------------------------------------------\n # Contact sheet\n # --------------------------------------------------------\n\n contact_count = min(len(images), 12)\n\n contact_indices = np.linspace(\n 0,\n len(images) - 1,\n contact_count,\n dtype=int\n )\n\n fig, axes = plt.subplots(\n 3,\n 4,\n figsize=(12, 9)\n )\n\n axes = axes.ravel()\n\n for ax in axes:\n ax.axis(\"off\")\n\n for ax, idx in zip(\n axes,\n contact_indices\n ):\n\n ax.imshow(\n display_volume[idx],\n cmap=\"gray\"\n )\n\n ax.set_title(\n f\"{idx + 1}/{len(images)}\"\n )\n\n ax.axis(\"off\")\n\n fig.suptitle(\n f\"Contact Sheet: {name}\",\n fontsize=14\n )\n\n plt.tight_layout()\n\n output_path = (\n AUDIT_DIR /\n f\"cell19_{name}_contact_sheet.png\"\n )\n\n plt.savefig(\n output_path,\n dpi=150,\n bbox_inches=\"tight\"\n )\n\n plt.show()\n\n\n# ------------------------------------------------------------\n# 5. Save summary\n# ------------------------------------------------------------\n\naudit_df = pd.DataFrame(audit_rows)\n\nsummary_path = (\n AUDIT_DIR /\n \"cell19_visual_dicom_audit.csv\"\n)\n\naudit_df.to_csv(\n summary_path,\n index=False\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 19 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"\\nSaved:\")\nprint(summary_path)\n\nprint(\"\\nGenerated visualization files:\")\nfor p in sorted(AUDIT_DIR.glob(\"cell19_*.png\")):\n print(p)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:12:21.967643Z","iopub.execute_input":"2026-08-11T08:12:21.968113Z","iopub.status.idle":"2026-08-11T08:30:51.624845Z","shell.execute_reply.started":"2026-08-11T08:12:21.968083Z","shell.execute_reply":"2026-08-11T08:30:51.623968Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 20: BROAD DICOM METADATA + PIXEL AUDIT\n# ============================================================\n\nimport os\nimport random\nimport numpy as np\nimport pandas as pd\nimport pydicom\n\nfrom pathlib import Path\n\nprint(\"=\" * 70)\nprint(\"BROAD DICOM METADATA + PIXEL AUDIT\")\nprint(\"=\" * 70)\n\n# ------------------------------------------------------------\n# Configuration\n# ------------------------------------------------------------\n\nSEED = 42\nrandom.seed(SEED)\nnp.random.seed(SEED)\n\nAUDIT_DIR = Path(\"/kaggle/working/rsna_knee_audit\")\n\nTRAIN_SERIES_DIR = Path(\n \"/kaggle/input/competitions/rsna-knee-abnormality-detection/train_series\"\n)\n\nTEST_SERIES_DIR = Path(\n \"/kaggle/input/competitions/rsna-knee-abnormality-detection/test_series\"\n)\n\nTRAIN_SERIES_CSV = (\n \"/kaggle/input/competitions/rsna-knee-abnormality-detection/train_series.csv\"\n)\n\nTEST_SERIES_CSV = (\n \"/kaggle/input/competitions/rsna-knee-abnormality-detection/test_series.csv\"\n)\n\nOUTPUT_FILE = AUDIT_DIR / \"dicom_audit.csv\"\n\nTRAIN_SAMPLE_SIZE = 250\n\n\n# ------------------------------------------------------------\n# 1. Load series metadata\n# ------------------------------------------------------------\n\ntrain_series = pd.read_csv(TRAIN_SERIES_CSV)\ntest_series = pd.read_csv(TEST_SERIES_CSV)\n\nprint(\"\\nTrain series rows:\", len(train_series))\nprint(\"Test series rows:\", len(test_series))\n\n\n# ------------------------------------------------------------\n# 2. Build representative training sample\n# ------------------------------------------------------------\n\nprint(\"\\nCreating stratified training sample...\")\n\nsample_parts = []\n\n# Slice-count audit is available from Cell 17.\n# Use it if available, otherwise calculate counts from filesystem.\n\nslice_audit_file = AUDIT_DIR / \"train_slice_audit.csv\"\n\nif slice_audit_file.exists():\n\n slice_audit = pd.read_csv(slice_audit_file)\n\n train_meta = train_series.merge(\n slice_audit[\n [\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n \"slice_count\"\n ]\n ],\n on=[\"StudyInstanceUID\", \"SeriesInstanceUID\"],\n how=\"left\"\n )\n\nelse:\n\n print(\n \"train_slice_audit.csv not found. \"\n \"Using metadata-only sampling.\"\n )\n\n train_meta = train_series.copy()\n train_meta[\"slice_count\"] = np.nan\n\n\n# ------------------------------------------------------------\n# Define sampling groups\n# ------------------------------------------------------------\n\ntrain_meta[\"plane_group\"] = (\n train_meta[\"Anatomical_Plane\"]\n .fillna(\"UNKNOWN\")\n .astype(str)\n)\n\ntrain_meta[\"sequence_group\"] = (\n train_meta[\"Fluid_Sensitive\"].fillna(-1).astype(str)\n + \"_\"\n + train_meta[\"Fat_Suppression\"].fillna(-1).astype(str)\n)\n\nif train_meta[\"slice_count\"].notna().any():\n\n train_meta[\"slice_group\"] = pd.cut(\n train_meta[\"slice_count\"],\n bins=[0, 15, 25, 35, 50, 100, 1000],\n labels=[\n \"11-15\",\n \"16-25\",\n \"26-35\",\n \"36-50\",\n \"51-100\",\n \"101+\"\n ],\n include_lowest=True\n )\n\nelse:\n\n train_meta[\"slice_group\"] = \"UNKNOWN\"\n\n\n# ------------------------------------------------------------\n# Stratified sample\n# ------------------------------------------------------------\n\ngrouped = train_meta.groupby(\n [\n \"plane_group\",\n \"sequence_group\",\n \"slice_group\"\n ],\n dropna=False,\n observed=True\n)\n\nn_groups = len(grouped)\n\nper_group = max(\n 1,\n TRAIN_SAMPLE_SIZE // max(n_groups, 1)\n)\n\nfor _, group in grouped:\n\n take = min(\n len(group),\n per_group\n )\n\n if take > 0:\n sample_parts.append(\n group.sample(\n n=take,\n random_state=SEED\n )\n )\n\ntrain_sample = pd.concat(\n sample_parts,\n ignore_index=True\n) if sample_parts else train_meta.sample(\n n=min(TRAIN_SAMPLE_SIZE, len(train_meta)),\n random_state=SEED\n)\n\n# Fill remaining slots if stratified sampling produced too few rows.\n\nif len(train_sample) < TRAIN_SAMPLE_SIZE:\n\n remaining = train_meta[\n ~train_meta[\"SeriesInstanceUID\"].isin(\n train_sample[\"SeriesInstanceUID\"]\n )\n ]\n\n extra_n = min(\n TRAIN_SAMPLE_SIZE - len(train_sample),\n len(remaining)\n )\n\n if extra_n > 0:\n\n extra = remaining.sample(\n n=extra_n,\n random_state=SEED\n )\n\n train_sample = pd.concat(\n [\n train_sample,\n extra\n ],\n ignore_index=True\n )\n\ntrain_sample = train_sample.drop_duplicates(\n subset=[\"SeriesInstanceUID\"]\n).head(TRAIN_SAMPLE_SIZE)\n\nprint(\n f\"Selected {len(train_sample)} \"\n \"training series for detailed DICOM audit.\"\n)\n\n\n# ------------------------------------------------------------\n# 3. Helper functions\n# ------------------------------------------------------------\n\ndef get_series_directory(root, study_uid, series_uid):\n\n return (\n Path(root)\n / str(study_uid)\n / str(series_uid)\n )\n\n\ndef choose_middle_file(series_dir):\n\n files = sorted(\n series_dir.glob(\"*.dcm\")\n )\n\n if not files:\n return None\n\n return files[len(files) // 2]\n\n\ndef safe_value(ds, field):\n\n try:\n value = getattr(ds, field, None)\n\n if value is None:\n return None\n\n if isinstance(value, (list, tuple)):\n return str(list(value))\n\n return str(value)\n\n except Exception:\n\n return None\n\n\ndef safe_float(ds, field):\n\n try:\n\n value = getattr(ds, field, None)\n\n if value is None:\n return np.nan\n\n return float(value)\n\n except Exception:\n\n return np.nan\n\n\ndef get_pixel_statistics(ds):\n\n try:\n\n if not hasattr(ds, \"PixelData\"):\n return {\n \"pixel_loaded\": False,\n \"pixel_min\": np.nan,\n \"pixel_max\": np.nan,\n \"pixel_mean\": np.nan,\n \"pixel_std\": np.nan,\n \"pixel_p01\": np.nan,\n \"pixel_p50\": np.nan,\n \"pixel_p99\": np.nan\n }\n\n arr = ds.pixel_array.astype(\n np.float32\n )\n\n # Apply rescale when present.\n\n slope = safe_float(\n ds,\n \"RescaleSlope\"\n )\n\n intercept = safe_float(\n ds,\n \"RescaleIntercept\"\n )\n\n if np.isfinite(slope):\n arr = arr * slope\n\n if np.isfinite(intercept):\n arr = arr + intercept\n\n return {\n \"pixel_loaded\": True,\n \"pixel_min\": float(np.min(arr)),\n \"pixel_max\": float(np.max(arr)),\n \"pixel_mean\": float(np.mean(arr)),\n \"pixel_std\": float(np.std(arr)),\n \"pixel_p01\": float(np.percentile(arr, 1)),\n \"pixel_p50\": float(np.percentile(arr, 50)),\n \"pixel_p99\": float(np.percentile(arr, 99))\n }\n\n except Exception:\n\n return {\n \"pixel_loaded\": False,\n \"pixel_min\": np.nan,\n \"pixel_max\": np.nan,\n \"pixel_mean\": np.nan,\n \"pixel_std\": np.nan,\n \"pixel_p01\": np.nan,\n \"pixel_p50\": np.nan,\n \"pixel_p99\": np.nan\n }\n\n\n# ------------------------------------------------------------\n# 4. Audit function\n# ------------------------------------------------------------\n\naudit_rows = []\n\nloaded_count = 0\nfailed_count = 0\nmissing_pixel_count = 0\n\n\ndef audit_series(\n root,\n study_uid,\n series_uid,\n source,\n metadata_row=None\n):\n\n global loaded_count\n global failed_count\n global missing_pixel_count\n\n series_dir = get_series_directory(\n root,\n study_uid,\n series_uid\n )\n\n dcm_file = choose_middle_file(\n series_dir\n )\n\n if dcm_file is None:\n\n failed_count += 1\n\n return {\n \"source\": source,\n \"StudyInstanceUID\": study_uid,\n \"SeriesInstanceUID\": series_uid,\n \"file_path\": str(series_dir),\n \"read_status\": \"NO_DICOM_FILE\"\n }\n\n try:\n\n ds = pydicom.dcmread(\n str(dcm_file),\n force=False\n )\n\n loaded_count += 1\n\n except Exception as e:\n\n failed_count += 1\n\n return {\n \"source\": source,\n \"StudyInstanceUID\": study_uid,\n \"SeriesInstanceUID\": series_uid,\n \"file_path\": str(dcm_file),\n \"read_status\": \"READ_FAILED\",\n \"error\": str(e)[:300]\n }\n\n\n if not hasattr(ds, \"PixelData\"):\n missing_pixel_count += 1\n\n\n row = {\n \"source\": source,\n\n \"StudyInstanceUID\": str(\n study_uid\n ),\n\n \"SeriesInstanceUID\": str(\n series_uid\n ),\n\n \"file_path\": str(\n dcm_file\n ),\n\n \"read_status\": \"OK\",\n\n \"Modality\": safe_value(\n ds,\n \"Modality\"\n ),\n\n \"SOPClassUID\": safe_value(\n ds,\n \"SOPClassUID\"\n ),\n\n \"TransferSyntaxUID\": safe_value(\n getattr(\n ds,\n \"file_meta\",\n None\n ),\n \"TransferSyntaxUID\"\n ) if hasattr(\n ds,\n \"file_meta\"\n ) else None,\n\n \"Rows\": safe_value(\n ds,\n \"Rows\"\n ),\n\n \"Columns\": safe_value(\n ds,\n \"Columns\"\n ),\n\n \"SamplesPerPixel\": safe_value(\n ds,\n \"SamplesPerPixel\"\n ),\n\n \"PhotometricInterpretation\": safe_value(\n ds,\n \"PhotometricInterpretation\"\n ),\n\n \"BitsAllocated\": safe_value(\n ds,\n \"BitsAllocated\"\n ),\n\n \"BitsStored\": safe_value(\n ds,\n \"BitsStored\"\n ),\n\n \"HighBit\": safe_value(\n ds,\n \"HighBit\"\n ),\n\n \"PixelRepresentation\": safe_value(\n ds,\n \"PixelRepresentation\"\n ),\n\n \"PixelSpacing\": safe_value(\n ds,\n \"PixelSpacing\"\n ),\n\n \"SliceThickness\": safe_value(\n ds,\n \"SliceThickness\"\n ),\n\n \"SpacingBetweenSlices\": safe_value(\n ds,\n \"SpacingBetweenSlices\"\n ),\n\n \"ImageOrientationPatient\": safe_value(\n ds,\n \"ImageOrientationPatient\"\n ),\n\n \"ImagePositionPatient\": safe_value(\n ds,\n \"ImagePositionPatient\"\n ),\n\n \"InstanceNumber\": safe_value(\n ds,\n \"InstanceNumber\"\n ),\n\n \"SeriesDescription\": safe_value(\n ds,\n \"SeriesDescription\"\n ),\n\n \"ProtocolName\": safe_value(\n ds,\n \"ProtocolName\"\n ),\n\n \"Manufacturer\": safe_value(\n ds,\n \"Manufacturer\"\n ),\n\n \"ManufacturerModelName\": safe_value(\n ds,\n \"ManufacturerModelName\"\n ),\n\n \"RescaleSlope\": safe_value(\n ds,\n \"RescaleSlope\"\n ),\n\n \"RescaleIntercept\": safe_value(\n ds,\n \"RescaleIntercept\"\n )\n }\n\n\n # --------------------------------------------------------\n # Pixel statistics\n # --------------------------------------------------------\n\n pixel_stats = get_pixel_statistics(\n ds\n )\n\n row.update(\n pixel_stats\n )\n\n\n # --------------------------------------------------------\n # CSV metadata\n # --------------------------------------------------------\n\n if metadata_row is not None:\n\n row[\"csv_Fluid_Sensitive\"] = (\n metadata_row.get(\n \"Fluid_Sensitive\",\n np.nan\n )\n )\n\n row[\"csv_Fat_Suppression\"] = (\n metadata_row.get(\n \"Fat_Suppression\",\n np.nan\n )\n )\n\n row[\"csv_Anatomical_Plane\"] = (\n metadata_row.get(\n \"Anatomical_Plane\",\n None\n )\n )\n\n row[\"csv_slice_count\"] = (\n metadata_row.get(\n \"slice_count\",\n np.nan\n )\n )\n\n\n return row\n\n\n# ------------------------------------------------------------\n# 5. Audit training sample\n# ------------------------------------------------------------\n\nprint(\"\\nAuditing sampled training series...\")\n\nfor i, (_, r) in enumerate(\n train_sample.iterrows(),\n start=1\n):\n\n row = audit_series(\n TRAIN_SERIES_DIR,\n r[\"StudyInstanceUID\"],\n r[\"SeriesInstanceUID\"],\n \"train\",\n r\n )\n\n audit_rows.append(\n row\n )\n\n if i % 25 == 0:\n print(\n f\"Processed {i} / \"\n f\"{len(train_sample)} train series\"\n )\n\n\n# ------------------------------------------------------------\n# 6. Audit ALL test series\n# ------------------------------------------------------------\n\nprint(\"\\nAuditing ALL test series...\")\n\nfor i, (_, r) in enumerate(\n test_series.iterrows(),\n start=1\n):\n\n row = audit_series(\n TEST_SERIES_DIR,\n r[\"StudyInstanceUID\"],\n r[\"SeriesInstanceUID\"],\n \"test\",\n r\n )\n\n audit_rows.append(\n row\n )\n\n print(\n f\"Processed test series \"\n f\"{i} / {len(test_series)}\"\n )\n\n\n# ------------------------------------------------------------\n# 7. Save audit\n# ------------------------------------------------------------\n\ndicom_audit = pd.DataFrame(\n audit_rows\n)\n\ndicom_audit.to_csv(\n OUTPUT_FILE,\n index=False\n)\n\n\n# ------------------------------------------------------------\n# 8. Summary\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"DICOM AUDIT SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n \"Rows audited:\",\n len(dicom_audit)\n)\n\nprint(\n \"Successfully read:\",\n loaded_count\n)\n\nprint(\n \"Failed:\",\n failed_count\n)\n\nprint(\n \"Missing PixelData:\",\n missing_pixel_count\n)\n\n\nprint(\"\\nRead status:\")\n\nprint(\n dicom_audit[\n \"read_status\"\n ].value_counts(\n dropna=False\n )\n)\n\n\nprint(\"\\nDimensions:\")\n\nprint(\n dicom_audit[\n [\n \"Rows\",\n \"Columns\"\n ]\n ].value_counts(\n dropna=False\n ).head(20)\n)\n\n\nprint(\"\\nPhotometric interpretation:\")\n\nprint(\n dicom_audit[\n \"PhotometricInterpretation\"\n ].value_counts(\n dropna=False\n )\n)\n\n\nprint(\"\\nTransfer Syntax:\")\n\nprint(\n dicom_audit[\n \"TransferSyntaxUID\"\n ].value_counts(\n dropna=False\n ).head(20)\n)\n\n\nprint(\"\\nPixel representation:\")\n\nprint(\n dicom_audit[\n [\n \"BitsAllocated\",\n \"BitsStored\",\n \"PixelRepresentation\"\n ]\n ].value_counts(\n dropna=False\n )\n)\n\n\nprint(\"\\nCSV plane distribution in audited sample:\")\n\nif \"csv_Anatomical_Plane\" in dicom_audit.columns:\n\n print(\n dicom_audit[\n \"csv_Anatomical_Plane\"\n ].value_counts(\n dropna=False\n )\n )\n\n\nprint(\"\\nPixel intensity summary:\")\n\nprint(\n dicom_audit[\n [\n \"pixel_min\",\n \"pixel_max\",\n \"pixel_mean\",\n \"pixel_std\",\n \"pixel_p01\",\n \"pixel_p50\",\n \"pixel_p99\"\n ]\n ].describe()\n)\n\n\nprint(\"\\nMissing metadata:\")\n\nmetadata_fields = [\n \"Rows\",\n \"Columns\",\n \"PixelSpacing\",\n \"SliceThickness\",\n \"SpacingBetweenSlices\",\n \"ImageOrientationPatient\",\n \"ImagePositionPatient\",\n \"InstanceNumber\",\n \"PhotometricInterpretation\",\n \"BitsAllocated\",\n \"BitsStored\",\n \"PixelRepresentation\",\n \"TransferSyntaxUID\",\n \"SeriesDescription\"\n]\n\nmissing_summary = []\n\nfor field in metadata_fields:\n\n missing = (\n dicom_audit[field]\n .isna()\n |\n (\n dicom_audit[field]\n .astype(str)\n .str.strip()\n .isin(\n [\n \"\",\n \"None\",\n \"nan\"\n ]\n )\n )\n ).sum()\n\n missing_summary.append({\n \"field\": field,\n \"missing_count\": int(missing),\n \"missing_percentage\": (\n 100 * missing /\n len(dicom_audit)\n )\n })\n\nmissing_df = pd.DataFrame(\n missing_summary\n)\n\nprint(\n missing_df.to_string(\n index=False\n )\n)\n\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 20 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"\\nSaved:\")\nprint(OUTPUT_FILE)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:30:51.62638Z","iopub.execute_input":"2026-08-11T08:30:51.62676Z","iopub.status.idle":"2026-08-11T08:30:58.100067Z","shell.execute_reply.started":"2026-08-11T08:30:51.626722Z","shell.execute_reply":"2026-08-11T08:30:58.098967Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 21: REPORT SEMANTIC + TERMINOLOGY AUDIT\n# ============================================================\n\nimport re\nimport unicodedata\nimport numpy as np\nimport pandas as pd\nfrom pathlib import Path\nfrom collections import Counter\n\nprint(\"=\" * 70)\nprint(\"REPORT SEMANTIC + TERMINOLOGY AUDIT\")\nprint(\"=\" * 70)\n\n# ------------------------------------------------------------\n# 1. Configuration\n# ------------------------------------------------------------\n\nTRAIN_CSV = (\n \"/kaggle/input/competitions/rsna-knee-abnormality-detection/train.csv\"\n)\n\nAUDIT_DIR = Path(\n \"/kaggle/working/rsna_knee_audit\"\n)\n\nAUDIT_DIR.mkdir(\n parents=True,\n exist_ok=True\n)\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n]\n\n# ------------------------------------------------------------\n# 2. Load train.csv independently\n# ------------------------------------------------------------\n\ntrain = pd.read_csv(\n TRAIN_CSV\n)\n\nprint(\"\\nTrain shape:\", train.shape)\n\nprint(\n \"Reports available:\",\n train[\"Report\"].notna().sum()\n)\n\nprint(\n \"Missing reports:\",\n train[\"Report\"].isna().sum()\n)\n\n\n# ------------------------------------------------------------\n# 3. Basic report statistics\n# ------------------------------------------------------------\n\ntrain[\"Report\"] = (\n train[\"Report\"]\n .fillna(\"\")\n .astype(str)\n)\n\ntrain[\"report_chars\"] = (\n train[\"Report\"]\n .str.len()\n)\n\ntrain[\"report_words\"] = (\n train[\"Report\"]\n .str.split()\n .str.len()\n)\n\ntrain[\"report_lines\"] = (\n train[\"Report\"]\n .str.count(\"\\n\") + 1\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"REPORT LENGTH STATISTICS\")\nprint(\"=\" * 70)\n\nprint(\n train[\n [\n \"report_chars\",\n \"report_words\",\n \"report_lines\"\n ]\n ].describe()\n)\n\n\n# ------------------------------------------------------------\n# 4. Script detection\n# ------------------------------------------------------------\n\ndef detect_script(text):\n\n counts = Counter()\n\n for char in text:\n\n if char.isspace():\n continue\n\n name = unicodedata.name(\n char,\n \"\"\n )\n\n if \"LATIN\" in name:\n counts[\"LATIN\"] += 1\n\n elif \"GREEK\" in name:\n counts[\"GREEK\"] += 1\n\n elif \"CYRILLIC\" in name:\n counts[\"CYRILLIC\"] += 1\n\n elif \"ARABIC\" in name:\n counts[\"ARABIC\"] += 1\n\n elif \"HEBREW\" in name:\n counts[\"HEBREW\"] += 1\n\n elif \"DEVANAGARI\" in name:\n counts[\"DEVANAGARI\"] += 1\n\n else:\n counts[\"OTHER\"] += 1\n\n if not counts:\n return \"EMPTY\"\n\n return counts.most_common(1)[0][0]\n\n\ntrain[\"dominant_script\"] = (\n train[\"Report\"]\n .apply(detect_script)\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"DOMINANT REPORT SCRIPT\")\nprint(\"=\" * 70)\n\nprint(\n train[\"dominant_script\"]\n .value_counts()\n)\n\n\n# ------------------------------------------------------------\n# 5. Target terminology\n# ------------------------------------------------------------\n\n# These are deliberately broad candidate terms.\n# They are NOT labels.\n# Cell 22 will validate them against the 58 labeled studies.\n\nTERM_PATTERNS = {\n\n \"ACL\": [\n r\"\\bacl\\b\",\n r\"anterior cruciate\",\n r\"ligament croisé antérieur\",\n r\"ligamento cruzado anterior\",\n r\"vorderes kreuzband\"\n ],\n\n \"MCL\": [\n r\"\\bmcl\\b\",\n r\"medial collateral\",\n r\"medial collateral ligament\",\n r\"ligament collatéral médial\",\n r\"ligamento colateral medial\"\n ],\n\n \"Medial Meniscus\": [\n r\"medial meniscus\",\n r\"medial meniscal\",\n r\"menisque médial\",\n r\"menisco medial\",\n r\"innenmeniskus\"\n ],\n\n \"Lateral Meniscus\": [\n r\"lateral meniscus\",\n r\"lateral meniscal\",\n r\"menisque latéral\",\n r\"menisco lateral\",\n r\"außenmeniskus\"\n ],\n\n \"Medial OA\": [\n r\"medial compartment\",\n r\"medial femorotibial\",\n r\"medial osteoarthritis\",\n r\"medial arthrosis\",\n r\"medial gonarthrosis\",\n r\"medial joint space narrowing\"\n ],\n\n \"Lateral OA\": [\n r\"lateral compartment\",\n r\"lateral femorotibial\",\n r\"lateral osteoarthritis\",\n r\"lateral arthrosis\",\n r\"lateral gonarthrosis\",\n r\"lateral joint space narrowing\"\n ],\n\n \"PF OA\": [\n r\"patellofemoral\",\n r\"patello-femoral\",\n r\"patellofemoral osteoarthritis\",\n r\"patellofemoral arthrosis\",\n r\"retropatellar\",\n r\"patellar cartilage\"\n ],\n\n \"Effusion\": [\n r\"\\beffusion\\b\",\n r\"joint effusion\",\n r\"articular effusion\",\n r\"épanchement\",\n r\"gelenkerguss\"\n ],\n\n \"Synovitis\": [\n r\"\\bsynovitis\\b\",\n r\"synovial\",\n r\"synoviale\",\n r\"synoviale verdickung\"\n ],\n\n \"Baker's\": [\n r\"baker\",\n r\"popliteal cyst\",\n r\"popliteal cyste\",\n r\"cyste poplitée\",\n r\"baker cyst\"\n ],\n\n \"Contusion\": [\n r\"\\bcontusion\\b\",\n r\"bone bruise\",\n r\"bone marrow edema\",\n r\"marrow edema\",\n r\"œdème osseux\",\n r\"knochenödem\"\n ],\n\n \"Fracture\": [\n r\"\\bfracture\\b\",\n r\"\\bfractures\\b\",\n r\"fracture line\",\n r\"fracture du\",\n r\"fraktur\",\n r\"fractura\"\n ]\n}\n\n\n# ------------------------------------------------------------\n# 6. Compile patterns\n# ------------------------------------------------------------\n\ncompiled_patterns = {}\n\nfor target, patterns in TERM_PATTERNS.items():\n\n compiled_patterns[target] = [\n re.compile(\n pattern,\n flags=re.IGNORECASE\n )\n for pattern in patterns\n ]\n\n\n# ------------------------------------------------------------\n# 7. Negation / uncertainty patterns\n# ------------------------------------------------------------\n\nNEGATION_PATTERNS = [\n\n r\"\\bno\\b\",\n r\"\\bwithout\\b\",\n r\"\\bnot\\b\",\n r\"\\bnegative\\b\",\n r\"\\bnormal\\b\",\n r\"\\bintact\\b\",\n r\"\\bpreserved\\b\",\n r\"\\babsence\\b\",\n r\"\\babsent\\b\",\n\n r\"\\baucune\\b\",\n r\"\\baucun\\b\",\n r\"\\bsans\\b\",\n r\"\\bnormal\\b\",\n r\"\\bintact\\b\",\n\n r\"\\bkein\\b\",\n r\"\\bkeine\\b\",\n r\"\\bkeinen\\b\",\n r\"\\bintakt\\b\",\n\n r\"\\bno hay\\b\",\n r\"\\bsin\\b\",\n r\"\\bnormal\\b\",\n r\"\\bíntegro\\b\",\n r\"\\bintegro\\b\"\n]\n\nUNCERTAINTY_PATTERNS = [\n\n r\"\\bpossible\\b\",\n r\"\\bpossibly\\b\",\n r\"\\bprobable\\b\",\n r\"\\bprobable\\b\",\n r\"\\bsuspect\\b\",\n r\"\\bsuspicious\\b\",\n r\"\\bcannot exclude\\b\",\n r\"\\bmay represent\\b\",\n r\"\\bmay be\\b\",\n\n r\"\\bpossible\\b\",\n r\"\\bsuspect\\b\",\n r\"\\bà confirmer\\b\",\n r\"\\bne peut être exclu\\b\",\n\n r\"\\bmöglich\\b\",\n r\"\\bverdacht\\b\",\n\n r\"\\bposible\\b\",\n r\"\\bprobable\\b\",\n r\"\\bno se puede excluir\\b\"\n]\n\ncompiled_negation = [\n re.compile(\n p,\n flags=re.IGNORECASE\n )\n for p in NEGATION_PATTERNS\n]\n\ncompiled_uncertainty = [\n re.compile(\n p,\n flags=re.IGNORECASE\n )\n for p in UNCERTAINTY_PATTERNS\n]\n\n\n# ------------------------------------------------------------\n# 8. Find target terminology\n# ------------------------------------------------------------\n\ndef find_matches(text, patterns):\n\n matches = []\n\n for pattern in patterns:\n\n for match in pattern.finditer(text):\n\n start = max(\n 0,\n match.start() - 80\n )\n\n end = min(\n len(text),\n match.end() + 80\n )\n\n context = text[\n start:end\n ].replace(\n \"\\n\",\n \" \"\n )\n\n matches.append({\n \"term\": match.group(),\n \"context\": context\n })\n\n return matches\n\n\n# ------------------------------------------------------------\n# 9. Build terminology audit\n# ------------------------------------------------------------\n\naudit_rows = []\n\nfor _, row in train.iterrows():\n\n text = row[\"Report\"]\n\n for target in TARGETS:\n\n matches = find_matches(\n text,\n compiled_patterns[target]\n )\n\n if matches:\n\n context_text = \" || \".join(\n m[\"context\"]\n for m in matches[:5]\n )\n\n negation_found = any(\n pattern.search(\n context_text\n )\n for pattern in compiled_negation\n )\n\n uncertainty_found = any(\n pattern.search(\n context_text\n )\n for pattern in compiled_uncertainty\n )\n\n audit_rows.append({\n\n \"StudyInstanceUID\":\n row[\"StudyInstanceUID\"],\n\n \"target\":\n target,\n\n \"mention_count\":\n len(matches),\n\n \"matched_terms\":\n \" | \".join(\n sorted(\n set(\n m[\"term\"]\n for m in matches\n )\n )\n ),\n\n \"negation_context_found\":\n negation_found,\n\n \"uncertainty_context_found\":\n uncertainty_found,\n\n \"context\":\n context_text\n })\n\n\nreport_term_audit = pd.DataFrame(\n audit_rows\n)\n\n\n# ------------------------------------------------------------\n# 10. Summary by target\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TARGET TERMINOLOGY SUMMARY\")\nprint(\"=\" * 70)\n\nif len(report_term_audit) > 0:\n\n target_summary = (\n report_term_audit\n .groupby(\"target\")\n .agg(\n studies_with_mentions=(\n \"StudyInstanceUID\",\n \"nunique\"\n ),\n total_mentions=(\n \"mention_count\",\n \"sum\"\n ),\n negation_contexts=(\n \"negation_context_found\",\n \"sum\"\n ),\n uncertainty_contexts=(\n \"uncertainty_context_found\",\n \"sum\"\n )\n )\n .reindex(TARGETS)\n .fillna(0)\n )\n\n print(\n target_summary\n )\n\nelse:\n\n target_summary = pd.DataFrame()\n\n\n# ------------------------------------------------------------\n# 11. Explicitly labeled subset\n# ------------------------------------------------------------\n\nlabel_mask = train[\n TARGETS\n].notna().all(axis=1)\n\nlabeled_train = train.loc[\n label_mask\n].copy()\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"EXPLICITLY LABELED REPORT AUDIT\")\nprint(\"=\" * 70)\n\nprint(\n \"Explicitly labeled studies:\",\n len(labeled_train)\n)\n\nprint(\n \"Labeled studies with reports:\",\n labeled_train[\"Report\"]\n .notna()\n .sum()\n)\n\n\n# ------------------------------------------------------------\n# 12. Display representative reports\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"REPRESENTATIVE REPORTS FROM LABELED STUDIES\")\nprint(\"=\" * 70)\n\ndisplay_columns = [\n \"StudyInstanceUID\",\n \"Report\"\n] + TARGETS\n\nsample_labeled = labeled_train.sample(\n n=min(10, len(labeled_train)),\n random_state=42\n)\n\nfor i, (_, row) in enumerate(\n sample_labeled.iterrows(),\n start=1\n):\n\n print(\"\\n\" + \"-\" * 70)\n print(\n f\"Labeled report {i}\"\n )\n print(\n \"StudyInstanceUID:\",\n row[\"StudyInstanceUID\"]\n )\n\n print(\"\\nLabels:\")\n\n for target in TARGETS:\n\n print(\n f\" {target}:\",\n int(row[target])\n )\n\n print(\"\\nReport:\")\n print(\n row[\"Report\"]\n )\n\n\n# ------------------------------------------------------------\n# 13. Save outputs\n# ------------------------------------------------------------\n\nterm_file = (\n AUDIT_DIR /\n \"report_terminology_audit.csv\"\n)\n\nsummary_file = (\n AUDIT_DIR /\n \"report_terminology_summary.csv\"\n)\n\nreport_term_audit.to_csv(\n term_file,\n index=False\n)\n\ntarget_summary.to_csv(\n summary_file\n)\n\n\n# ------------------------------------------------------------\n# 14. Final output\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 21 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"\\nSaved:\")\nprint(term_file)\nprint(summary_file)\n\nprint(\"\\nTotal report-target mention records:\")\nprint(len(report_term_audit))\n\nprint(\"\\nImportant:\")\nprint(\n \"These terminology matches are CANDIDATE EVIDENCE only.\"\n)\n\nprint(\n \"They must NOT be treated as ground-truth labels yet.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:33:25.955998Z","iopub.execute_input":"2026-08-11T08:33:25.956335Z","iopub.status.idle":"2026-08-11T08:33:34.59629Z","shell.execute_reply.started":"2026-08-11T08:33:25.956306Z","shell.execute_reply":"2026-08-11T08:33:34.595402Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# CELL 22: REPORT ↔ LABEL CONSISTENCY AUDIT\n# ============================================================\n\nimport re\nimport numpy as np\nimport pandas as pd\nfrom pathlib import Path\n\nprint(\"=\" * 70)\nprint(\"REPORT ↔ LABEL CONSISTENCY AUDIT\")\nprint(\"=\" * 70)\n\n# ------------------------------------------------------------\n# 1. Configuration\n# ------------------------------------------------------------\n\nTRAIN_CSV = (\n \"/kaggle/input/competitions/rsna-knee-abnormality-detection/train.csv\"\n)\n\nAUDIT_DIR = Path(\n \"/kaggle/working/rsna_knee_audit\"\n)\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n]\n\n# ------------------------------------------------------------\n# 2. Load data\n# ------------------------------------------------------------\n\ntrain = pd.read_csv(\n TRAIN_CSV\n)\n\ntrain[\"Report\"] = (\n train[\"Report\"]\n .fillna(\"\")\n .astype(str)\n)\n\n# Explicitly labeled studies\nlabeled = train[\n train[TARGETS].notna().all(axis=1)\n].copy()\n\nprint(\n \"\\nExplicitly labeled studies:\",\n len(labeled)\n)\n\n# ------------------------------------------------------------\n# 3. Target-specific positive evidence\n# ------------------------------------------------------------\n\nPOSITIVE_PATTERNS = {\n\n \"ACL\": [\n r\"acl.*tear\",\n r\"tear.*acl\",\n r\"acl.*ruptur\",\n r\"ruptur.*acl\",\n r\"acl.*injur\",\n r\"acl.*sprain\",\n r\"anterior cruciate ligament.*tear\",\n r\"anterior cruciate ligament.*ruptur\",\n r\"anterior cruciate ligament.*injur\",\n r\"ligamento cruzado anterior.*rotura\",\n r\"ligament.*croisé antérieur.*rupture\"\n ],\n\n \"MCL\": [\n r\"mcl.*tear\",\n r\"mcl.*injur\",\n r\"mcl.*sprain\",\n r\"medial collateral ligament.*tear\",\n r\"medial collateral ligament.*injur\",\n r\"medial collateral ligament.*sprain\",\n r\"ligamento colateral medial.*esguince\",\n r\"ligamento colateral medial.*rotura\",\n r\"ligamento colateral medial.*lesion\",\n r\"ligament.*collatéral médial.*lésion\"\n ],\n\n \"Medial Meniscus\": [\n r\"medial meniscus.*tear\",\n r\"medial meniscal.*tear\",\n r\"medial meniscus.*ruptur\",\n r\"medial meniscus.*lesion\",\n r\"meniscus.*medial.*tear\",\n r\"menisco medial.*rotura\",\n r\"menisco medial.*desgarro\",\n r\"menisque médial.*rupture\",\n r\"innenmeniskus.*riss\"\n ],\n\n \"Lateral Meniscus\": [\n r\"lateral meniscus.*tear\",\n r\"lateral meniscal.*tear\",\n r\"lateral meniscus.*ruptur\",\n r\"lateral meniscus.*lesion\",\n r\"meniscus.*lateral.*tear\",\n r\"menisco lateral.*rotura\",\n r\"menisco lateral.*desgarro\",\n r\"menisque latéral.*rupture\",\n r\"außenmeniskus.*riss\"\n ],\n\n \"Medial OA\": [\n r\"medial compartment.*osteoarthritis\",\n r\"medial compartment.*arthrosis\",\n r\"medial compartment.*chondrosis\",\n r\"medial compartment.*chondral\",\n r\"medial.*joint space narrowing\",\n r\"medial.*cartilage loss\",\n r\"medial.*osteophyte\",\n r\"medial tibiofemoral.*osteoarthritis\",\n r\"medial femorotibial.*osteoarthritis\",\n r\"osteoarthritis.*medial compartment\",\n r\"arthrosis.*medial compartment\"\n ],\n\n \"Lateral OA\": [\n r\"lateral compartment.*osteoarthritis\",\n r\"lateral compartment.*arthrosis\",\n r\"lateral compartment.*chondrosis\",\n r\"lateral compartment.*chondral\",\n r\"lateral.*joint space narrowing\",\n r\"lateral.*cartilage loss\",\n r\"lateral.*osteophyte\",\n r\"lateral tibiofemoral.*osteoarthritis\",\n r\"lateral femorotibial.*osteoarthritis\",\n r\"osteoarthritis.*lateral compartment\",\n r\"arthrosis.*lateral compartment\"\n ],\n\n \"PF OA\": [\n r\"patellofemoral.*osteoarthritis\",\n r\"patellofemoral.*arthrosis\",\n r\"patellofemoral.*chondrosis\",\n r\"patellofemoral.*chondral\",\n r\"patellofemoral.*cartilage loss\",\n r\"patellofemoral.*cartilage.*damage\",\n r\"patellar.*cartilage loss\",\n r\"patellar.*chondrosis\",\n r\"retropatellar.*chondrosis\",\n r\"patellofemoral compartment.*degener\"\n ],\n\n \"Effusion\": [\n r\"\\beffusion\\b\",\n r\"joint effusion\",\n r\"knee effusion\",\n r\"articular effusion\",\n r\"joint fluid\",\n r\"fluid.*joint\",\n r\"derrame articular\",\n r\"derrame de la articulación\",\n r\"épanchement articulaire\",\n r\"gelenkerguss\"\n ],\n\n \"Synovitis\": [\n r\"\\bsynovitis\\b\",\n r\"synovial.*thickening\",\n r\"synovial.*proliferation\",\n r\"synoviale.*verdick\",\n r\"sinovitis\"\n ],\n\n \"Baker's\": [\n r\"baker.?s cyst\",\n r\"baker cyst\",\n r\"popliteal cyst\",\n r\"popliteal cyste\",\n r\"cyste poplitée\",\n r\"quiste poplíteo\",\n r\"baker.*kyst\"\n ],\n\n \"Contusion\": [\n r\"bone contusion\",\n r\"bone bruise\",\n r\"bone marrow edema\",\n r\"marrow edema\",\n r\"osteochondral impaction injury\",\n r\"contusion\",\n r\"contusion.*bone\",\n r\"bone.*contusion\",\n r\"œdème osseux\",\n r\"edema.*médula ósea\"\n ],\n\n \"Fracture\": [\n r\"\\bfracture\\b\",\n r\"\\bfractures\\b\",\n r\"fracture line\",\n r\"fracture.*bone\",\n r\"insufficiency fracture\",\n r\"stress fracture\",\n r\"microfracture\",\n r\"microtrabecular fracture\",\n r\"fraktur\",\n r\"fractura\"\n ]\n}\n\n\n# ------------------------------------------------------------\n# 4. Target-specific negative evidence\n# ------------------------------------------------------------\n\nNEGATIVE_PATTERNS = {\n\n \"ACL\": [\n r\"acl.*normal\",\n r\"acl.*intact\",\n r\"acl.*preserved\",\n r\"acl.*no tear\",\n r\"acl.*without tear\",\n r\"no.*acl.*tear\",\n r\"anterior cruciate ligament.*normal\",\n r\"anterior cruciate ligament.*intact\",\n r\"anterior cruciate ligament.*no tear\"\n ],\n\n \"MCL\": [\n r\"mcl.*normal\",\n r\"mcl.*intact\",\n r\"mcl.*preserved\",\n r\"mcl.*no tear\",\n r\"no.*mcl.*tear\",\n r\"medial collateral ligament.*normal\",\n r\"medial collateral ligament.*intact\",\n r\"medial collateral ligament.*no tear\"\n ],\n\n \"Medial Meniscus\": [\n r\"medial meniscus.*normal\",\n r\"medial meniscus.*intact\",\n r\"medial meniscus.*no tear\",\n r\"medial meniscus.*not torn\",\n r\"medial meniscus.*without tear\",\n r\"meniscus.*medial.*normal\",\n r\"menisco medial.*normal\",\n r\"menisco medial.*sin.*rotura\"\n ],\n\n \"Lateral Meniscus\": [\n r\"lateral meniscus.*normal\",\n r\"lateral meniscus.*intact\",\n r\"lateral meniscus.*no tear\",\n r\"lateral meniscus.*not torn\",\n r\"lateral meniscus.*without tear\",\n r\"meniscus.*lateral.*normal\",\n r\"menisco lateral.*normal\"\n ],\n\n \"Medial OA\": [\n r\"medial compartment.*normal\",\n r\"medial compartment.*intact\",\n r\"medial compartment.*no.*chondrosis\",\n r\"medial compartment.*no.*chondral\",\n r\"medial compartment.*cartilage.*intact\",\n r\"medial compartment.*without.*chondral\"\n ],\n\n \"Lateral OA\": [\n r\"lateral compartment.*normal\",\n r\"lateral compartment.*intact\",\n r\"lateral compartment.*no.*chondrosis\",\n r\"lateral compartment.*no.*chondral\",\n r\"lateral compartment.*cartilage.*intact\",\n r\"lateral compartment.*without.*chondral\"\n ],\n\n \"PF OA\": [\n r\"patellofemoral.*normal\",\n r\"patellofemoral.*intact\",\n r\"patellofemoral.*without.*chondral\",\n r\"patellofemoral.*no.*chondral\",\n r\"patellar.*cartilage.*normal\",\n r\"patellar.*cartilage.*intact\"\n ],\n\n \"Effusion\": [\n r\"no joint effusion\",\n r\"no knee effusion\",\n r\"no articular effusion\",\n r\"without.*effusion\",\n r\"no.*joint fluid\",\n r\"no.*fluid.*joint\",\n r\"no hay.*derrame\",\n r\"sin.*derrame\",\n r\"no.*épanchement\"\n ],\n\n \"Synovitis\": [\n r\"no synovitis\",\n r\"without synovitis\",\n r\"no.*synovial.*thickening\",\n r\"sinovitis.*no\",\n r\"no evidence.*synovitis\"\n ],\n\n \"Baker's\": [\n r\"no baker.?s cyst\",\n r\"no popliteal cyst\",\n r\"without.*baker.?s cyst\",\n r\"no.*popliteal cyst\",\n r\"no hay.*quiste poplíteo\",\n r\"sin.*quiste poplíteo\"\n ],\n\n \"Contusion\": [\n r\"no bone contusion\",\n r\"no bone bruise\",\n r\"no marrow edema\",\n r\"without.*bone contusion\",\n r\"without.*bone bruise\",\n r\"no.*osseous.*contusion\",\n r\"no hay.*contusion\"\n ],\n\n \"Fracture\": [\n r\"no fracture\",\n r\"no fractures\",\n r\"without fracture\",\n r\"without.*fracture\",\n r\"no acute fracture\",\n r\"no evidence.*fracture\",\n r\"sin fractura\",\n r\"sin.*fractura\",\n r\"keine fraktur\"\n ]\n}\n\n\n# ------------------------------------------------------------\n# 5. Compile regex patterns\n# ------------------------------------------------------------\n\ncompiled_positive = {\n target: [\n re.compile(\n pattern,\n flags=re.IGNORECASE\n )\n for pattern in patterns\n ]\n for target, patterns in POSITIVE_PATTERNS.items()\n}\n\ncompiled_negative = {\n target: [\n re.compile(\n pattern,\n flags=re.IGNORECASE\n )\n for pattern in patterns\n ]\n for target, patterns in NEGATIVE_PATTERNS.items()\n}\n\n\n# ------------------------------------------------------------\n# 6. Sentence splitting\n# ------------------------------------------------------------\n\ndef split_sentences(text):\n\n text = re.sub(\n r\"\\s+\",\n \" \",\n text\n ).strip()\n\n if not text:\n return []\n\n return [\n sentence.strip()\n for sentence in re.split(\n r\"(?<=[.!?])\\s+\",\n text\n )\n if sentence.strip()\n ]\n\n\n# ------------------------------------------------------------\n# 7. Detect evidence\n# ------------------------------------------------------------\n\ndef detect_evidence(\n text,\n target\n):\n\n sentences = split_sentences(\n text\n )\n\n positive_hits = []\n negative_hits = []\n\n for sentence in sentences:\n\n for pattern in compiled_positive[target]:\n\n if pattern.search(sentence):\n\n positive_hits.append(\n sentence\n )\n break\n\n for pattern in compiled_negative[target]:\n\n if pattern.search(sentence):\n\n negative_hits.append(\n sentence\n )\n break\n\n return (\n positive_hits,\n negative_hits\n )\n\n\n# ------------------------------------------------------------\n# 8. Audit all 58 labeled studies\n# ------------------------------------------------------------\n\naudit_rows = []\n\nfor _, row in labeled.iterrows():\n\n text = row[\"Report\"]\n\n for target in TARGETS:\n\n label = int(\n row[target]\n )\n\n positive_hits, negative_hits = (\n detect_evidence(\n text,\n target\n )\n )\n\n has_positive = (\n len(positive_hits) > 0\n )\n\n has_negative = (\n len(negative_hits) > 0\n )\n\n if has_positive and has_negative:\n\n evidence_status = \"MIXED\"\n\n elif has_positive:\n\n evidence_status = \"POSITIVE_EVIDENCE\"\n\n elif has_negative:\n\n evidence_status = \"NEGATIVE_EVIDENCE\"\n\n else:\n\n evidence_status = \"NO_CLEAR_EVIDENCE\"\n\n\n # ----------------------------------------------------\n # Compare evidence with actual label\n # ----------------------------------------------------\n\n if label == 1:\n\n if has_positive:\n consistency = \"AGREEMENT_POSITIVE\"\n\n elif has_negative:\n consistency = \"CONFLICT_LABEL_1_REPORT_NEGATIVE\"\n\n else:\n consistency = \"UNCLEAR_LABEL_1\"\n\n else:\n\n if has_positive:\n consistency = \"CONFLICT_LABEL_0_REPORT_POSITIVE\"\n\n elif has_negative:\n consistency = \"AGREEMENT_NEGATIVE\"\n\n else:\n consistency = \"UNCLEAR_LABEL_0\"\n\n\n audit_rows.append({\n\n \"StudyInstanceUID\":\n row[\"StudyInstanceUID\"],\n\n \"target\":\n target,\n\n \"label\":\n label,\n\n \"positive_evidence\":\n has_positive,\n\n \"negative_evidence\":\n has_negative,\n\n \"evidence_status\":\n evidence_status,\n\n \"consistency\":\n consistency,\n\n \"positive_context\":\n \" || \".join(\n positive_hits[:3]\n ),\n\n \"negative_context\":\n \" || \".join(\n negative_hits[:3]\n )\n })\n\n\nconsistency_df = pd.DataFrame(\n audit_rows\n)\n\n\n# ------------------------------------------------------------\n# 9. Overall consistency summary\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OVERALL CONSISTENCY SUMMARY\")\nprint(\"=\" * 70)\n\noverall_summary = (\n consistency_df[\n \"consistency\"\n ]\n .value_counts()\n)\n\nprint(\n overall_summary\n)\n\n\n# ------------------------------------------------------------\n# 10. Consistency by target\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CONSISTENCY BY TARGET\")\nprint(\"=\" * 70)\n\ntarget_consistency = pd.crosstab(\n consistency_df[\"target\"],\n consistency_df[\"consistency\"]\n)\n\ntarget_consistency = (\n target_consistency\n .reindex(TARGETS)\n .fillna(0)\n .astype(int)\n)\n\nprint(\n target_consistency\n)\n\n\n# ------------------------------------------------------------\n# 11. Calculate agreement metrics\n# ------------------------------------------------------------\n\nmetrics = []\n\nfor target in TARGETS:\n\n subset = consistency_df[\n consistency_df[\"target\"] == target\n ]\n\n label_1 = subset[\n subset[\"label\"] == 1\n ]\n\n label_0 = subset[\n subset[\"label\"] == 0\n ]\n\n tp_like = (\n label_1[\"positive_evidence\"]\n .sum()\n )\n\n fn_like = (\n (~label_1[\"positive_evidence\"])\n .sum()\n )\n\n fp_like = (\n label_0[\"positive_evidence\"]\n .sum()\n )\n\n tn_like = (\n (~label_0[\"positive_evidence\"])\n .sum()\n )\n\n total = len(subset)\n\n evidence_agreement = (\n (\n tp_like +\n tn_like\n ) / total\n if total > 0\n else np.nan\n )\n\n metrics.append({\n\n \"target\":\n target,\n\n \"n\":\n total,\n\n \"label_positive\":\n len(label_1),\n\n \"label_negative\":\n len(label_0),\n\n \"positive_evidence_when_label_1\":\n int(tp_like),\n\n \"positive_evidence_when_label_0\":\n int(fp_like),\n\n \"negative_or_no_positive_evidence_when_label_0\":\n int(tn_like),\n\n \"label_1_without_positive_evidence\":\n int(fn_like),\n\n \"positive_evidence_agreement\":\n evidence_agreement\n })\n\n\nmetrics_df = pd.DataFrame(\n metrics\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"REPORT EVIDENCE AGREEMENT METRICS\")\nprint(\"=\" * 70)\n\nprint(\n metrics_df.to_string(\n index=False\n )\n)\n\n\n# ------------------------------------------------------------\n# 12. Show conflicts\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"LABEL 0 / REPORT POSITIVE CONFLICTS\")\nprint(\"=\" * 70)\n\nconflict_0 = consistency_df[\n consistency_df[\"consistency\"]\n == \"CONFLICT_LABEL_0_REPORT_POSITIVE\"\n]\n\nprint(\n \"Count:\",\n len(conflict_0)\n)\n\nif len(conflict_0) > 0:\n\n display(\n conflict_0[\n [\n \"StudyInstanceUID\",\n \"target\",\n \"label\",\n \"positive_context\"\n ]\n ].head(30)\n )\n\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"LABEL 1 / REPORT NEGATIVE CONFLICTS\")\nprint(\"=\" * 70)\n\nconflict_1 = consistency_df[\n consistency_df[\"consistency\"]\n == \"CONFLICT_LABEL_1_REPORT_NEGATIVE\"\n]\n\nprint(\n \"Count:\",\n len(conflict_1)\n)\n\nif len(conflict_1) > 0:\n\n display(\n conflict_1[\n [\n \"StudyInstanceUID\",\n \"target\",\n \"label\",\n \"negative_context\"\n ]\n ].head(30)\n )\n\n\n# ------------------------------------------------------------\n# 13. Save outputs\n# ------------------------------------------------------------\n\nconsistency_file = (\n AUDIT_DIR /\n \"report_label_consistency.csv\"\n)\n\nmetrics_file = (\n AUDIT_DIR /\n \"report_label_consistency_summary.csv\"\n)\n\nconflict_file = (\n AUDIT_DIR /\n \"report_label_conflicts.csv\"\n)\n\nconsistency_df.to_csv(\n consistency_file,\n index=False\n)\n\nmetrics_df.to_csv(\n metrics_file,\n index=False\n)\n\npd.concat(\n [\n conflict_0,\n conflict_1\n ],\n ignore_index=True\n).to_csv(\n conflict_file,\n index=False\n)\n\n\n# ------------------------------------------------------------\n# 14. Final status\n# ------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 22 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"\\nSaved:\")\nprint(consistency_file)\nprint(metrics_file)\nprint(conflict_file)\n\nprint(\"\\nIMPORTANT:\")\nprint(\n \"This audit measures report-label agreement.\"\n)\n\nprint(\n \"It does NOT create pseudo-labels.\"\n)\n\nprint(\n \"Do not train on report-derived labels yet.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:36:10.929556Z","iopub.execute_input":"2026-08-11T08:36:10.929968Z","iopub.status.idle":"2026-08-11T08:36:11.444912Z","shell.execute_reply.started":"2026-08-11T08:36:10.929937Z","shell.execute_reply":"2026-08-11T08:36:11.444023Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 23 - IMPROVED REPORT EVIDENCE AUDIT\n# ================================================================\n\nimport os\nimport re\nimport numpy as np\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"IMPROVED REPORT EVIDENCE AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# 1. Make sure train and target definitions exist\n# ----------------------------------------------------------------\n\nif \"train\" not in globals():\n train_csv_path = \"/kaggle/input/competitions/rsna-knee-abnormality-detection/train.csv\"\n\n if not os.path.exists(train_csv_path):\n raise FileNotFoundError(\n f\"Could not find train.csv at:\\n{train_csv_path}\"\n )\n\n train = pd.read_csv(train_csv_path)\n\nif \"TARGETS\" not in globals():\n TARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n ]\n\nprint(f\"Train studies: {len(train)}\")\nprint(f\"Explicitly labeled studies: {train[TARGETS].notna().all(axis=1).sum()}\")\n\n\n# ----------------------------------------------------------------\n# 2. Conservative target aliases\n# ----------------------------------------------------------------\n\nTARGET_ALIASES = {\n \"ACL\": [\n r\"\\bacl\\b\",\n r\"anterior cruciate ligament\"\n ],\n\n \"MCL\": [\n r\"\\bmcl\\b\",\n r\"medial collateral ligament\"\n ],\n\n \"Medial Meniscus\": [\n r\"medial meniscus\",\n r\"meniscus medialis\"\n ],\n\n \"Lateral Meniscus\": [\n r\"lateral meniscus\",\n r\"meniscus lateralis\"\n ],\n\n \"Medial OA\": [\n r\"medial osteoarthritis\",\n r\"medial compartment.*(?:osteoarthritis|arthrosis|degenerative)\",\n r\"medial compartment.*(?:joint space narrowing|osteophyte)\"\n ],\n\n \"Lateral OA\": [\n r\"lateral osteoarthritis\",\n r\"lateral compartment.*(?:osteoarthritis|arthrosis|degenerative)\",\n r\"lateral compartment.*(?:joint space narrowing|osteophyte)\"\n ],\n\n \"PF OA\": [\n r\"patellofemoral osteoarthritis\",\n r\"patellofemoral.*(?:osteoarthritis|arthrosis|degenerative)\",\n r\"patellofemoral.*(?:joint space narrowing|osteophyte)\"\n ],\n\n \"Effusion\": [\n r\"\\beffusion\\b\",\n r\"joint effusion\"\n ],\n\n \"Synovitis\": [\n r\"\\bsynovitis\\b\",\n r\"synovial thickening\"\n ],\n\n \"Baker's\": [\n r\"baker.?s cyst\",\n r\"popliteal cyst\"\n ],\n\n \"Contusion\": [\n r\"\\bcontusion\\b\",\n r\"bone bruise\",\n r\"bone contusion\"\n ],\n\n \"Fracture\": [\n r\"\\bfracture\\b\",\n r\"fractured\"\n ]\n}\n\n\n# ----------------------------------------------------------------\n# 3. Explicit negative language\n# ----------------------------------------------------------------\n\nNEGATIVE_PATTERNS = [\n r\"\\bno\\b\",\n r\"\\bnot\\b\",\n r\"\\bwithout\\b\",\n r\"\\bnegative for\\b\",\n r\"\\bno evidence of\\b\",\n r\"\\bno sign of\\b\",\n r\"\\bno signs of\\b\",\n r\"\\bno acute\\b\",\n r\"\\bintact\\b\",\n r\"\\bnormal\\b\",\n r\"\\bunremarkable\\b\",\n r\"\\bpreserved\\b\",\n r\"\\bconserved\\b\",\n r\"\\bconservation\\b\",\n r\"\\babsence of\\b\",\n r\"\\babsent\\b\",\n r\"\\bnot seen\\b\",\n r\"\\bnone\\b\"\n]\n\n\n# ----------------------------------------------------------------\n# 4. Explicit positive/pathological language\n# ----------------------------------------------------------------\n\nPOSITIVE_PATTERNS = [\n r\"\\btear\\b\",\n r\"\\btear(?:ed|ing)?\\b\",\n r\"\\bru(p|pture|ptured)\\b\",\n r\"\\blésion\\b\",\n r\"\\blesion\\b\",\n r\"\\babnormal\\b\",\n r\"\\bpathologic\\b\",\n r\"\\bpathological\\b\",\n r\"\\bdegenerative\\b\",\n r\"\\bdegeneration\\b\",\n r\"\\bosteoarthritis\\b\",\n r\"\\barthrosis\\b\",\n r\"\\bosteoarthritic\\b\",\n r\"\\bswelling\\b\",\n r\"\\bfluid\\b\",\n r\"\\bthickening\\b\",\n r\"\\bcontusion\\b\",\n r\"\\bbruise\\b\",\n r\"\\bfracture\\b\",\n r\"\\bcyst\\b\",\n r\"\\beffusion\\b\",\n r\"\\bsynovitis\\b\",\n r\"\\bsignal abnormality\\b\",\n r\"\\bhyperintense\\b\",\n r\"\\bdisruption\\b\",\n r\"\\bdisrupted\\b\"\n]\n\n\n# ----------------------------------------------------------------\n# 5. Sentence splitter\n# ----------------------------------------------------------------\n\ndef split_sentences(text):\n if pd.isna(text):\n return []\n\n text = str(text)\n\n # Normalize line breaks\n text = re.sub(r\"[\\r\\n]+\", \" \", text)\n\n # Basic sentence segmentation\n sentences = re.split(\n r\"(?<=[.!?])\\s+|;\\s+|\\|\\|\",\n text\n )\n\n sentences = [\n s.strip()\n for s in sentences\n if s.strip()\n ]\n\n return sentences\n\n\n# ----------------------------------------------------------------\n# 6. Evidence classifier\n# ----------------------------------------------------------------\n\ndef classify_sentence(sentence, target):\n sentence_lower = sentence.lower()\n\n aliases = TARGET_ALIASES[target]\n\n target_found = any(\n re.search(pattern, sentence_lower, flags=re.IGNORECASE)\n for pattern in aliases\n )\n\n if not target_found:\n return \"NO_TARGET_MENTION\"\n\n # Check nearby negative context\n negative_found = any(\n re.search(pattern, sentence_lower, flags=re.IGNORECASE)\n for pattern in NEGATIVE_PATTERNS\n )\n\n # Check positive/pathological context\n positive_found = any(\n re.search(pattern, sentence_lower, flags=re.IGNORECASE)\n for pattern in POSITIVE_PATTERNS\n )\n\n # Strong negative constructions take precedence\n if negative_found:\n return \"NEGATIVE\"\n\n if positive_found:\n return \"POSITIVE\"\n\n return \"UNCERTAIN\"\n\n\n# ----------------------------------------------------------------\n# 7. Audit explicitly labeled studies\n# ----------------------------------------------------------------\n\nlabel_mask = train[TARGETS].notna().all(axis=1)\nlabeled_train = train.loc[label_mask].copy()\n\nrecords = []\n\nfor _, row in labeled_train.iterrows():\n\n study_id = row[\"StudyInstanceUID\"]\n report = row[\"Report\"]\n\n sentences = split_sentences(report)\n\n for target in TARGETS:\n\n label = int(row[target])\n\n target_results = []\n\n for sentence in sentences:\n result = classify_sentence(sentence, target)\n\n if result != \"NO_TARGET_MENTION\":\n target_results.append(\n {\n \"sentence\": sentence,\n \"evidence\": result\n }\n )\n\n # --------------------------------------------------------\n # Aggregate sentence-level evidence\n # --------------------------------------------------------\n\n if not target_results:\n overall_evidence = \"NO_EVIDENCE\"\n\n else:\n evidence_types = [\n x[\"evidence\"]\n for x in target_results\n ]\n\n if \"POSITIVE\" in evidence_types and \"NEGATIVE\" in evidence_types:\n overall_evidence = \"CONFLICTING_EVIDENCE\"\n\n elif \"POSITIVE\" in evidence_types:\n overall_evidence = \"POSITIVE\"\n\n elif \"NEGATIVE\" in evidence_types:\n overall_evidence = \"NEGATIVE\"\n\n else:\n overall_evidence = \"UNCERTAIN\"\n\n # --------------------------------------------------------\n # Compare against actual dataset label\n # --------------------------------------------------------\n\n if overall_evidence == \"POSITIVE\":\n if label == 1:\n agreement = \"AGREEMENT_POSITIVE\"\n else:\n agreement = \"LABEL_0_REPORT_POSITIVE\"\n\n elif overall_evidence == \"NEGATIVE\":\n if label == 0:\n agreement = \"AGREEMENT_NEGATIVE\"\n else:\n agreement = \"LABEL_1_REPORT_NEGATIVE\"\n\n elif overall_evidence == \"CONFLICTING_EVIDENCE\":\n agreement = \"CONFLICTING_REPORT\"\n\n else:\n agreement = \"UNCLEAR\"\n\n records.append(\n {\n \"StudyInstanceUID\": study_id,\n \"target\": target,\n \"label\": label,\n \"report_evidence\": overall_evidence,\n \"agreement\": agreement,\n \"evidence_sentence_count\": len(target_results),\n \"evidence_text\": \" || \".join(\n x[\"sentence\"]\n for x in target_results\n )\n }\n )\n\n\nevidence_df = pd.DataFrame(records)\n\n\n# ----------------------------------------------------------------\n# 8. Print overall summary\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"OVERALL EVIDENCE SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n evidence_df[\"report_evidence\"]\n .value_counts()\n .to_string()\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"LABEL / REPORT AGREEMENT\")\nprint(\"=\" * 70)\n\nprint(\n evidence_df[\"agreement\"]\n .value_counts()\n .to_string()\n)\n\n\n# ----------------------------------------------------------------\n# 9. Per-target summary\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"PER-TARGET SUMMARY\")\nprint(\"=\" * 70)\n\ntarget_summary = (\n evidence_df\n .groupby([\"target\", \"agreement\"])\n .size()\n .unstack(fill_value=0)\n)\n\nprint(target_summary)\n\n\n# ----------------------------------------------------------------\n# 10. Show high-confidence examples\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"SAMPLE POSITIVE AGREEMENTS\")\nprint(\"=\" * 70)\n\npositive_examples = evidence_df[\n evidence_df[\"agreement\"] == \"AGREEMENT_POSITIVE\"\n].head(20)\n\nif len(positive_examples) > 0:\n print(\n positive_examples[\n [\n \"target\",\n \"label\",\n \"evidence_text\"\n ]\n ].to_string(index=False)\n )\nelse:\n print(\"No positive agreements found.\")\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"SAMPLE NEGATIVE AGREEMENTS\")\nprint(\"=\" * 70)\n\nnegative_examples = evidence_df[\n evidence_df[\"agreement\"] == \"AGREEMENT_NEGATIVE\"\n].head(20)\n\nif len(negative_examples) > 0:\n print(\n negative_examples[\n [\n \"target\",\n \"label\",\n \"evidence_text\"\n ]\n ].to_string(index=False)\n )\nelse:\n print(\"No negative agreements found.\")\n\n\n# ----------------------------------------------------------------\n# 11. Show conflicts\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"LABEL 0 / REPORT POSITIVE\")\nprint(\"=\" * 70)\n\nconflict_0 = evidence_df[\n evidence_df[\"agreement\"] == \"LABEL_0_REPORT_POSITIVE\"\n]\n\nprint(f\"Count: {len(conflict_0)}\")\n\nif len(conflict_0) > 0:\n print(\n conflict_0[\n [\n \"StudyInstanceUID\",\n \"target\",\n \"label\",\n \"report_evidence\",\n \"evidence_text\"\n ]\n ].head(30).to_string(index=False)\n )\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"LABEL 1 / REPORT NEGATIVE\")\nprint(\"=\" * 70)\n\nconflict_1 = evidence_df[\n evidence_df[\"agreement\"] == \"LABEL_1_REPORT_NEGATIVE\"\n]\n\nprint(f\"Count: {len(conflict_1)}\")\n\nif len(conflict_1) > 0:\n print(\n conflict_1[\n [\n \"StudyInstanceUID\",\n \"target\",\n \"label\",\n \"report_evidence\",\n \"evidence_text\"\n ]\n ].to_string(index=False)\n )\n\n\n# ----------------------------------------------------------------\n# 12. Save outputs\n# ----------------------------------------------------------------\n\nOUTPUT_DIR = \"/kaggle/working/rsna_knee_audit\"\nos.makedirs(OUTPUT_DIR, exist_ok=True)\n\nevidence_path = os.path.join(\n OUTPUT_DIR,\n \"cell23_improved_report_evidence.csv\"\n)\n\nsummary_path = os.path.join(\n OUTPUT_DIR,\n \"cell23_improved_report_evidence_summary.csv\"\n)\n\nevidence_df.to_csv(\n evidence_path,\n index=False\n)\n\ntarget_summary.to_csv(\n summary_path\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 23 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(f\"Saved:\")\nprint(evidence_path)\nprint(summary_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:38:37.168729Z","iopub.execute_input":"2026-08-11T08:38:37.169121Z","iopub.status.idle":"2026-08-11T08:38:37.342981Z","shell.execute_reply.started":"2026-08-11T08:38:37.169091Z","shell.execute_reply":"2026-08-11T08:38:37.341684Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 24 - TRAIN / TEST LEAKAGE AND STUDY INTEGRITY AUDIT\n# ================================================================\n\nimport os\nimport pandas as pd\nimport numpy as np\n\nprint(\"=\" * 70)\nprint(\"TRAIN / TEST LEAKAGE AND STUDY INTEGRITY AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# 1. Load required CSVs if necessary\n# ----------------------------------------------------------------\n\nBASE_DIR = \"/kaggle/input/competitions/rsna-knee-abnormality-detection\"\n\ntrain_csv = os.path.join(BASE_DIR, \"train.csv\")\ntest_csv = os.path.join(BASE_DIR, \"test.csv\")\ntrain_series_csv = os.path.join(BASE_DIR, \"train_series.csv\")\ntest_series_csv = os.path.join(BASE_DIR, \"test_series.csv\")\n\nif \"train\" not in globals():\n train = pd.read_csv(train_csv)\n\nif \"test\" not in globals():\n test = pd.read_csv(test_csv)\n\nif \"train_series\" not in globals():\n train_series = pd.read_csv(train_series_csv)\n\nif \"test_series\" not in globals():\n test_series = pd.read_csv(test_series_csv)\n\nprint(f\"Train studies: {len(train)}\")\nprint(f\"Test studies: {len(test)}\")\nprint(f\"Train series rows: {len(train_series)}\")\nprint(f\"Test series rows: {len(test_series)}\")\n\n\n# ----------------------------------------------------------------\n# 2. StudyInstanceUID overlap\n# ----------------------------------------------------------------\n\ntrain_studies = set(\n train[\"StudyInstanceUID\"].astype(str)\n)\n\ntest_studies = set(\n test[\"StudyInstanceUID\"].astype(str)\n)\n\nstudy_overlap = train_studies.intersection(test_studies)\n\nprint()\nprint(\"=\" * 70)\nprint(\"STUDY INSTANCE UID OVERLAP\")\nprint(\"=\" * 70)\n\nprint(f\"Train unique studies: {len(train_studies)}\")\nprint(f\"Test unique studies: {len(test_studies)}\")\nprint(f\"Overlapping studies: {len(study_overlap)}\")\n\nif study_overlap:\n print(\"WARNING: StudyInstanceUID leakage detected.\")\n print(list(study_overlap)[:20])\nelse:\n print(\"PASS: No StudyInstanceUID overlap.\")\n\n\n# ----------------------------------------------------------------\n# 3. SeriesInstanceUID overlap\n# ----------------------------------------------------------------\n\ntrain_series_ids = set(\n train_series[\"SeriesInstanceUID\"].astype(str)\n)\n\ntest_series_ids = set(\n test_series[\"SeriesInstanceUID\"].astype(str)\n)\n\nseries_overlap = train_series_ids.intersection(\n test_series_ids\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"SERIES INSTANCE UID OVERLAP\")\nprint(\"=\" * 70)\n\nprint(f\"Train unique series: {len(train_series_ids)}\")\nprint(f\"Test unique series: {len(test_series_ids)}\")\nprint(f\"Overlapping series: {len(series_overlap)}\")\n\nif series_overlap:\n print(\"WARNING: SeriesInstanceUID leakage detected.\")\n print(list(series_overlap)[:20])\nelse:\n print(\"PASS: No SeriesInstanceUID overlap.\")\n\n\n# ----------------------------------------------------------------\n# 4. Verify every train series belongs to train\n# ----------------------------------------------------------------\n\ntrain_study_ids_from_series = set(\n train_series[\"StudyInstanceUID\"].astype(str)\n)\n\ntest_study_ids_from_series = set(\n test_series[\"StudyInstanceUID\"].astype(str)\n)\n\ntrain_series_orphan_studies = (\n train_study_ids_from_series - train_studies\n)\n\ntest_series_orphan_studies = (\n test_study_ids_from_series - test_studies\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"SERIES -> STUDY CONSISTENCY\")\nprint(\"=\" * 70)\n\nprint(\n \"Train-series studies not present in train.csv:\",\n len(train_series_orphan_studies)\n)\n\nprint(\n \"Test-series studies not present in test.csv:\",\n len(test_series_orphan_studies)\n)\n\nif len(train_series_orphan_studies) == 0:\n print(\"PASS: Every train series maps to a train study.\")\n\nif len(test_series_orphan_studies) == 0:\n print(\"PASS: Every test series maps to a test study.\")\n\n\n# ----------------------------------------------------------------\n# 5. Check series counts per study\n# ----------------------------------------------------------------\n\ntrain_series_counts = (\n train_series\n .groupby(\"StudyInstanceUID\")[\"SeriesInstanceUID\"]\n .nunique()\n)\n\ntest_series_counts = (\n test_series\n .groupby(\"StudyInstanceUID\")[\"SeriesInstanceUID\"]\n .nunique()\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"SERIES PER STUDY\")\nprint(\"=\" * 70)\n\nprint(\"TRAIN:\")\nprint(train_series_counts.describe().to_string())\n\nprint()\nprint(\"TEST:\")\nprint(test_series_counts.describe().to_string())\n\n\n# ----------------------------------------------------------------\n# 6. Check duplicate StudyInstanceUID rows\n# ----------------------------------------------------------------\n\ntrain_duplicate_rows = train[\n train[\"StudyInstanceUID\"].duplicated(keep=False)\n]\n\ntest_duplicate_rows = test[\n test[\"StudyInstanceUID\"].duplicated(keep=False)\n]\n\nprint()\nprint(\"=\" * 70)\nprint(\"DUPLICATE STUDY ROWS\")\nprint(\"=\" * 70)\n\nprint(\n f\"Duplicate train study rows: {len(train_duplicate_rows)}\"\n)\n\nprint(\n f\"Duplicate test study rows: {len(test_duplicate_rows)}\"\n)\n\n\n# ----------------------------------------------------------------\n# 7. Check duplicate SeriesInstanceUID rows\n# ----------------------------------------------------------------\n\ntrain_series_duplicate_rows = train_series[\n train_series[\"SeriesInstanceUID\"].duplicated(keep=False)\n]\n\ntest_series_duplicate_rows = test_series[\n test_series[\"SeriesInstanceUID\"].duplicated(keep=False)\n]\n\nprint()\nprint(\"=\" * 70)\nprint(\"DUPLICATE SERIES ROWS\")\nprint(\"=\" * 70)\n\nprint(\n f\"Duplicate train series rows: {len(train_series_duplicate_rows)}\"\n)\n\nprint(\n f\"Duplicate test series rows: {len(test_series_duplicate_rows)}\"\n)\n\n\n# ----------------------------------------------------------------\n# 8. Verify StudyInstanceUID -> series mapping\n# ----------------------------------------------------------------\n\ntrain_mapping_counts = (\n train_series\n .groupby(\"SeriesInstanceUID\")[\"StudyInstanceUID\"]\n .nunique()\n)\n\ntest_mapping_counts = (\n test_series\n .groupby(\"SeriesInstanceUID\")[\"StudyInstanceUID\"]\n .nunique()\n)\n\ntrain_bad_series_mapping = (\n train_mapping_counts[\n train_mapping_counts != 1\n ]\n)\n\ntest_bad_series_mapping = (\n test_mapping_counts[\n test_mapping_counts != 1\n ]\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"SERIES -> STUDY ONE-TO-ONE MAPPING\")\nprint(\"=\" * 70)\n\nprint(\n \"Train series with multiple studies:\",\n len(train_bad_series_mapping)\n)\n\nprint(\n \"Test series with multiple studies:\",\n len(test_bad_series_mapping)\n)\n\nif len(train_bad_series_mapping) == 0:\n print(\"PASS: Every train series maps to exactly one study.\")\n\nif len(test_bad_series_mapping) == 0:\n print(\"PASS: Every test series maps to exactly one study.\")\n\n\n# ----------------------------------------------------------------\n# 9. Labelled studies and series\n# ----------------------------------------------------------------\n\nif \"TARGETS\" not in globals():\n TARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n ]\n\nlabeled_mask = train[TARGETS].notna().all(axis=1)\n\nlabeled_studies = set(\n train.loc[\n labeled_mask,\n \"StudyInstanceUID\"\n ].astype(str)\n)\n\nlabeled_series = train_series[\n train_series[\"StudyInstanceUID\"].astype(str).isin(\n labeled_studies\n )\n].copy()\n\nprint()\nprint(\"=\" * 70)\nprint(\"LABELED STUDY COVERAGE\")\nprint(\"=\" * 70)\n\nprint(f\"Fully labeled studies: {len(labeled_studies)}\")\nprint(f\"Series belonging to labeled studies: {len(labeled_series)}\")\nprint(\n f\"Unique series in labeled studies: \"\n f\"{labeled_series['SeriesInstanceUID'].nunique()}\"\n)\n\n\n# ----------------------------------------------------------------\n# 10. Metadata overlap between train and test\n# ----------------------------------------------------------------\n\nmetadata_columns = [\n \"Fluid_Sensitive\",\n \"Fat_Suppression\",\n \"Anatomical_Plane\"\n]\n\nprint()\nprint(\"=\" * 70)\nprint(\"TRAIN / TEST SERIES METADATA DISTRIBUTION\")\nprint(\"=\" * 70)\n\nfor col in metadata_columns:\n\n print()\n print(f\"--- {col} ---\")\n\n train_dist = (\n train_series[col]\n .value_counts(dropna=False)\n .rename(\"train_count\")\n )\n\n test_dist = (\n test_series[col]\n .value_counts(dropna=False)\n .rename(\"test_count\")\n )\n\n comparison = pd.concat(\n [\n train_dist,\n test_dist\n ],\n axis=1\n ).fillna(0)\n\n print(comparison)\n\n\n# ----------------------------------------------------------------\n# 11. Final integrity verdict\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL INTEGRITY VERDICT\")\nprint(\"=\" * 70)\n\nchecks = {\n \"Study UID leakage\": len(study_overlap) == 0,\n \"Series UID leakage\": len(series_overlap) == 0,\n \"Train series orphan studies\": len(train_series_orphan_studies) == 0,\n \"Test series orphan studies\": len(test_series_orphan_studies) == 0,\n \"Duplicate train studies\": len(train_duplicate_rows) == 0,\n \"Duplicate test studies\": len(test_duplicate_rows) == 0,\n \"Duplicate train series\": len(train_series_duplicate_rows) == 0,\n \"Duplicate test series\": len(test_series_duplicate_rows) == 0,\n \"Train series mapping\": len(train_bad_series_mapping) == 0,\n \"Test series mapping\": len(test_bad_series_mapping) == 0\n}\n\nfor check_name, passed in checks.items():\n print(\n f\"{'PASS' if passed else 'FAIL'} - {check_name}\"\n )\n\nall_pass = all(checks.values())\n\nprint()\nprint(\n \"OVERALL:\",\n \"PASS\" if all_pass else \"REVIEW REQUIRED\"\n)\n\n\n# ----------------------------------------------------------------\n# 12. Save audit\n# ----------------------------------------------------------------\n\nOUTPUT_DIR = \"/kaggle/working/rsna_knee_audit\"\nos.makedirs(OUTPUT_DIR, exist_ok=True)\n\naudit_rows = []\n\nfor check_name, passed in checks.items():\n audit_rows.append(\n {\n \"check\": check_name,\n \"passed\": passed\n }\n )\n\naudit_df = pd.DataFrame(audit_rows)\n\naudit_path = os.path.join(\n OUTPUT_DIR,\n \"cell24_train_test_integrity_audit.csv\"\n)\n\naudit_df.to_csv(\n audit_path,\n index=False\n)\n\nprint()\nprint(f\"Saved: {audit_path}\")\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 24 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:39:39.259355Z","iopub.execute_input":"2026-08-11T08:39:39.259651Z","iopub.status.idle":"2026-08-11T08:39:39.382223Z","shell.execute_reply.started":"2026-08-11T08:39:39.259627Z","shell.execute_reply":"2026-08-11T08:39:39.381352Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 25 - LABEL-AWARE VALIDATION FEASIBILITY AUDIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nfrom sklearn.model_selection import StratifiedKFold\nfrom sklearn.metrics import roc_auc_score\n\nprint(\"=\" * 70)\nprint(\"LABEL-AWARE VALIDATION FEASIBILITY AUDIT\")\nprint(\"=\" * 70)\n\n\n# ----------------------------------------------------------------\n# 1. Load train data if necessary\n# ----------------------------------------------------------------\n\nBASE_DIR = \"/kaggle/input/competitions/rsna-knee-abnormality-detection\"\n\ntrain_csv = os.path.join(BASE_DIR, \"train.csv\")\n\nif \"train\" not in globals():\n train = pd.read_csv(train_csv)\n\nif \"TARGETS\" not in globals():\n TARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n ]\n\n\n# ----------------------------------------------------------------\n# 2. Select fully labeled studies\n# ----------------------------------------------------------------\n\nlabel_mask = train[TARGETS].notna().all(axis=1)\n\nlabeled_train = train.loc[label_mask].copy()\n\nprint(f\"Total train studies: {len(train)}\")\nprint(f\"Fully labeled studies: {len(labeled_train)}\")\n\n\n# ----------------------------------------------------------------\n# 3. Basic label counts\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"LABEL COUNTS\")\nprint(\"=\" * 70)\n\nlabel_counts = []\n\nfor target in TARGETS:\n\n positive = int(\n (labeled_train[target] == 1).sum()\n )\n\n negative = int(\n (labeled_train[target] == 0).sum()\n )\n\n label_counts.append(\n {\n \"target\": target,\n \"positive\": positive,\n \"negative\": negative,\n \"total\": positive + negative,\n \"positive_rate\": positive / (positive + negative)\n }\n )\n\nlabel_counts_df = pd.DataFrame(label_counts)\n\nprint(\n label_counts_df.to_string(\n index=False\n )\n)\n\n\n# ----------------------------------------------------------------\n# 4. Determine feasible number of folds\n# ----------------------------------------------------------------\n\nmin_positive = label_counts_df[\"positive\"].min()\nmin_negative = label_counts_df[\"negative\"].min()\n\nprint()\nprint(\"=\" * 70)\nprint(\"FOLD FEASIBILITY\")\nprint(\"=\" * 70)\n\nprint(f\"Minimum positive count across targets: {min_positive}\")\nprint(f\"Minimum negative count across targets: {min_negative}\")\n\nprint()\nprint(\n \"Important: For ordinary stratification of a binary target, \"\n \"the number of folds cannot exceed the smaller class count.\"\n)\n\n\n# ----------------------------------------------------------------\n# 5. Examine candidate fold counts\n# ----------------------------------------------------------------\n\ncandidate_folds = [2, 3, 4, 5]\n\nfold_feasibility = []\n\nfor n_splits in candidate_folds:\n\n feasible = (\n min_positive >= n_splits\n and min_negative >= n_splits\n )\n\n fold_feasibility.append(\n {\n \"n_splits\": n_splits,\n \"feasible_for_all_targets\": feasible\n }\n )\n\nfold_feasibility_df = pd.DataFrame(\n fold_feasibility\n)\n\nprint(\n fold_feasibility_df.to_string(\n index=False\n )\n)\n\n\n# ----------------------------------------------------------------\n# 6. Build multilabel stratification signature\n#\n# We create a compact representation of the 12-label combination.\n# This is NOT our final CV splitter. It is only a feasibility\n# diagnostic.\n# ----------------------------------------------------------------\n\nlabel_signature = (\n labeled_train[TARGETS]\n .astype(int)\n .astype(str)\n .agg(\"\".join, axis=1)\n)\n\nsignature_counts = (\n label_signature\n .value_counts()\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"MULTILABEL COMBINATION FREQUENCY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Unique 12-target label combinations: \"\n f\"{len(signature_counts)}\"\n)\n\nprint()\nprint(\"Most common combinations:\")\n\nprint(\n signature_counts.head(20).to_string()\n)\n\n\n# ----------------------------------------------------------------\n# 7. Check how many unique combinations occur only once\n# ----------------------------------------------------------------\n\nsingletons = int(\n (signature_counts == 1).sum()\n)\n\nrare_combinations = int(\n (signature_counts <= 2).sum()\n)\n\nprint()\nprint(\n f\"Combinations occurring exactly once: {singletons}\"\n)\n\nprint(\n f\"Combinations occurring at most twice: {rare_combinations}\"\n)\n\n\n# ----------------------------------------------------------------\n# 8. Simulate multilabel-aware fold assignment\n#\n# We use an iterative greedy assignment based on label frequencies.\n# This is a diagnostic splitter, not the final production splitter.\n# ----------------------------------------------------------------\n\ndef iterative_multilabel_split(\n labels,\n n_splits=3,\n random_state=42\n):\n \"\"\"\n Simple iterative multilabel stratification diagnostic.\n\n Attempts to distribute rare positive labels across folds\n while keeping fold sizes approximately balanced.\n \"\"\"\n\n rng = np.random.default_rng(random_state)\n\n Y = np.asarray(labels).astype(int)\n\n n_samples = Y.shape[0]\n\n folds = [\n []\n for _ in range(n_splits)\n ]\n\n # Start with rarest positive label\n remaining = set(range(n_samples))\n\n label_counts = Y.sum(axis=0)\n\n # Sort samples by rarity of their positive labels\n sample_priority = []\n\n for i in range(n_samples):\n\n positive_labels = np.where(\n Y[i] == 1\n )[0]\n\n if len(positive_labels) == 0:\n rarity_score = 0\n else:\n rarity_score = sum(\n 1.0 / max(label_counts[j], 1)\n for j in positive_labels\n )\n\n sample_priority.append(\n rarity_score\n )\n\n order = np.argsort(\n -np.asarray(sample_priority)\n )\n\n # Assign samples greedily\n fold_label_counts = np.zeros(\n (n_splits, Y.shape[1]),\n dtype=int\n )\n\n fold_sizes = np.zeros(\n n_splits,\n dtype=int\n )\n\n for idx in order:\n\n sample = Y[idx]\n\n scores = []\n\n for fold_idx in range(n_splits):\n\n label_score = np.sum(\n fold_label_counts[fold_idx]\n * sample\n )\n\n size_score = fold_sizes[fold_idx]\n\n scores.append(\n label_score * 10.0\n + size_score\n )\n\n min_score = min(scores)\n\n candidates = [\n i\n for i, score in enumerate(scores)\n if score == min_score\n ]\n\n chosen = rng.choice(\n candidates\n )\n\n folds[chosen].append(idx)\n\n fold_label_counts[\n chosen\n ] += sample\n\n fold_sizes[\n chosen\n ] += 1\n\n return folds\n\n\n# ----------------------------------------------------------------\n# 9. Analyze candidate fold structures\n# ----------------------------------------------------------------\n\nfold_diagnostics = []\n\nfor n_splits in [3, 4, 5]:\n\n folds = iterative_multilabel_split(\n labeled_train[TARGETS].values,\n n_splits=n_splits,\n random_state=42\n )\n\n print()\n print(\"=\" * 70)\n print(f\"{n_splits}-FOLD MULTILABEL DIAGNOSTIC\")\n print(\"=\" * 70)\n\n for fold_idx, indices in enumerate(folds):\n\n fold_data = labeled_train.iloc[\n indices\n ]\n\n print()\n print(\n f\"Fold {fold_idx + 1}: \"\n f\"{len(fold_data)} studies\"\n )\n\n for target in TARGETS:\n\n positives = int(\n (fold_data[target] == 1).sum()\n )\n\n negatives = int(\n (fold_data[target] == 0).sum()\n )\n\n print(\n f\"{target:20s} \"\n f\"pos={positives:2d} \"\n f\"neg={negatives:2d}\"\n )\n\n fold_diagnostics.append(\n {\n \"n_splits\": n_splits,\n \"fold\": fold_idx + 1,\n \"target\": target,\n \"positive\": positives,\n \"negative\": negatives\n }\n )\n\n\n# ----------------------------------------------------------------\n# 10. Identify impossible AUC folds\n# ----------------------------------------------------------------\n\ndiagnostics_df = pd.DataFrame(\n fold_diagnostics\n)\n\ndiagnostics_df[\"auc_possible\"] = (\n (diagnostics_df[\"positive\"] > 0)\n &\n (diagnostics_df[\"negative\"] > 0)\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"AUC FEASIBILITY\")\nprint(\"=\" * 70)\n\nfor n_splits in [3, 4, 5]:\n\n subset = diagnostics_df[\n diagnostics_df[\"n_splits\"] == n_splits\n ]\n\n impossible = subset[\n ~subset[\"auc_possible\"]\n ]\n\n print(\n f\"{n_splits}-fold: \"\n f\"{len(impossible)} target-fold combinations \"\n f\"without both classes\"\n )\n\n if len(impossible) > 0:\n print(\n impossible[\n [\n \"fold\",\n \"target\",\n \"positive\",\n \"negative\"\n ]\n ].to_string(index=False)\n )\n\n\n# ----------------------------------------------------------------\n# 11. Estimate minimum positive examples per validation fold\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"MINIMUM POSITIVE COUNT PER FOLD\")\nprint(\"=\" * 70)\n\nfor n_splits in [3, 4, 5]:\n\n subset = diagnostics_df[\n diagnostics_df[\"n_splits\"] == n_splits\n ]\n\n minimum_positive = (\n subset[\"positive\"].min()\n )\n\n minimum_negative = (\n subset[\"negative\"].min()\n )\n\n print(\n f\"{n_splits}-fold -> \"\n f\"minimum positive = {minimum_positive}, \"\n f\"minimum negative = {minimum_negative}\"\n )\n\n\n# ----------------------------------------------------------------\n# 12. Save outputs\n# ----------------------------------------------------------------\n\nOUTPUT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nos.makedirs(\n OUTPUT_DIR,\n exist_ok=True\n)\n\nlabel_counts_path = os.path.join(\n OUTPUT_DIR,\n \"cell25_label_counts.csv\"\n)\n\nfold_feasibility_path = os.path.join(\n OUTPUT_DIR,\n \"cell25_fold_feasibility.csv\"\n)\n\nsignature_path = os.path.join(\n OUTPUT_DIR,\n \"cell25_label_signature_counts.csv\"\n)\n\ndiagnostics_path = os.path.join(\n OUTPUT_DIR,\n \"cell25_fold_diagnostics.csv\"\n)\n\nlabel_counts_df.to_csv(\n label_counts_path,\n index=False\n)\n\nfold_feasibility_df.to_csv(\n fold_feasibility_path,\n index=False\n)\n\nsignature_counts.rename(\n \"count\"\n).to_csv(\n signature_path\n)\n\ndiagnostics_df.to_csv(\n diagnostics_path,\n index=False\n)\n\n\n# ----------------------------------------------------------------\n# 13. Final message\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 25 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"Saved:\")\nprint(label_counts_path)\nprint(fold_feasibility_path)\nprint(signature_path)\nprint(diagnostics_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:40:56.94178Z","iopub.execute_input":"2026-08-11T08:40:56.942152Z","iopub.status.idle":"2026-08-11T08:40:58.216362Z","shell.execute_reply.started":"2026-08-11T08:40:56.942123Z","shell.execute_reply":"2026-08-11T08:40:58.215276Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 26 - LABELED STUDY MRI SEQUENCE / SERIES AUDIT\n# ================================================================\n\nimport os\nimport pandas as pd\nimport numpy as np\n\nprint(\"=\" * 70)\nprint(\"LABELED STUDY MRI SEQUENCE / SERIES AUDIT\")\nprint(\"=\" * 70)\n\n\n# ----------------------------------------------------------------\n# 1. Load required data\n# ----------------------------------------------------------------\n\nBASE_DIR = \"/kaggle/input/competitions/rsna-knee-abnormality-detection\"\n\nif \"train\" not in globals():\n train = pd.read_csv(\n os.path.join(BASE_DIR, \"train.csv\")\n )\n\nif \"train_series\" not in globals():\n train_series = pd.read_csv(\n os.path.join(BASE_DIR, \"train_series.csv\")\n )\n\nif \"TARGETS\" not in globals():\n TARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n ]\n\n\n# ----------------------------------------------------------------\n# 2. Select the 58 fully labeled studies\n# ----------------------------------------------------------------\n\nlabel_mask = train[TARGETS].notna().all(axis=1)\n\nlabeled_train = train.loc[\n label_mask\n].copy()\n\nlabeled_study_ids = set(\n labeled_train[\n \"StudyInstanceUID\"\n ].astype(str)\n)\n\nlabeled_series = train_series[\n train_series[\n \"StudyInstanceUID\"\n ].astype(str).isin(\n labeled_study_ids\n )\n].copy()\n\nprint(\n f\"Fully labeled studies: \"\n f\"{len(labeled_train)}\"\n)\n\nprint(\n f\"Labeled-study series rows: \"\n f\"{len(labeled_series)}\"\n)\n\nprint(\n f\"Unique labeled-study series: \"\n f\"{labeled_series['SeriesInstanceUID'].nunique()}\"\n)\n\n\n# ----------------------------------------------------------------\n# 3. Series per labeled study\n# ----------------------------------------------------------------\n\nseries_per_study = (\n labeled_series\n .groupby(\"StudyInstanceUID\")\n [\"SeriesInstanceUID\"]\n .nunique()\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"SERIES PER LABELED STUDY\")\nprint(\"=\" * 70)\n\nprint(\n series_per_study.describe().to_string()\n)\n\n\n# ----------------------------------------------------------------\n# 4. Plane distribution\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"LABELED-STUDY PLANE DISTRIBUTION\")\nprint(\"=\" * 70)\n\nplane_counts = (\n labeled_series[\n \"Anatomical_Plane\"\n ]\n .value_counts(dropna=False)\n)\n\nplane_percent = (\n plane_counts\n / plane_counts.sum()\n * 100\n)\n\nplane_table = pd.DataFrame({\n \"count\": plane_counts,\n \"percentage\": plane_percent.round(2)\n})\n\nprint(\n plane_table.to_string()\n)\n\n\n# ----------------------------------------------------------------\n# 5. Sequence property distribution\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FLUID-SENSITIVE DISTRIBUTION\")\nprint(\"=\" * 70)\n\nfluid_counts = (\n labeled_series[\n \"Fluid_Sensitive\"\n ]\n .value_counts(dropna=False)\n)\n\nfluid_table = pd.DataFrame({\n \"count\": fluid_counts,\n \"percentage\": (\n fluid_counts\n / fluid_counts.sum()\n * 100\n ).round(2)\n})\n\nprint(\n fluid_table.to_string()\n)\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"FAT-SUPPRESSION DISTRIBUTION\")\nprint(\"=\" * 70)\n\nfat_counts = (\n labeled_series[\n \"Fat_Suppression\"\n ]\n .value_counts(dropna=False)\n)\n\nfat_table = pd.DataFrame({\n \"count\": fat_counts,\n \"percentage\": (\n fat_counts\n / fat_counts.sum()\n * 100\n ).round(2)\n})\n\nprint(\n fat_table.to_string()\n)\n\n\n# ----------------------------------------------------------------\n# 6. Plane + sequence combinations\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"PLANE + SEQUENCE COMBINATIONS\")\nprint(\"=\" * 70)\n\ncombo = (\n labeled_series\n .groupby(\n [\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\"\n ],\n dropna=False\n )\n .size()\n .reset_index(\n name=\"series_count\"\n )\n .sort_values(\n \"series_count\",\n ascending=False\n )\n)\n\ncombo[\"percentage\"] = (\n combo[\"series_count\"]\n / len(labeled_series)\n * 100\n).round(2)\n\nprint(\n combo.to_string(index=False)\n)\n\n\n# ----------------------------------------------------------------\n# 7. Series descriptions\n#\n# train_series.csv does not necessarily contain descriptions.\n# Therefore we only inspect columns that actually exist.\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"AVAILABLE TRAIN SERIES COLUMNS\")\nprint(\"=\" * 70)\n\nprint(\n list(train_series.columns)\n)\n\nif \"SeriesDescription\" in labeled_series.columns:\n\n print()\n print(\"=\" * 70)\n print(\"TOP SERIES DESCRIPTIONS\")\n print(\"=\" * 70)\n\n descriptions = (\n labeled_series[\n \"SeriesDescription\"\n ]\n .fillna(\"MISSING\")\n .astype(str)\n .value_counts()\n .head(50)\n )\n\n print(\n descriptions.to_string()\n )\n\nelse:\n\n print()\n print(\n \"SeriesDescription is not available \"\n \"in train_series.csv.\"\n )\n\n\n# ----------------------------------------------------------------\n# 8. Per-study plane coverage\n# ----------------------------------------------------------------\n\nstudy_plane = (\n labeled_series\n .pivot_table(\n index=\"StudyInstanceUID\",\n columns=\"Anatomical_Plane\",\n values=\"SeriesInstanceUID\",\n aggfunc=\"nunique\",\n fill_value=0\n )\n)\n\nfor plane in [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n]:\n if plane not in study_plane.columns:\n study_plane[plane] = 0\n\nstudy_plane = study_plane[\n [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n ]\n]\n\nprint()\nprint(\"=\" * 70)\nprint(\"PLANE COVERAGE PER LABELED STUDY\")\nprint(\"=\" * 70)\n\nprint(\n study_plane.describe().to_string()\n)\n\n\n# ----------------------------------------------------------------\n# 9. Count studies containing each plane\n# ----------------------------------------------------------------\n\nplane_presence = pd.DataFrame(\n {\n \"plane\": [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n ],\n \"studies_with_plane\": [\n int(\n (study_plane[\"Sagittal\"] > 0).sum()\n ),\n int(\n (study_plane[\"Coronal\"] > 0).sum()\n ),\n int(\n (study_plane[\"Axial\"] > 0).sum()\n )\n ]\n }\n)\n\nplane_presence[\"percentage_of_labeled_studies\"] = (\n plane_presence[\"studies_with_plane\"]\n / len(labeled_train)\n * 100\n).round(2)\n\nprint()\nprint(\n plane_presence.to_string(\n index=False\n )\n)\n\n\n# ----------------------------------------------------------------\n# 10. Count studies containing each sequence property\n# ----------------------------------------------------------------\n\nfluid_presence = (\n labeled_series\n .groupby(\"StudyInstanceUID\")\n [\"Fluid_Sensitive\"]\n .apply(\n lambda x: int((x == 1).any())\n )\n)\n\nfat_presence = (\n labeled_series\n .groupby(\"StudyInstanceUID\")\n [\"Fat_Suppression\"]\n .apply(\n lambda x: int((x == 1).any())\n )\n)\n\nsequence_presence = pd.DataFrame({\n \"Fluid_Sensitive_present\": fluid_presence,\n \"Fat_Suppression_present\": fat_presence\n})\n\nprint()\nprint(\"=\" * 70)\nprint(\"SEQUENCE PROPERTY PRESENCE PER LABELED STUDY\")\nprint(\"=\" * 70)\n\nprint(\n sequence_presence.sum().to_string()\n)\n\nprint()\n\nprint(\n (\n sequence_presence.mean() * 100\n ).round(2).to_string()\n)\n\n\n# ----------------------------------------------------------------\n# 11. Studies with complete three-plane coverage\n# ----------------------------------------------------------------\n\ncomplete_three_plane = (\n (study_plane[\"Sagittal\"] > 0)\n &\n (study_plane[\"Coronal\"] > 0)\n &\n (study_plane[\"Axial\"] > 0)\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"THREE-PLANE COVERAGE\")\nprint(\"=\" * 70)\n\nprint(\n \"Studies containing all three planes:\",\n int(complete_three_plane.sum())\n)\n\nprint(\n \"Percentage:\",\n round(\n complete_three_plane.mean() * 100,\n 2\n )\n)\n\n\n# ----------------------------------------------------------------\n# 12. Studies with multiple series of same plane\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"MULTIPLE SERIES PER PLANE\")\nprint(\"=\" * 70)\n\nfor plane in [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n]:\n\n values = study_plane[plane]\n\n print()\n print(\n f\"{plane}:\"\n )\n\n print(\n f\" Studies with 0 series: \"\n f\"{int((values == 0).sum())}\"\n )\n\n print(\n f\" Studies with 1 series: \"\n f\"{int((values == 1).sum())}\"\n )\n\n print(\n f\" Studies with >1 series: \"\n f\"{int((values > 1).sum())}\"\n )\n\n print(\n f\" Maximum series in one study: \"\n f\"{int(values.max())}\"\n )\n\n\n# ----------------------------------------------------------------\n# 13. Create study-level sequence summary\n# ----------------------------------------------------------------\n\nstudy_sequence_summary = (\n labeled_series\n .groupby(\"StudyInstanceUID\")\n .agg(\n total_series=(\n \"SeriesInstanceUID\",\n \"nunique\"\n ),\n fluid_sensitive_series=(\n \"Fluid_Sensitive\",\n lambda x: int((x == 1).sum())\n ),\n fat_suppressed_series=(\n \"Fat_Suppression\",\n lambda x: int((x == 1).sum())\n ),\n sagittal_series=(\n \"Anatomical_Plane\",\n lambda x: int((x == \"Sagittal\").sum())\n ),\n coronal_series=(\n \"Anatomical_Plane\",\n lambda x: int((x == \"Coronal\").sum())\n ),\n axial_series=(\n \"Anatomical_Plane\",\n lambda x: int((x == \"Axial\").sum())\n )\n )\n .reset_index()\n)\n\n\n# ----------------------------------------------------------------\n# 14. Merge labels with study-level sequence information\n# ----------------------------------------------------------------\n\nstudy_sequence_summary = study_sequence_summary.merge(\n labeled_train[\n [\"StudyInstanceUID\"] + TARGETS\n ],\n on=\"StudyInstanceUID\",\n how=\"left\"\n)\n\n\n# ----------------------------------------------------------------\n# 15. Save outputs\n# ----------------------------------------------------------------\n\nOUTPUT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nos.makedirs(\n OUTPUT_DIR,\n exist_ok=True\n)\n\nstudy_summary_path = os.path.join(\n OUTPUT_DIR,\n \"cell26_labeled_study_sequence_summary.csv\"\n)\n\ncombo_path = os.path.join(\n OUTPUT_DIR,\n \"cell26_sequence_combinations.csv\"\n)\n\nplane_presence_path = os.path.join(\n OUTPUT_DIR,\n \"cell26_plane_presence.csv\"\n)\n\nstudy_sequence_summary.to_csv(\n study_summary_path,\n index=False\n)\n\ncombo.to_csv(\n combo_path,\n index=False\n)\n\nplane_presence.to_csv(\n plane_presence_path,\n index=False\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 26 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"Saved:\")\nprint(study_summary_path)\nprint(combo_path)\nprint(plane_presence_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:42:31.650844Z","iopub.execute_input":"2026-08-11T08:42:31.651233Z","iopub.status.idle":"2026-08-11T08:42:31.778553Z","shell.execute_reply.started":"2026-08-11T08:42:31.651203Z","shell.execute_reply":"2026-08-11T08:42:31.777579Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 27 - MODEL INPUT SERIES SELECTION AUDIT\n# ================================================================\n\nimport os\nimport glob\nimport pandas as pd\nimport numpy as np\nimport pydicom\n\nprint(\"=\" * 70)\nprint(\"MODEL INPUT SERIES SELECTION AUDIT\")\nprint(\"=\" * 70)\n\n\n# ----------------------------------------------------------------\n# 1. Load required data\n# ----------------------------------------------------------------\n\nBASE_DIR = \"/kaggle/input/competitions/rsna-knee-abnormality-detection\"\n\nTRAIN_SERIES_DIR = os.path.join(\n BASE_DIR,\n \"train_series\"\n)\n\nTEST_SERIES_DIR = os.path.join(\n BASE_DIR,\n \"test_series\"\n)\n\nif \"train\" not in globals():\n train = pd.read_csv(\n os.path.join(BASE_DIR, \"train.csv\")\n )\n\nif \"train_series\" not in globals():\n train_series = pd.read_csv(\n os.path.join(BASE_DIR, \"train_series.csv\")\n )\n\nif \"test\" not in globals():\n test = pd.read_csv(\n os.path.join(BASE_DIR, \"test.csv\")\n )\n\nif \"test_series\" not in globals():\n test_series = pd.read_csv(\n os.path.join(BASE_DIR, \"test_series.csv\")\n )\n\nif \"TARGETS\" not in globals():\n TARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n ]\n\n\n# ----------------------------------------------------------------\n# 2. Identify fully labeled studies\n# ----------------------------------------------------------------\n\nlabel_mask = train[TARGETS].notna().all(axis=1)\n\nlabeled_train = train.loc[\n label_mask\n].copy()\n\nlabeled_study_ids = set(\n labeled_train[\n \"StudyInstanceUID\"\n ].astype(str)\n)\n\nlabeled_series = train_series[\n train_series[\n \"StudyInstanceUID\"\n ].astype(str).isin(\n labeled_study_ids\n )\n].copy()\n\nprint(\n \"Fully labeled studies:\",\n len(labeled_study_ids)\n)\n\nprint(\n \"Labeled series:\",\n len(labeled_series)\n)\n\n\n# ----------------------------------------------------------------\n# 3. Count DICOM files for each series\n# ----------------------------------------------------------------\n\ndef get_series_dir(\n root_dir,\n study_uid,\n series_uid\n):\n return os.path.join(\n root_dir,\n str(study_uid),\n str(series_uid)\n )\n\n\ndef count_dicom_files(\n series_dir\n):\n if not os.path.isdir(series_dir):\n return 0\n\n return len(\n glob.glob(\n os.path.join(\n series_dir,\n \"*.dcm\"\n )\n )\n )\n\n\n# ----------------------------------------------------------------\n# 4. Build series-level audit table\n# ----------------------------------------------------------------\n\nrows = []\n\nprint()\nprint(\"Auditing labeled-study series...\")\n\nfor idx, row in labeled_series.iterrows():\n\n study_uid = str(\n row[\"StudyInstanceUID\"]\n )\n\n series_uid = str(\n row[\"SeriesInstanceUID\"]\n )\n\n series_dir = get_series_dir(\n TRAIN_SERIES_DIR,\n study_uid,\n series_uid\n )\n\n slice_count = count_dicom_files(\n series_dir\n )\n\n first_dicom = None\n\n files = glob.glob(\n os.path.join(\n series_dir,\n \"*.dcm\"\n )\n )\n\n if files:\n try:\n first_dicom = pydicom.dcmread(\n files[0],\n stop_before_pixels=True\n )\n except Exception:\n first_dicom = None\n\n if first_dicom is not None:\n\n rows_count = getattr(\n first_dicom,\n \"Rows\",\n np.nan\n )\n\n cols_count = getattr(\n first_dicom,\n \"Columns\",\n np.nan\n )\n\n pixel_spacing = getattr(\n first_dicom,\n \"PixelSpacing\",\n [np.nan, np.nan]\n )\n\n if pixel_spacing is not None:\n try:\n pixel_spacing_x = float(\n pixel_spacing[0]\n )\n pixel_spacing_y = float(\n pixel_spacing[1]\n )\n except Exception:\n pixel_spacing_x = np.nan\n pixel_spacing_y = np.nan\n else:\n pixel_spacing_x = np.nan\n pixel_spacing_y = np.nan\n\n slice_thickness = getattr(\n first_dicom,\n \"SliceThickness\",\n np.nan\n )\n\n spacing_between = getattr(\n first_dicom,\n \"SpacingBetweenSlices\",\n np.nan\n )\n\n series_description = getattr(\n first_dicom,\n \"SeriesDescription\",\n \"\"\n )\n\n image_type = getattr(\n first_dicom,\n \"ImageType\",\n \"\"\n )\n\n if isinstance(\n image_type,\n (list, tuple)\n ):\n image_type = \"|\".join(\n map(str, image_type)\n )\n\n else:\n\n rows_count = np.nan\n cols_count = np.nan\n pixel_spacing_x = np.nan\n pixel_spacing_y = np.nan\n slice_thickness = np.nan\n spacing_between = np.nan\n series_description = \"\"\n image_type = \"\"\n\n rows.append(\n {\n \"StudyInstanceUID\": study_uid,\n \"SeriesInstanceUID\": series_uid,\n \"Anatomical_Plane\": row[\n \"Anatomical_Plane\"\n ],\n \"Fluid_Sensitive\": row[\n \"Fluid_Sensitive\"\n ],\n \"Fat_Suppression\": row[\n \"Fat_Suppression\"\n ],\n \"slice_count\": slice_count,\n \"Rows\": rows_count,\n \"Columns\": cols_count,\n \"PixelSpacingX\": pixel_spacing_x,\n \"PixelSpacingY\": pixel_spacing_y,\n \"SliceThickness\": slice_thickness,\n \"SpacingBetweenSlices\": spacing_between,\n \"SeriesDescription\": str(\n series_description\n ),\n \"ImageType\": str(\n image_type\n )\n }\n )\n\n if (\n len(rows) % 50 == 0\n ):\n print(\n f\"Processed {len(rows)} / \"\n f\"{len(labeled_series)}\"\n )\n\n\nseries_audit = pd.DataFrame(\n rows\n)\n\nprint()\nprint(\n \"Completed series audit:\",\n len(series_audit)\n)\n\n\n# ----------------------------------------------------------------\n# 5. Create candidate ranking\n#\n# Priority:\n# 1. Fluid sensitive\n# 2. Fat suppression\n# 3. Reasonable slice count\n# 4. Higher spatial resolution\n#\n# We do NOT automatically claim this is the final model choice.\n# This is an audit to identify strong candidate series.\n# ----------------------------------------------------------------\n\nseries_audit[\"resolution_score\"] = (\n series_audit[\"Rows\"].fillna(0)\n *\n series_audit[\"Columns\"].fillna(0)\n)\n\nseries_audit[\"sequence_score\"] = (\n series_audit[\"Fluid_Sensitive\"].fillna(0)\n +\n series_audit[\"Fat_Suppression\"].fillna(0)\n)\n\nseries_audit[\"candidate_score\"] = (\n series_audit[\"sequence_score\"] * 100000000\n +\n series_audit[\"resolution_score\"]\n)\n\n\n# ----------------------------------------------------------------\n# 6. Rank candidates within each study and plane\n# ----------------------------------------------------------------\n\nseries_audit = (\n series_audit\n .sort_values(\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\",\n \"candidate_score\",\n \"slice_count\"\n ],\n ascending=[\n True,\n True,\n False,\n False\n ]\n )\n)\n\nseries_audit[\"rank_within_plane\"] = (\n series_audit\n .groupby(\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\"\n ]\n )\n .cumcount()\n + 1\n)\n\n\n# ----------------------------------------------------------------\n# 7. Show top candidates from labeled studies\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TOP SERIES CANDIDATES\")\nprint(\"=\" * 70)\n\ndisplay_columns = [\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\",\n \"slice_count\",\n \"Rows\",\n \"Columns\",\n \"PixelSpacingX\",\n \"PixelSpacingY\",\n \"SliceThickness\",\n \"SpacingBetweenSlices\",\n \"SeriesDescription\",\n \"rank_within_plane\"\n]\n\nprint(\n series_audit[\n display_columns\n ]\n .head(40)\n .to_string(index=False)\n)\n\n\n# ----------------------------------------------------------------\n# 8. Count what the top-ranked candidate looks like\n# ----------------------------------------------------------------\n\ntop_candidates = series_audit[\n series_audit[\n \"rank_within_plane\"\n ] == 1\n].copy()\n\nprint()\nprint(\"=\" * 70)\nprint(\"TOP-RANKED SERIES DISTRIBUTION\")\nprint(\"=\" * 70)\n\nprint(\n top_candidates[\n [\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\"\n ]\n ]\n .value_counts()\n .to_string()\n)\n\n\n# ----------------------------------------------------------------\n# 9. Check whether every labeled study has a candidate\n# in all three planes\n# ----------------------------------------------------------------\n\ncandidate_plane_counts = (\n top_candidates\n .groupby(\n \"StudyInstanceUID\"\n )[\"Anatomical_Plane\"]\n .nunique()\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"TOP-CANDIDATE THREE-PLANE COVERAGE\")\nprint(\"=\" * 70)\n\nprint(\n \"Studies:\",\n len(candidate_plane_counts)\n)\n\nprint(\n \"Studies with 3 planes:\",\n int(\n (\n candidate_plane_counts == 3\n ).sum()\n )\n)\n\nprint(\n \"Studies missing at least one plane:\",\n int(\n (\n candidate_plane_counts < 3\n ).sum()\n )\n)\n\n\n# ----------------------------------------------------------------\n# 10. Test-set series audit\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TEST SERIES SUMMARY\")\nprint(\"=\" * 70)\n\ntest_summary = (\n test_series\n .groupby(\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\"\n ]\n )\n .size()\n .reset_index(\n name=\"series_count\"\n )\n)\n\nprint(\n test_summary.to_string(\n index=False\n )\n)\n\n\n# ----------------------------------------------------------------\n# 11. Save\n# ----------------------------------------------------------------\n\nOUTPUT_DIR = (\n \"/kaggle/working/rsna_knee_audit\"\n)\n\nos.makedirs(\n OUTPUT_DIR,\n exist_ok=True\n)\n\nseries_audit_path = os.path.join(\n OUTPUT_DIR,\n \"cell27_labeled_series_candidate_audit.csv\"\n)\n\ntop_candidates_path = os.path.join(\n OUTPUT_DIR,\n \"cell27_top_series_candidates.csv\"\n)\n\ntest_summary_path = os.path.join(\n OUTPUT_DIR,\n \"cell27_test_series_summary.csv\"\n)\n\nseries_audit.to_csv(\n series_audit_path,\n index=False\n)\n\ntop_candidates.to_csv(\n top_candidates_path,\n index=False\n)\n\ntest_summary.to_csv(\n test_summary_path,\n index=False\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 27 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"Saved:\")\nprint(series_audit_path)\nprint(top_candidates_path)\nprint(test_summary_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:43:29.151065Z","iopub.execute_input":"2026-08-11T08:43:29.151425Z","iopub.status.idle":"2026-08-11T08:43:32.967956Z","shell.execute_reply.started":"2026-08-11T08:43:29.151361Z","shell.execute_reply":"2026-08-11T08:43:32.966813Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 28 - MODELING ENVIRONMENT AND GPU CHECK\n# ================================================================\n\nimport os\nimport sys\nimport platform\nimport subprocess\nimport importlib.util\n\nprint(\"=\" * 70)\nprint(\"MODELING ENVIRONMENT CHECK\")\nprint(\"=\" * 70)\n\n# ---------------------------------------------------------------\n# Python\n# ---------------------------------------------------------------\n\nprint(\"\\nPython:\")\nprint(sys.version)\n\nprint(\"\\nPlatform:\")\nprint(platform.platform())\n\n\n# ---------------------------------------------------------------\n# PyTorch\n# ---------------------------------------------------------------\n\nprint(\"\\nChecking PyTorch...\")\n\nif importlib.util.find_spec(\"torch\") is None:\n print(\"ERROR: PyTorch is not installed.\")\nelse:\n import torch\n\n print(\"PyTorch version:\", torch.__version__)\n print(\"CUDA available:\", torch.cuda.is_available())\n\n if torch.cuda.is_available():\n\n print(\"CUDA version:\", torch.version.cuda)\n print(\n \"GPU count:\",\n torch.cuda.device_count()\n )\n\n for i in range(torch.cuda.device_count()):\n\n props = torch.cuda.get_device_properties(i)\n\n print()\n print(f\"GPU {i}:\")\n print(\"Name:\", props.name)\n print(\n \"VRAM:\",\n round(\n props.total_memory / (1024 ** 3),\n 2\n ),\n \"GB\"\n )\n\n print(\n \"Compute capability:\",\n f\"{props.major}.{props.minor}\"\n )\n\n else:\n print(\n \"WARNING: CUDA is not available.\"\n )\n\n\n# ---------------------------------------------------------------\n# Important packages\n# ---------------------------------------------------------------\n\npackages = [\n \"numpy\",\n \"pandas\",\n \"pydicom\",\n \"sklearn\",\n \"PIL\",\n \"torch\",\n \"torchvision\"\n]\n\nprint()\nprint(\"=\" * 70)\nprint(\"PACKAGE CHECK\")\nprint(\"=\" * 70)\n\nfor package in packages:\n\n installed = (\n importlib.util.find_spec(package)\n is not None\n )\n\n print(\n f\"{package:15s}: \"\n f\"{'AVAILABLE' if installed else 'MISSING'}\"\n )\n\n\n# ---------------------------------------------------------------\n# Output/storage check\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"STORAGE CHECK\")\nprint(\"=\" * 70)\n\npaths_to_check = [\n \"/kaggle/input\",\n \"/kaggle/working\",\n \"/kaggle/working/rsna_knee_audit\"\n]\n\nfor path in paths_to_check:\n\n if os.path.exists(path):\n\n try:\n\n usage = (\n subprocess.check_output(\n [\"df\", \"-h\", path],\n text=True\n )\n .strip()\n .split(\"\\n\")\n )\n\n print()\n print(path)\n print(usage[-1])\n\n except Exception as e:\n\n print(\n path,\n \"exists, but disk usage \"\n \"could not be read.\"\n )\n\n else:\n\n print(\n path,\n \"does not exist.\"\n )\n\n\n# ---------------------------------------------------------------\n# Existing audit files\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"AUDIT FILES\")\nprint(\"=\" * 70)\n\nAUDIT_DIR = (\n \"/kaggle/working/rsna_knee_audit\"\n)\n\nif os.path.isdir(AUDIT_DIR):\n\n audit_files = sorted(\n os.listdir(AUDIT_DIR)\n )\n\n print(\n \"Audit files found:\",\n len(audit_files)\n )\n\n for filename in audit_files:\n\n print(\n \" -\",\n filename\n )\n\nelse:\n\n print(\n \"Audit directory not found.\"\n )\n\n\n# ---------------------------------------------------------------\n# Final verdict\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 28 COMPLETE\")\nprint(\"=\" * 70)\n\nif (\n importlib.util.find_spec(\"torch\")\n is not None\n):\n\n import torch\n\n if torch.cuda.is_available():\n\n print(\n \"READY: CUDA GPU is available.\"\n )\n\n else:\n\n print(\n \"WARNING: CUDA GPU is not available.\"\n )\n\nelse:\n\n print(\n \"STOP: PyTorch is missing.\"\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:44:53.640288Z","iopub.execute_input":"2026-08-11T08:44:53.640624Z","iopub.status.idle":"2026-08-11T08:44:57.862975Z","shell.execute_reply.started":"2026-08-11T08:44:53.640597Z","shell.execute_reply":"2026-08-11T08:44:57.862053Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 29 - BUILD STUDY-LEVEL MODELING TABLE + 3-FOLD SPLIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nfrom collections import defaultdict\n\n# ---------------------------------------------------------------\n# Paths\n# ---------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nTRAIN_CSV = (\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection/train.csv\"\n)\n\nLABELED_CSV = os.path.join(\n AUDIT_DIR,\n \"explicitly_labeled_studies.csv\"\n)\n\nTOP_CANDIDATES_CSV = os.path.join(\n AUDIT_DIR,\n \"cell27_top_series_candidates.csv\"\n)\n\nOUTPUT_MODEL_TABLE = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nOUTPUT_FOLDS = os.path.join(\n AUDIT_DIR,\n \"cell29_study_folds.csv\"\n)\n\n# ---------------------------------------------------------------\n# Target columns\n# ---------------------------------------------------------------\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n]\n\nPLANES = [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n]\n\nSEED = 42\nN_FOLDS = 3\n\nprint(\"=\" * 70)\nprint(\"CELL 29 - STUDY-LEVEL MODELING TABLE\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Load files\n# ---------------------------------------------------------------\n\nprint(\"\\nLoading training data...\")\n\ntrain = pd.read_csv(TRAIN_CSV)\nlabeled = pd.read_csv(LABELED_CSV)\ncandidates = pd.read_csv(TOP_CANDIDATES_CSV)\n\nprint(\"Train rows:\", len(train))\nprint(\"Labeled rows:\", len(labeled))\nprint(\"Candidate rows:\", len(candidates))\n\n\n# ---------------------------------------------------------------\n# Basic validation\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"BASIC INPUT VALIDATION\")\nprint(\"=\" * 70)\n\nmissing_targets = [\n col for col in TARGETS\n if col not in train.columns\n]\n\nif missing_targets:\n raise ValueError(\n f\"Missing target columns: {missing_targets}\"\n )\n\nif \"StudyInstanceUID\" not in train.columns:\n raise ValueError(\n \"StudyInstanceUID missing from train.csv\"\n )\n\nif \"StudyInstanceUID\" not in candidates.columns:\n raise ValueError(\n \"StudyInstanceUID missing from candidate file\"\n )\n\nif \"SeriesInstanceUID\" not in candidates.columns:\n raise ValueError(\n \"SeriesInstanceUID missing from candidate file\"\n )\n\nprint(\"All target columns present.\")\nprint(\"StudyInstanceUID present.\")\nprint(\"SeriesInstanceUID present.\")\n\n\n# ---------------------------------------------------------------\n# Restrict to explicitly labeled studies\n# ---------------------------------------------------------------\n\nlabeled_ids = (\n labeled[\"StudyInstanceUID\"]\n .astype(str)\n .unique()\n)\n\nprint()\nprint(\"Explicitly labeled studies:\", len(labeled_ids))\n\nlabeled_train = train[\n train[\"StudyInstanceUID\"]\n .astype(str)\n .isin(labeled_ids)\n].copy()\n\nprint(\n \"Matching train studies:\",\n labeled_train[\"StudyInstanceUID\"].nunique()\n)\n\nif labeled_train[\"StudyInstanceUID\"].nunique() != 58:\n raise ValueError(\n \"Expected exactly 58 fully labeled studies.\"\n )\n\n\n# ---------------------------------------------------------------\n# Select top candidate per study + plane\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"SELECTING PRIMARY SERIES\")\nprint(\"=\" * 70)\n\ncandidate_subset = candidates[\n candidates[\"Anatomical_Plane\"].isin(PLANES)\n].copy()\n\ncandidate_subset = candidate_subset.sort_values(\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\",\n \"rank_within_plane\"\n ]\n)\n\nprimary = (\n candidate_subset\n .groupby(\n [\"StudyInstanceUID\", \"Anatomical_Plane\"],\n as_index=False\n )\n .first()\n)\n\nprint(\n \"Primary series rows:\",\n len(primary)\n)\n\n\n# ---------------------------------------------------------------\n# Check plane coverage\n# ---------------------------------------------------------------\n\nplane_counts = (\n primary\n .groupby(\"StudyInstanceUID\")[\"Anatomical_Plane\"]\n .nunique()\n)\n\nprint(\n \"Studies with all 3 primary planes:\",\n int((plane_counts == 3).sum())\n)\n\nprint(\n \"Studies missing at least one primary plane:\",\n int((plane_counts < 3).sum())\n)\n\nmissing_plane_studies = plane_counts[\n plane_counts < 3\n]\n\nif len(missing_plane_studies) > 0:\n\n print(\"\\nWARNING: Missing-plane studies:\")\n\n for study_id in missing_plane_studies.index:\n present = set(\n primary.loc[\n primary[\"StudyInstanceUID\"] == study_id,\n \"Anatomical_Plane\"\n ]\n )\n\n missing = set(PLANES) - present\n\n print(\n study_id,\n \"missing:\",\n sorted(missing)\n )\n\n raise ValueError(\n \"Every labeled study must have all 3 primary planes.\"\n )\n\n\n# ---------------------------------------------------------------\n# Pivot primary series into one study-level row\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"BUILDING STUDY-LEVEL SERIES TABLE\")\nprint(\"=\" * 70)\n\nmodel_table = labeled_train[\n [\"StudyInstanceUID\"] + TARGETS\n].copy()\n\nmodel_table[\"StudyInstanceUID\"] = (\n model_table[\"StudyInstanceUID\"].astype(str)\n)\n\nfor plane in PLANES:\n\n plane_df = primary[\n primary[\"Anatomical_Plane\"] == plane\n ][\n [\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n \"slice_count\",\n \"Rows\",\n \"Columns\",\n \"PixelSpacingX\",\n \"PixelSpacingY\",\n \"SliceThickness\",\n \"SpacingBetweenSlices\",\n \"SeriesDescription\"\n ]\n ].copy()\n\n plane_df = plane_df.rename(\n columns={\n \"SeriesInstanceUID\":\n f\"{plane}_SeriesInstanceUID\",\n \"slice_count\":\n f\"{plane}_slice_count\",\n \"Rows\":\n f\"{plane}_Rows\",\n \"Columns\":\n f\"{plane}_Columns\",\n \"PixelSpacingX\":\n f\"{plane}_PixelSpacingX\",\n \"PixelSpacingY\":\n f\"{plane}_PixelSpacingY\",\n \"SliceThickness\":\n f\"{plane}_SliceThickness\",\n \"SpacingBetweenSlices\":\n f\"{plane}_SpacingBetweenSlices\",\n \"SeriesDescription\":\n f\"{plane}_SeriesDescription\"\n }\n )\n\n model_table = model_table.merge(\n plane_df,\n on=\"StudyInstanceUID\",\n how=\"left\",\n validate=\"one_to_one\"\n )\n\n\n# ---------------------------------------------------------------\n# Final study-level integrity checks\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"STUDY-LEVEL INTEGRITY CHECK\")\nprint(\"=\" * 70)\n\nprint(\n \"Modeling studies:\",\n model_table[\"StudyInstanceUID\"].nunique()\n)\n\nprint(\n \"Modeling rows:\",\n len(model_table)\n)\n\nif len(model_table) != 58:\n raise ValueError(\n \"Modeling table must contain exactly 58 studies.\"\n )\n\nif model_table[\"StudyInstanceUID\"].duplicated().any():\n raise ValueError(\n \"Duplicate StudyInstanceUID detected.\"\n )\n\n\n# Check primary series IDs\nfor plane in PLANES:\n\n col = f\"{plane}_SeriesInstanceUID\"\n\n missing = model_table[col].isna().sum()\n\n unique = model_table[col].nunique()\n\n print(\n f\"{plane}: \"\n f\"missing={missing}, \"\n f\"unique_series={unique}\"\n )\n\n if missing != 0:\n raise ValueError(\n f\"Missing primary {plane} series.\"\n )\n\n\n# ---------------------------------------------------------------\n# Verify labels are binary and complete\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TARGET VALIDATION\")\nprint(\"=\" * 70)\n\nfor target in TARGETS:\n\n values = sorted(\n model_table[target]\n .dropna()\n .unique()\n .tolist()\n )\n\n print(\n f\"{target:20s}: {values}\"\n )\n\n if set(values) != {0, 1}:\n raise ValueError(\n f\"Unexpected values in target {target}: {values}\"\n )\n\n if model_table[target].isna().any():\n raise ValueError(\n f\"Missing labels in target {target}.\"\n )\n\n\n# ---------------------------------------------------------------\n# Create deterministic multilabel-aware folds\n# ---------------------------------------------------------------\n#\n# We use a greedy iterative assignment.\n#\n# Goal:\n# - keep each target's positive/negative counts reasonably\n# balanced across 3 folds\n# - keep each fold approximately equal in size\n# - preserve complete studies as the split unit\n#\n# This is intentionally study-level.\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CREATING 3-FOLD STUDY-LEVEL SPLIT\")\nprint(\"=\" * 70)\n\nrng = np.random.RandomState(SEED)\n\nwork = model_table[\n [\"StudyInstanceUID\"] + TARGETS\n].copy()\n\nwork = work.sample(\n frac=1.0,\n random_state=SEED\n).reset_index(drop=True)\n\nY = work[TARGETS].astype(int).values\n\nn_samples = len(work)\n\n# Desired number of samples per fold\nbase_size = n_samples // N_FOLDS\nremainder = n_samples % N_FOLDS\n\ndesired_sizes = np.array([\n base_size + (1 if i < remainder else 0)\n for i in range(N_FOLDS)\n])\n\n# Desired positive counts per target per fold\ntarget_totals = Y.sum(axis=0)\n\ndesired_positive = (\n target_totals / N_FOLDS\n)\n\nfold_indices = [\n [] for _ in range(N_FOLDS)\n]\n\nfold_positive = np.zeros(\n (N_FOLDS, len(TARGETS)),\n dtype=float\n)\n\nfold_sizes = np.zeros(\n N_FOLDS,\n dtype=int\n)\n\n# Process studies with more abnormalities first.\n# This helps distribute rare multilabel signatures.\nseverity = Y.sum(axis=1)\n\norder = np.argsort(\n -severity +\n rng.uniform(\n 0,\n 0.01,\n size=n_samples\n )\n)\n\nfor idx in order:\n\n sample_labels = Y[idx]\n\n best_fold = None\n best_score = None\n\n for fold in range(N_FOLDS):\n\n if (\n fold_sizes[fold]\n >= desired_sizes[fold]\n ):\n continue\n\n new_positive = (\n fold_positive[fold]\n + sample_labels\n )\n\n # Positive-count deviation\n positive_error = np.sum(\n (\n new_positive\n - desired_positive\n ) ** 2\n )\n\n # Fold-size penalty\n size_error = (\n (\n fold_sizes[fold] + 1\n )\n - desired_sizes[fold]\n ) ** 2\n\n # Small penalty for already-large folds\n load_penalty = (\n fold_sizes[fold]\n / max(1, desired_sizes.max())\n )\n\n score = (\n positive_error\n + 0.25 * size_error\n + 0.05 * load_penalty\n )\n\n if (\n best_score is None\n or score < best_score\n ):\n best_score = score\n best_fold = fold\n\n fold_indices[best_fold].append(idx)\n\n fold_positive[best_fold] += sample_labels\n fold_sizes[best_fold] += 1\n\n\n# ---------------------------------------------------------------\n# Convert assignments back to StudyInstanceUID\n# ---------------------------------------------------------------\n\nfold_assignment = {}\n\nfor fold in range(N_FOLDS):\n\n for idx in fold_indices[fold]:\n\n study_id = work.loc[\n idx,\n \"StudyInstanceUID\"\n ]\n\n fold_assignment[\n study_id\n ] = fold\n\n\nmodel_table[\"fold\"] = (\n model_table[\"StudyInstanceUID\"]\n .map(fold_assignment)\n .astype(int)\n)\n\n\n# ---------------------------------------------------------------\n# Fold diagnostics\n# ---------------------------------------------------------------\n\nprint()\n\nfor fold in range(N_FOLDS):\n\n fold_df = model_table[\n model_table[\"fold\"] == fold\n ]\n\n print(\n f\"Fold {fold + 1}: \"\n f\"{len(fold_df)} studies\"\n )\n\n for target in TARGETS:\n\n pos = int(\n fold_df[target].sum()\n )\n\n neg = int(\n len(fold_df) - pos\n )\n\n print(\n f\" {target:20s} \"\n f\"pos={pos:2d} \"\n f\"neg={neg:2d}\"\n )\n\n print()\n\n\n# ---------------------------------------------------------------\n# Verify every fold has both classes\n# ---------------------------------------------------------------\n\nprint(\"=\" * 70)\nprint(\"CLASS COVERAGE CHECK\")\nprint(\"=\" * 70)\n\nbad = []\n\nfor fold in range(N_FOLDS):\n\n fold_df = model_table[\n model_table[\"fold\"] == fold\n ]\n\n for target in TARGETS:\n\n pos = int(\n fold_df[target].sum()\n )\n\n neg = int(\n len(fold_df) - pos\n )\n\n if pos == 0 or neg == 0:\n\n bad.append(\n (\n fold,\n target,\n pos,\n neg\n )\n )\n\nif bad:\n\n print(\n \"WARNING: Some target/fold combinations \"\n \"do not contain both classes.\"\n )\n\n for item in bad:\n print(item)\n\nelse:\n\n print(\n \"PASS: Every target has both classes \"\n \"in every fold.\"\n )\n\n\n# ---------------------------------------------------------------\n# Verify no study leakage\n# ---------------------------------------------------------------\n\nfold_counts = (\n model_table\n .groupby(\"StudyInstanceUID\")[\"fold\"]\n .nunique()\n)\n\nif (fold_counts != 1).any():\n\n raise ValueError(\n \"Study leakage detected across folds.\"\n )\n\nprint(\n \"PASS: Every study belongs to exactly one fold.\"\n)\n\n\n# ---------------------------------------------------------------\n# Save\n# ---------------------------------------------------------------\n\nmodel_table.to_csv(\n OUTPUT_MODEL_TABLE,\n index=False\n)\n\nmodel_table[\n [\"StudyInstanceUID\", \"fold\"]\n].to_csv(\n OUTPUT_FOLDS,\n index=False\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 29 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Saved:\",\n OUTPUT_MODEL_TABLE\n)\n\nprint(\n \"Saved:\",\n OUTPUT_FOLDS\n)\n\nprint()\nprint(\n \"Modeling table shape:\",\n model_table.shape\n)\n\nprint(\n \"Fold distribution:\"\n)\n\nprint(\n model_table[\"fold\"]\n .value_counts()\n .sort_index()\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:46:48.768813Z","iopub.execute_input":"2026-08-11T08:46:48.769625Z","iopub.status.idle":"2026-08-11T08:46:48.966949Z","shell.execute_reply.started":"2026-08-11T08:46:48.769584Z","shell.execute_reply":"2026-08-11T08:46:48.965996Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 29B - ROBUST MULTILABEL 3-FOLD SPLIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\n# ---------------------------------------------------------------\n# Configuration\n# ---------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nMODEL_TABLE_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nOUTPUT_MODEL_TABLE = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nOUTPUT_FOLDS = os.path.join(\n AUDIT_DIR,\n \"cell29_study_folds.csv\"\n)\n\nSEED = 42\nN_FOLDS = 3\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n]\n\nprint(\"=\" * 70)\nprint(\"CELL 29B - ROBUST MULTILABEL 3-FOLD SPLIT\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Load existing modeling table\n# ---------------------------------------------------------------\n\nmodel_table = pd.read_csv(\n MODEL_TABLE_PATH\n)\n\nprint(\n \"\\nLoaded modeling table:\",\n model_table.shape\n)\n\nif len(model_table) != 58:\n raise ValueError(\n f\"Expected 58 labeled studies, got {len(model_table)}\"\n )\n\nif model_table[\"StudyInstanceUID\"].duplicated().any():\n raise ValueError(\n \"Duplicate StudyInstanceUID detected.\"\n )\n\n\n# ---------------------------------------------------------------\n# Prepare labels\n# ---------------------------------------------------------------\n\nY = (\n model_table[TARGETS]\n .astype(int)\n .values\n)\n\nn_samples = len(model_table)\nn_targets = len(TARGETS)\n\nif n_samples != 58:\n raise ValueError(\n \"Expected exactly 58 labeled studies.\"\n )\n\n\n# ---------------------------------------------------------------\n# Objective function\n# ---------------------------------------------------------------\n#\n# We want:\n#\n# 1. Every fold to have both classes for every target.\n# 2. Positive counts to be close to target_total / 3.\n# 3. Fold sizes close to 58 / 3.\n#\n# Invalid folds receive a very large penalty.\n# ---------------------------------------------------------------\n\ntarget_totals = Y.sum(axis=0)\n\ndesired_positive = (\n target_totals / N_FOLDS\n)\n\ndesired_size = (\n n_samples / N_FOLDS\n)\n\n\ndef evaluate_assignment(\n assignment\n):\n\n assignment = np.asarray(\n assignment\n )\n\n score = 0.0\n\n fold_sizes = np.zeros(\n N_FOLDS,\n dtype=int\n )\n\n fold_positive = np.zeros(\n (N_FOLDS, n_targets),\n dtype=int\n )\n\n for i in range(n_samples):\n\n fold = assignment[i]\n\n fold_sizes[fold] += 1\n fold_positive[fold] += Y[i]\n\n # -----------------------------------------------------------\n # Fold-size balance\n # -----------------------------------------------------------\n\n score += np.sum(\n (\n fold_sizes\n - desired_size\n ) ** 2\n ) * 2.0\n\n # -----------------------------------------------------------\n # Positive-count balance\n # -----------------------------------------------------------\n\n for fold in range(N_FOLDS):\n\n diff = (\n fold_positive[fold]\n - desired_positive\n )\n\n # Normalize by target frequency so rare targets\n # receive appropriate importance.\n normalized = (\n diff ** 2\n / np.maximum(\n desired_positive,\n 1.0\n )\n )\n\n score += np.sum(\n normalized\n )\n\n # -----------------------------------------------------------\n # Hard class-coverage constraints\n # -----------------------------------------------------------\n\n invalid_count = 0\n\n for fold in range(N_FOLDS):\n\n positives = fold_positive[fold]\n\n fold_size = fold_sizes[fold]\n\n negatives = (\n fold_size - positives\n )\n\n invalid_count += np.sum(\n positives == 0\n )\n\n invalid_count += np.sum(\n negatives == 0\n )\n\n # Huge penalty for invalid target/fold combinations.\n score += (\n invalid_count * 10000.0\n )\n\n return (\n score,\n fold_sizes,\n fold_positive\n )\n\n\n# ---------------------------------------------------------------\n# Randomized local-search optimization\n# ---------------------------------------------------------------\n\nrng = np.random.RandomState(\n SEED\n)\n\nbest_assignment = None\nbest_score = np.inf\nbest_details = None\n\nN_RESTARTS = 300\nMAX_ITERATIONS = 5000\n\nprint()\nprint(\n \"Searching for a valid multilabel split...\"\n)\n\nfor restart in range(N_RESTARTS):\n\n # Start with balanced fold assignment.\n assignment = np.arange(\n n_samples\n ) % N_FOLDS\n\n rng.shuffle(\n assignment\n )\n\n current_score, _, _ = evaluate_assignment(\n assignment\n )\n\n for iteration in range(\n MAX_ITERATIONS\n ):\n\n improved = False\n\n # Randomly inspect study pairs.\n pairs = rng.randint(\n 0,\n n_samples,\n size=(80, 2)\n )\n\n for i, j in pairs:\n\n if i == j:\n continue\n\n if assignment[i] == assignment[j]:\n continue\n\n old_i = assignment[i]\n old_j = assignment[j]\n\n assignment[i] = old_j\n assignment[j] = old_i\n\n new_score, _, _ = evaluate_assignment(\n assignment\n )\n\n if new_score < current_score:\n\n current_score = new_score\n improved = True\n\n else:\n\n assignment[i] = old_i\n assignment[j] = old_j\n\n if not improved:\n break\n\n final_score, fold_sizes, fold_positive = (\n evaluate_assignment(\n assignment\n )\n )\n\n if final_score < best_score:\n\n best_score = final_score\n best_assignment = assignment.copy()\n\n best_details = (\n fold_sizes.copy(),\n fold_positive.copy()\n )\n\n # Stop immediately once a valid and reasonably balanced\n # solution is found.\n if (\n np.all(\n best_details[1] > 0\n )\n and np.all(\n best_details[1]\n < best_details[0][:, None]\n )\n ):\n\n # We have a valid solution.\n if restart >= 10:\n break\n\n\n# ---------------------------------------------------------------\n# Safety check\n# ---------------------------------------------------------------\n\nif best_assignment is None:\n raise RuntimeError(\n \"Could not generate a fold assignment.\"\n )\n\nfold_sizes, fold_positive = best_details\n\nprint()\nprint(\n \"Best objective score:\",\n best_score\n)\n\nprint(\n \"Fold sizes:\",\n fold_sizes.tolist()\n)\n\n\n# ---------------------------------------------------------------\n# Validate class coverage\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL FOLD DIAGNOSTICS\")\nprint(\"=\" * 70)\n\ninvalid = []\n\nfor fold in range(N_FOLDS):\n\n positives = fold_positive[fold]\n\n negatives = (\n fold_sizes[fold]\n - positives\n )\n\n print()\n print(\n f\"Fold {fold + 1}: \"\n f\"{fold_sizes[fold]} studies\"\n )\n\n for t, target in enumerate(TARGETS):\n\n pos = int(\n positives[t]\n )\n\n neg = int(\n negatives[t]\n )\n\n print(\n f\" {target:20s} \"\n f\"pos={pos:2d} \"\n f\"neg={neg:2d}\"\n )\n\n if pos == 0 or neg == 0:\n\n invalid.append(\n (\n fold,\n target,\n pos,\n neg\n )\n )\n\n\n# ---------------------------------------------------------------\n# Apply fold assignment\n# ---------------------------------------------------------------\n\nmodel_table[\"fold\"] = (\n best_assignment\n)\n\n\n# ---------------------------------------------------------------\n# Verify no study leakage\n# ---------------------------------------------------------------\n\nfold_counts = (\n model_table\n .groupby(\n \"StudyInstanceUID\"\n )[\"fold\"]\n .nunique()\n)\n\nif (fold_counts != 1).any():\n\n raise ValueError(\n \"A study appears in multiple folds.\"\n )\n\n\n# ---------------------------------------------------------------\n# Final verdict\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CLASS COVERAGE VERDICT\")\nprint(\"=\" * 70)\n\nif invalid:\n\n print(\n \"FAIL: Invalid target/fold combinations:\"\n )\n\n for item in invalid:\n print(item)\n\n raise RuntimeError(\n \"Could not construct a valid 3-fold split.\"\n )\n\nelse:\n\n print(\n \"PASS: Every target has both positive \"\n \"and negative samples in every fold.\"\n )\n\nprint(\n \"PASS: Every study belongs to exactly one fold.\"\n)\n\n\n# ---------------------------------------------------------------\n# Save\n# ---------------------------------------------------------------\n\nmodel_table.to_csv(\n OUTPUT_MODEL_TABLE,\n index=False\n)\n\nmodel_table[\n [\n \"StudyInstanceUID\",\n \"fold\"\n ]\n].to_csv(\n OUTPUT_FOLDS,\n index=False\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 29B COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Saved:\",\n OUTPUT_MODEL_TABLE\n)\n\nprint(\n \"Saved:\",\n OUTPUT_FOLDS\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:47:40.660563Z","iopub.execute_input":"2026-08-11T08:47:40.660949Z","iopub.status.idle":"2026-08-11T08:47:41.196402Z","shell.execute_reply.started":"2026-08-11T08:47:40.660908Z","shell.execute_reply":"2026-08-11T08:47:41.195463Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 30 - ROBUST DICOM SERIES LOADER\n# ================================================================\n\nimport os\nimport gc\nimport numpy as np\nimport pandas as pd\nimport pydicom\nimport torch\n\nfrom PIL import Image\n\n\n# ---------------------------------------------------------------\n# Configuration\n# ---------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nTRAIN_SERIES_DIR = (\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection/train_series\"\n)\n\nMODEL_TABLE_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nIMAGE_SIZE = 224\nNUM_SLICES = 5\n\nPLANES = [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n]\n\nSEED = 42\n\nprint(\"=\" * 70)\nprint(\"CELL 30 - DICOM MODEL INPUT LOADER\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Load modeling table\n# ---------------------------------------------------------------\n\nmodel_table = pd.read_csv(\n MODEL_TABLE_PATH\n)\n\nprint(\n \"\\nModeling studies:\",\n len(model_table)\n)\n\nprint(\n \"Image size:\",\n IMAGE_SIZE\n)\n\nprint(\n \"Slices per series:\",\n NUM_SLICES\n)\n\nprint(\n \"Planes:\",\n PLANES\n)\n\n\n# ---------------------------------------------------------------\n# Utility: find DICOM files\n# ---------------------------------------------------------------\n\ndef get_dicom_files(series_path):\n \"\"\"\n Return DICOM files from a series directory.\n \"\"\"\n\n if not os.path.isdir(series_path):\n return []\n\n files = []\n\n for filename in os.listdir(series_path):\n\n path = os.path.join(\n series_path,\n filename\n )\n\n if os.path.isfile(path):\n files.append(path)\n\n return files\n\n\n# ---------------------------------------------------------------\n# Utility: calculate slice position\n# ---------------------------------------------------------------\n\ndef get_slice_position(ds):\n \"\"\"\n Calculate the physical position of a DICOM slice.\n\n Uses ImagePositionPatient and\n ImageOrientationPatient when available.\n \"\"\"\n\n position = getattr(\n ds,\n \"ImagePositionPatient\",\n None\n )\n\n orientation = getattr(\n ds,\n \"ImageOrientationPatient\",\n None\n )\n\n if (\n position is not None\n and orientation is not None\n and len(position) >= 3\n and len(orientation) >= 6\n ):\n\n try:\n\n row_cosines = np.asarray(\n orientation[:3],\n dtype=np.float64\n )\n\n col_cosines = np.asarray(\n orientation[3:6],\n dtype=np.float64\n )\n\n normal = np.cross(\n row_cosines,\n col_cosines\n )\n\n position = np.asarray(\n position,\n dtype=np.float64\n )\n\n return float(\n np.dot(\n position,\n normal\n )\n )\n\n except Exception:\n pass\n\n # Fallback to InstanceNumber\n instance = getattr(\n ds,\n \"InstanceNumber\",\n None\n )\n\n if instance is not None:\n\n try:\n return float(instance)\n except Exception:\n pass\n\n return 0.0\n\n\n# ---------------------------------------------------------------\n# Utility: sort DICOM files\n# ---------------------------------------------------------------\n\ndef sort_dicom_files(dicom_files):\n \"\"\"\n Sort DICOM files using physical slice position.\n \"\"\"\n\n records = []\n\n for path in dicom_files:\n\n try:\n\n ds = pydicom.dcmread(\n path,\n stop_before_pixels=True,\n force=True\n )\n\n position = get_slice_position(\n ds\n )\n\n records.append(\n (\n position,\n path\n )\n )\n\n except Exception:\n continue\n\n records.sort(\n key=lambda x: x[0]\n )\n\n return [\n path\n for _, path in records\n ]\n\n\n# ---------------------------------------------------------------\n# Utility: robust intensity normalization\n# ---------------------------------------------------------------\n\ndef normalize_dicom_image(\n pixel_array\n):\n \"\"\"\n Robust percentile-based normalization.\n\n The purpose is to reduce sensitivity to\n different MRI intensity scales.\n \"\"\"\n\n image = np.asarray(\n pixel_array,\n dtype=np.float32\n )\n\n image = np.nan_to_num(\n image,\n nan=0.0,\n posinf=0.0,\n neginf=0.0\n )\n\n # Remove extreme values.\n low = np.percentile(\n image,\n 1\n )\n\n high = np.percentile(\n image,\n 99\n )\n\n if high <= low:\n\n low = float(\n image.min()\n )\n\n high = float(\n image.max()\n )\n\n if high <= low:\n\n return np.zeros_like(\n image,\n dtype=np.float32\n )\n\n image = np.clip(\n image,\n low,\n high\n )\n\n image = (\n image - low\n ) / (\n high - low\n )\n\n return image.astype(\n np.float32\n )\n\n\n# ---------------------------------------------------------------\n# Utility: resize image\n# ---------------------------------------------------------------\n\ndef resize_image(\n image,\n size=IMAGE_SIZE\n):\n \"\"\"\n Resize normalized image to model resolution.\n \"\"\"\n\n image_uint8 = (\n np.clip(\n image,\n 0.0,\n 1.0\n )\n * 255.0\n ).astype(\n np.uint8\n )\n\n pil_image = Image.fromarray(\n image_uint8\n )\n\n pil_image = pil_image.resize(\n (size, size),\n resample=Image.BILINEAR\n )\n\n output = np.asarray(\n pil_image,\n dtype=np.float32\n ) / 255.0\n\n return output\n\n\n# ---------------------------------------------------------------\n# Select representative slice indices\n# ---------------------------------------------------------------\n\ndef select_slice_indices(\n n_slices,\n num_slices=NUM_SLICES\n):\n \"\"\"\n Select approximately uniformly spaced slices.\n \"\"\"\n\n if n_slices <= 0:\n return []\n\n if n_slices <= num_slices:\n\n return list(\n range(n_slices)\n )\n\n indices = np.linspace(\n 0,\n n_slices - 1,\n num=num_slices\n )\n\n indices = np.round(\n indices\n ).astype(int)\n\n return list(\n np.unique(indices)\n )\n\n\n# ---------------------------------------------------------------\n# Load a single series\n# ---------------------------------------------------------------\n\ndef load_series(\n study_id,\n series_id,\n num_slices=NUM_SLICES,\n image_size=IMAGE_SIZE\n):\n \"\"\"\n Load representative slices from one DICOM series.\n\n Returns:\n tensor-like numpy array:\n [N, H, W]\n \"\"\"\n\n series_path = os.path.join(\n TRAIN_SERIES_DIR,\n str(study_id),\n str(series_id)\n )\n\n files = get_dicom_files(\n series_path\n )\n\n if len(files) == 0:\n\n raise FileNotFoundError(\n f\"No DICOM files found: \"\n f\"{series_path}\"\n )\n\n sorted_files = sort_dicom_files(\n files\n )\n\n if len(sorted_files) == 0:\n\n raise RuntimeError(\n f\"Could not sort/read DICOM files: \"\n f\"{series_path}\"\n )\n\n selected_indices = (\n select_slice_indices(\n len(sorted_files),\n num_slices\n )\n )\n\n images = []\n\n for index in selected_indices:\n\n path = sorted_files[index]\n\n try:\n\n ds = pydicom.dcmread(\n path,\n force=True\n )\n\n if not hasattr(\n ds,\n \"PixelData\"\n ):\n\n continue\n\n image = ds.pixel_array\n\n # Apply DICOM rescaling when available.\n slope = float(\n getattr(\n ds,\n \"RescaleSlope\",\n 1.0\n )\n )\n\n intercept = float(\n getattr(\n ds,\n \"RescaleIntercept\",\n 0.0\n )\n )\n\n image = (\n image.astype(\n np.float32\n )\n * slope\n + intercept\n )\n\n # Handle MONOCHROME1 safely.\n photometric = getattr(\n ds,\n \"PhotometricInterpretation\",\n \"MONOCHROME2\"\n )\n\n if photometric == \"MONOCHROME1\":\n\n image = (\n image.max()\n - image\n )\n\n image = normalize_dicom_image(\n image\n )\n\n image = resize_image(\n image,\n image_size\n )\n\n images.append(\n image\n )\n\n except Exception as e:\n\n print(\n \"WARNING: Failed to read:\",\n path,\n \"|\",\n type(e).__name__\n )\n\n if len(images) == 0:\n\n raise RuntimeError(\n f\"No usable slices found for \"\n f\"{study_id} / {series_id}\"\n )\n\n return np.stack(\n images,\n axis=0\n ).astype(\n np.float32\n )\n\n\n# ---------------------------------------------------------------\n# Test loader on one study\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TESTING LOADER\")\nprint(\"=\" * 70)\n\ntest_row = model_table.iloc[0]\n\nstudy_id = str(\n test_row[\"StudyInstanceUID\"]\n)\n\nprint(\n \"Test study:\",\n study_id\n)\n\nloaded_series = {}\n\nfor plane in PLANES:\n\n series_id = str(\n test_row[\n f\"{plane}_SeriesInstanceUID\"\n ]\n )\n\n print()\n print(\n f\"Loading {plane}:\"\n )\n\n print(\n \"Series:\",\n series_id\n )\n\n volume = load_series(\n study_id,\n series_id\n )\n\n loaded_series[\n plane\n ] = volume\n\n print(\n \"Loaded shape:\",\n volume.shape\n )\n\n print(\n \"Min:\",\n float(volume.min())\n )\n\n print(\n \"Max:\",\n float(volume.max())\n )\n\n print(\n \"Mean:\",\n float(volume.mean())\n )\n\n print(\n \"Std:\",\n float(volume.std())\n )\n\n\n# ---------------------------------------------------------------\n# Combine all planes\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"COMBINING PRIMARY PLANES\")\nprint(\"=\" * 70)\n\ncombined = np.concatenate(\n [\n loaded_series[\"Sagittal\"],\n loaded_series[\"Coronal\"],\n loaded_series[\"Axial\"]\n ],\n axis=0\n)\n\nprint(\n \"Combined shape:\",\n combined.shape\n)\n\nprint(\n \"Expected approximately:\",\n NUM_SLICES * len(PLANES),\n \"slices\"\n)\n\nprint(\n \"Tensor memory:\",\n round(\n combined.nbytes\n / (1024 ** 2),\n 4\n ),\n \"MB\"\n)\n\n\n# ---------------------------------------------------------------\n# Convert to PyTorch tensor\n# ---------------------------------------------------------------\n\ntensor = torch.from_numpy(\n combined\n)\n\nprint(\n \"PyTorch tensor shape:\",\n tuple(tensor.shape)\n)\n\nprint(\n \"PyTorch dtype:\",\n tensor.dtype\n)\n\n\n# ---------------------------------------------------------------\n# Cleanup\n# ---------------------------------------------------------------\n\ndel loaded_series\ndel combined\ndel tensor\n\ngc.collect()\n\n\n# ---------------------------------------------------------------\n# Final\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 30 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"DICOM loader test completed successfully.\"\n)\n\nprint(\n \"No full dataset was loaded into memory.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:49:06.321099Z","iopub.execute_input":"2026-08-11T08:49:06.321393Z","iopub.status.idle":"2026-08-11T08:49:06.917031Z","shell.execute_reply.started":"2026-08-11T08:49:06.321359Z","shell.execute_reply":"2026-08-11T08:49:06.916223Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 31 - VISUALIZE EXACT MODEL INPUTS\n# ================================================================\n\nimport os\nimport math\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\n# ---------------------------------------------------------------\n# Configuration\n# ---------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nMODEL_TABLE_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nOUTPUT_IMAGE = os.path.join(\n AUDIT_DIR,\n \"cell31_model_input_contact_sheet.png\"\n)\n\nIMAGE_SIZE = 224\nNUM_SLICES = 5\n\nPLANES = [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n]\n\nprint(\"=\" * 70)\nprint(\"CELL 31 - MODEL INPUT VISUALIZATION\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Load modeling table\n# ---------------------------------------------------------------\n\nmodel_table = pd.read_csv(\n MODEL_TABLE_PATH\n)\n\ntest_row = model_table.iloc[0]\n\nstudy_id = str(\n test_row[\"StudyInstanceUID\"]\n)\n\nprint(\"\\nStudy:\")\nprint(study_id)\n\n\n# ---------------------------------------------------------------\n# Load series using Cell 30 loader\n# ---------------------------------------------------------------\n\nloaded = {}\n\nfor plane in PLANES:\n\n series_id = str(\n test_row[\n f\"{plane}_SeriesInstanceUID\"\n ]\n )\n\n print(\n f\"\\nLoading {plane}:\"\n )\n\n volume = load_series(\n study_id,\n series_id,\n num_slices=NUM_SLICES,\n image_size=IMAGE_SIZE\n )\n\n loaded[plane] = volume\n\n print(\n \"Shape:\",\n volume.shape\n )\n\n\n# ---------------------------------------------------------------\n# Create contact sheet\n# ---------------------------------------------------------------\n\nfig, axes = plt.subplots(\n len(PLANES),\n NUM_SLICES,\n figsize=(15, 9)\n)\n\nif len(PLANES) == 1:\n axes = np.expand_dims(\n axes,\n axis=0\n )\n\nfor row_idx, plane in enumerate(PLANES):\n\n volume = loaded[plane]\n\n for col_idx in range(NUM_SLICES):\n\n ax = axes[\n row_idx,\n col_idx\n ]\n\n ax.imshow(\n volume[col_idx],\n cmap=\"gray\",\n vmin=0,\n vmax=1\n )\n\n ax.axis(\"off\")\n\n if col_idx == 0:\n\n ax.set_title(\n plane,\n fontsize=12\n )\n\n else:\n\n ax.set_title(\n f\"Slice {col_idx + 1}\",\n fontsize=10\n )\n\n\nfig.suptitle(\n \"Exact Model Inputs - 5 Slices per Primary Plane\",\n fontsize=15\n)\n\nplt.tight_layout()\n\nplt.savefig(\n OUTPUT_IMAGE,\n dpi=150,\n bbox_inches=\"tight\"\n)\n\nplt.show()\n\nplt.close(fig)\n\n\n# ---------------------------------------------------------------\n# Print numerical summary\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"INPUT SUMMARY\")\nprint(\"=\" * 70)\n\nfor plane in PLANES:\n\n volume = loaded[plane]\n\n print(\n f\"{plane:10s} | \"\n f\"shape={volume.shape} | \"\n f\"min={volume.min():.4f} | \"\n f\"max={volume.max():.4f} | \"\n f\"mean={volume.mean():.4f} | \"\n f\"std={volume.std():.4f}\"\n )\n\n\n# ---------------------------------------------------------------\n# Save summary\n# ---------------------------------------------------------------\n\nsummary_rows = []\n\nfor plane in PLANES:\n\n volume = loaded[plane]\n\n summary_rows.append({\n \"StudyInstanceUID\": study_id,\n \"Plane\": plane,\n \"num_slices\": volume.shape[0],\n \"height\": volume.shape[1],\n \"width\": volume.shape[2],\n \"min\": float(volume.min()),\n \"max\": float(volume.max()),\n \"mean\": float(volume.mean()),\n \"std\": float(volume.std())\n })\n\nsummary_df = pd.DataFrame(\n summary_rows\n)\n\nsummary_path = os.path.join(\n AUDIT_DIR,\n \"cell31_model_input_summary.csv\"\n)\n\nsummary_df.to_csv(\n summary_path,\n index=False\n)\n\n\n# ---------------------------------------------------------------\n# Final\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 31 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Saved visualization:\",\n OUTPUT_IMAGE\n)\n\nprint(\n \"Saved summary:\",\n summary_path\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:50:14.870825Z","iopub.execute_input":"2026-08-11T08:50:14.871412Z","iopub.status.idle":"2026-08-11T08:50:18.166582Z","shell.execute_reply.started":"2026-08-11T08:50:14.871381Z","shell.execute_reply":"2026-08-11T08:50:18.165696Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 32 - SLICE SAMPLING STRATEGY AUDIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport pydicom\n\n\n# ---------------------------------------------------------------\n# Configuration\n# ---------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nMODEL_TABLE_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nIMAGE_SIZE = 224\n\nPLANES = [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n]\n\n# Number of studies to visually inspect.\nNUM_STUDIES = 3\n\n# Number of slices to select.\nNUM_SLICES = 5\n\nprint(\"=\" * 70)\nprint(\"CELL 32 - SLICE SAMPLING STRATEGY AUDIT\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Sampling strategies\n# ---------------------------------------------------------------\n\ndef uniform_indices(\n n,\n k=5\n):\n \"\"\"\n Current strategy used in Cell 30.\n \"\"\"\n\n if n <= k:\n return list(range(n))\n\n indices = np.linspace(\n 0,\n n - 1,\n k\n )\n\n return list(\n np.unique(\n np.round(indices).astype(int)\n )\n )\n\n\ndef center_indices(\n n,\n k=5,\n center_fraction=0.50\n):\n \"\"\"\n Select k slices from the central portion\n of the series.\n \"\"\"\n\n if n <= k:\n return list(range(n))\n\n width = int(\n round(\n n * center_fraction\n )\n )\n\n width = max(\n width,\n k\n )\n\n start = max(\n 0,\n (n - width) // 2\n )\n\n end = min(\n n - 1,\n start + width - 1\n )\n\n indices = np.linspace(\n start,\n end,\n k\n )\n\n return list(\n np.unique(\n np.round(indices).astype(int)\n )\n )\n\n\ndef hybrid_indices(\n n,\n k=5,\n margin_fraction=0.10\n):\n \"\"\"\n Avoid extreme ends while retaining a wider\n portion of the anatomical coverage.\n \"\"\"\n\n if n <= k:\n return list(range(n))\n\n margin = int(\n round(\n n * margin_fraction\n )\n )\n\n start = margin\n end = n - 1 - margin\n\n if end <= start:\n return uniform_indices(\n n,\n k\n )\n\n indices = np.linspace(\n start,\n end,\n k\n )\n\n return list(\n np.unique(\n np.round(indices).astype(int)\n )\n )\n\n\n# ---------------------------------------------------------------\n# DICOM file discovery\n# ---------------------------------------------------------------\n\ndef get_sorted_dicom_files(\n study_id,\n series_id\n):\n \"\"\"\n Return DICOM files sorted using physical\n slice position.\n \"\"\"\n\n series_path = os.path.join(\n TRAIN_SERIES_DIR,\n str(study_id),\n str(series_id)\n )\n\n files = []\n\n for filename in os.listdir(\n series_path\n ):\n\n path = os.path.join(\n series_path,\n filename\n )\n\n if os.path.isfile(path):\n files.append(path)\n\n records = []\n\n for path in files:\n\n try:\n\n ds = pydicom.dcmread(\n path,\n stop_before_pixels=True,\n force=True\n )\n\n position = get_slice_position(\n ds\n )\n\n records.append(\n (\n position,\n path\n )\n )\n\n except Exception:\n continue\n\n records.sort(\n key=lambda x: x[0]\n )\n\n return [\n path\n for _, path in records\n ]\n\n\n# ---------------------------------------------------------------\n# Load selected slices\n# ---------------------------------------------------------------\n\ndef load_selected_images(\n study_id,\n series_id,\n indices\n):\n \"\"\"\n Load and normalize selected DICOM slices.\n \"\"\"\n\n files = get_sorted_dicom_files(\n study_id,\n series_id\n )\n\n images = []\n\n for index in indices:\n\n if index >= len(files):\n continue\n\n try:\n\n ds = pydicom.dcmread(\n files[index],\n force=True\n )\n\n image = ds.pixel_array.astype(\n np.float32\n )\n\n slope = float(\n getattr(\n ds,\n \"RescaleSlope\",\n 1.0\n )\n )\n\n intercept = float(\n getattr(\n ds,\n \"RescaleIntercept\",\n 0.0\n )\n )\n\n image = (\n image * slope\n + intercept\n )\n\n photometric = getattr(\n ds,\n \"PhotometricInterpretation\",\n \"MONOCHROME2\"\n )\n\n if photometric == \"MONOCHROME1\":\n\n image = (\n image.max()\n - image\n )\n\n image = normalize_dicom_image(\n image\n )\n\n image = resize_image(\n image,\n IMAGE_SIZE\n )\n\n images.append(\n image\n )\n\n except Exception:\n continue\n\n return images, len(files)\n\n\n# ---------------------------------------------------------------\n# Load modeling table\n# ---------------------------------------------------------------\n\nmodel_table = pd.read_csv(\n MODEL_TABLE_PATH\n)\n\nprint(\n \"\\nTotal modeling studies:\",\n len(model_table)\n)\n\nstudy_rows = model_table.iloc[\n :NUM_STUDIES\n]\n\n\n# ---------------------------------------------------------------\n# Compare strategies\n# ---------------------------------------------------------------\n\nstrategies = {\n \"Uniform\": uniform_indices,\n \"Center\": center_indices,\n \"Hybrid\": hybrid_indices\n}\n\nresults = []\n\n\nfor study_number, (_, row) in enumerate(\n study_rows.iterrows(),\n start=1\n):\n\n study_id = str(\n row[\"StudyInstanceUID\"]\n )\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n f\"Study {study_number}:\"\n )\n\n print(\n study_id\n )\n\n for plane in PLANES:\n\n series_id = str(\n row[\n f\"{plane}_SeriesInstanceUID\"\n ]\n )\n\n files = get_sorted_dicom_files(\n study_id,\n series_id\n )\n\n n = len(files)\n\n print()\n print(\n f\"{plane} | total slices = {n}\"\n )\n\n for strategy_name, strategy_function in strategies.items():\n\n indices = strategy_function(\n n,\n NUM_SLICES\n )\n\n results.append({\n \"StudyInstanceUID\": study_id,\n \"Plane\": plane,\n \"Strategy\": strategy_name,\n \"TotalSlices\": n,\n \"SelectedIndices\": \",\".join(\n map(\n str,\n indices\n )\n )\n })\n\n print(\n f\"{strategy_name:8s}:\",\n indices\n )\n\n\n# ---------------------------------------------------------------\n# Create visual comparison\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CREATING VISUAL COMPARISON\")\nprint(\"=\" * 70)\n\n\n# Use first study and all three planes.\nrow = study_rows.iloc[0]\n\nstudy_id = str(\n row[\"StudyInstanceUID\"]\n)\n\n\nfig, axes = plt.subplots(\n len(PLANES) * 3,\n NUM_SLICES,\n figsize=(15, 24)\n)\n\n\nfor plane_idx, plane in enumerate(\n PLANES\n):\n\n series_id = str(\n row[\n f\"{plane}_SeriesInstanceUID\"\n ]\n )\n\n files = get_sorted_dicom_files(\n study_id,\n series_id\n )\n\n n = len(files)\n\n for strategy_idx, (\n strategy_name,\n strategy_function\n ) in enumerate(\n strategies.items()\n ):\n\n row_idx = (\n plane_idx * 3\n + strategy_idx\n )\n\n indices = strategy_function(\n n,\n NUM_SLICES\n )\n\n images, _ = load_selected_images(\n study_id,\n series_id,\n indices\n )\n\n for col_idx in range(\n NUM_SLICES\n ):\n\n ax = axes[\n row_idx,\n col_idx\n ]\n\n ax.axis(\"off\")\n\n if col_idx < len(images):\n\n ax.imshow(\n images[col_idx],\n cmap=\"gray\",\n vmin=0,\n vmax=1\n )\n\n if col_idx == 0:\n\n ax.set_ylabel(\n f\"{plane}\\n{strategy_name}\",\n fontsize=11\n )\n\n\nfig.suptitle(\n \"Slice Sampling Comparison\",\n fontsize=16\n)\n\nplt.tight_layout()\n\nOUTPUT_IMAGE = os.path.join(\n AUDIT_DIR,\n \"cell32_slice_sampling_comparison.png\"\n)\n\nplt.savefig(\n OUTPUT_IMAGE,\n dpi=150,\n bbox_inches=\"tight\"\n)\n\nplt.show()\n\nplt.close(fig)\n\n\n# ---------------------------------------------------------------\n# Save strategy table\n# ---------------------------------------------------------------\n\nresults_df = pd.DataFrame(\n results\n)\n\nOUTPUT_CSV = os.path.join(\n AUDIT_DIR,\n \"cell32_slice_sampling_indices.csv\"\n)\n\nresults_df.to_csv(\n OUTPUT_CSV,\n index=False\n)\n\n\n# ---------------------------------------------------------------\n# Final\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 32 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Saved visualization:\",\n OUTPUT_IMAGE\n)\n\nprint(\n \"Saved sampling table:\",\n OUTPUT_CSV\n)\n\nprint()\nprint(\n \"Compare Uniform vs Center vs Hybrid \"\n \"in the generated image.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:52:27.14496Z","iopub.execute_input":"2026-08-11T08:52:27.146033Z","iopub.status.idle":"2026-08-11T08:52:37.639057Z","shell.execute_reply.started":"2026-08-11T08:52:27.145997Z","shell.execute_reply":"2026-08-11T08:52:37.637679Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 33 - PYTORCH DATASET + DATALOADER\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom torch.utils.data import Dataset, DataLoader\n\n\n# ---------------------------------------------------------------\n# Configuration\n# ---------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nMODEL_TABLE_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nFOLD_TABLE_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_study_folds.csv\"\n)\n\nIMAGE_SIZE = 224\n\nSLICES_PER_PLANE = 7\n\nPLANES = [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n]\n\nBATCH_SIZE = 2\n\nNUM_WORKERS = 0\n\nTARGET_COLUMNS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n]\n\n\nprint(\"=\" * 70)\nprint(\"CELL 33 - PYTORCH DATASET + DATALOADER\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Verify required functions from previous cells\n# ---------------------------------------------------------------\n\nrequired_functions = [\n \"get_slice_position\",\n \"normalize_dicom_image\",\n \"resize_image\"\n]\n\nmissing_functions = [\n name\n for name in required_functions\n if name not in globals()\n]\n\nif missing_functions:\n\n raise RuntimeError(\n \"Missing functions from previous cells: \"\n + str(missing_functions)\n )\n\n\n# ---------------------------------------------------------------\n# Hybrid sampling\n# ---------------------------------------------------------------\n\ndef hybrid_indices_7(\n n,\n k=SLICES_PER_PLANE,\n margin_fraction=0.10\n):\n \"\"\"\n Select k slices while avoiding the extreme\n first/last portions of the series.\n \"\"\"\n\n if n <= k:\n return list(range(n))\n\n margin = int(\n round(\n n * margin_fraction\n )\n )\n\n start = margin\n end = n - 1 - margin\n\n if end <= start:\n\n indices = np.linspace(\n 0,\n n - 1,\n k\n )\n\n else:\n\n indices = np.linspace(\n start,\n end,\n k\n )\n\n indices = np.round(\n indices\n ).astype(int)\n\n return list(\n np.unique(indices)\n )\n\n\n# ---------------------------------------------------------------\n# DICOM file discovery\n# ---------------------------------------------------------------\n\ndef get_sorted_dicom_files_33(\n study_id,\n series_id\n):\n\n series_path = os.path.join(\n TRAIN_SERIES_DIR,\n str(study_id),\n str(series_id)\n )\n\n if not os.path.isdir(series_path):\n\n raise FileNotFoundError(\n f\"Series directory not found: {series_path}\"\n )\n\n files = []\n\n for filename in os.listdir(\n series_path\n ):\n\n path = os.path.join(\n series_path,\n filename\n )\n\n if os.path.isfile(path):\n\n files.append(path)\n\n records = []\n\n for path in files:\n\n try:\n\n ds = pydicom.dcmread(\n path,\n stop_before_pixels=True,\n force=True\n )\n\n position = get_slice_position(\n ds\n )\n\n records.append(\n (\n position,\n path\n )\n )\n\n except Exception:\n\n continue\n\n records.sort(\n key=lambda x: x[0]\n )\n\n return [\n path\n for _, path in records\n ]\n\n\n# ---------------------------------------------------------------\n# Load selected DICOM slices\n# ---------------------------------------------------------------\n\ndef load_series_33(\n study_id,\n series_id,\n num_slices=SLICES_PER_PLANE\n):\n\n files = get_sorted_dicom_files_33(\n study_id,\n series_id\n )\n\n if len(files) == 0:\n\n raise RuntimeError(\n f\"No readable DICOM files: \"\n f\"{study_id}/{series_id}\"\n )\n\n indices = hybrid_indices_7(\n len(files),\n num_slices\n )\n\n images = []\n\n for index in indices:\n\n ds = pydicom.dcmread(\n files[index],\n force=True\n )\n\n image = ds.pixel_array.astype(\n np.float32\n )\n\n slope = float(\n getattr(\n ds,\n \"RescaleSlope\",\n 1.0\n )\n )\n\n intercept = float(\n getattr(\n ds,\n \"RescaleIntercept\",\n 0.0\n )\n )\n\n image = (\n image * slope\n + intercept\n )\n\n photometric = getattr(\n ds,\n \"PhotometricInterpretation\",\n \"MONOCHROME2\"\n )\n\n if photometric == \"MONOCHROME1\":\n\n image = (\n image.max()\n - image\n )\n\n image = normalize_dicom_image(\n image\n )\n\n image = resize_image(\n image,\n IMAGE_SIZE\n )\n\n images.append(\n image.astype(\n np.float32\n )\n )\n\n # If a very short series results in fewer\n # unique indices, pad by repeating the\n # last available image.\n\n while len(images) < num_slices:\n\n images.append(\n images[-1].copy()\n )\n\n images = images[\n :num_slices\n ]\n\n return np.stack(\n images,\n axis=0\n ).astype(\n np.float32\n )\n\n\n# ---------------------------------------------------------------\n# Study-level dataset\n# ---------------------------------------------------------------\n\nclass KneeStudyDataset(\n Dataset\n):\n\n def __init__(\n self,\n dataframe,\n targets=TARGET_COLUMNS\n ):\n\n self.df = dataframe.reset_index(\n drop=True\n )\n\n self.targets = targets\n\n def __len__(\n self\n ):\n\n return len(self.df)\n\n def __getitem__(\n self,\n index\n ):\n\n row = self.df.iloc[\n index\n ]\n\n study_id = str(\n row[\n \"StudyInstanceUID\"\n ]\n )\n\n plane_images = []\n\n for plane in PLANES:\n\n series_column = (\n f\"{plane}_SeriesInstanceUID\"\n )\n\n series_id = str(\n row[\n series_column\n ]\n )\n\n image_stack = load_series_33(\n study_id,\n series_id,\n SLICES_PER_PLANE\n )\n\n plane_images.append(\n image_stack\n )\n\n # -------------------------------------------------------\n # Shape:\n # 3 planes × 7 slices × 224 × 224\n # becomes:\n # 21 × 224 × 224\n # -------------------------------------------------------\n\n image_tensor = np.concatenate(\n plane_images,\n axis=0\n )\n\n image_tensor = torch.from_numpy(\n image_tensor\n ).float()\n\n labels = np.asarray(\n [\n float(row[target])\n for target in self.targets\n ],\n dtype=np.float32\n )\n\n label_tensor = torch.from_numpy(\n labels\n ).float()\n\n return {\n \"image\": image_tensor,\n \"label\": label_tensor,\n \"study_id\": study_id\n }\n\n\n# ---------------------------------------------------------------\n# Load modeling data\n# ---------------------------------------------------------------\n\nmodeling_df = pd.read_csv(\n MODEL_TABLE_PATH\n)\n\nfold_df = pd.read_csv(\n FOLD_TABLE_PATH\n)\n\n\nprint(\n \"\\nModeling table:\",\n modeling_df.shape\n)\n\nprint(\n \"Fold table:\",\n fold_df.shape\n)\n\n\n# ---------------------------------------------------------------\n# Merge fold information\n# ---------------------------------------------------------------\n\nif \"fold\" not in fold_df.columns:\n\n raise ValueError(\n \"Fold column not found in fold table.\"\n )\n\n\nif (\n \"StudyInstanceUID\"\n not in fold_df.columns\n):\n\n raise ValueError(\n \"StudyInstanceUID missing from fold table.\"\n )\n\n\nmodeling_df = modeling_df.drop(\n columns=[\"fold\"],\n errors=\"ignore\"\n)\n\nmodeling_df = modeling_df.merge(\n fold_df[\n [\n \"StudyInstanceUID\",\n \"fold\"\n ]\n ],\n on=\"StudyInstanceUID\",\n how=\"left\",\n validate=\"one_to_one\"\n)\n\n\nif modeling_df[\"fold\"].isna().any():\n\n raise RuntimeError(\n \"Some modeling studies do not have a fold assignment.\"\n )\n\n\n# ---------------------------------------------------------------\n# Integrity checks\n# ---------------------------------------------------------------\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"DATASET INTEGRITY\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Studies:\",\n len(modeling_df)\n)\n\nprint(\n \"Unique studies:\",\n modeling_df[\n \"StudyInstanceUID\"\n ].nunique()\n)\n\nprint(\n \"Fold distribution:\"\n)\n\nprint(\n modeling_df[\n \"fold\"\n ].value_counts().sort_index()\n)\n\n\nfor plane in PLANES:\n\n column = (\n f\"{plane}_SeriesInstanceUID\"\n )\n\n missing = modeling_df[\n column\n ].isna().sum()\n\n unique = modeling_df[\n column\n ].nunique()\n\n print(\n f\"{plane:8s}: \"\n f\"missing={missing}, \"\n f\"unique={unique}\"\n )\n\n\n# ---------------------------------------------------------------\n# Check targets\n# ---------------------------------------------------------------\n\nfor target in TARGET_COLUMNS:\n\n values = set(\n modeling_df[\n target\n ].dropna().unique()\n )\n\n if not values.issubset(\n {0, 1}\n ):\n\n raise ValueError(\n f\"Invalid values in target: {target}: {values}\"\n )\n\n\nprint(\n \"\\nTargets:\",\n len(TARGET_COLUMNS)\n)\n\nprint(\n \"All target columns contain only 0/1.\"\n)\n\n\n# ---------------------------------------------------------------\n# Create fold datasets\n# ---------------------------------------------------------------\n\ntrain_df = modeling_df[\n modeling_df[\"fold\"] != 0\n].reset_index(\n drop=True\n)\n\nval_df = modeling_df[\n modeling_df[\"fold\"] == 0\n].reset_index(\n drop=True\n)\n\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"FOLD 0 DATASET\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Training studies:\",\n len(train_df)\n)\n\nprint(\n \"Validation studies:\",\n len(val_df)\n)\n\nprint(\n \"Training/validation overlap:\",\n len(\n set(\n train_df[\n \"StudyInstanceUID\"\n ]\n )\n &\n set(\n val_df[\n \"StudyInstanceUID\"\n ]\n )\n )\n)\n\n\n# ---------------------------------------------------------------\n# Create datasets\n# ---------------------------------------------------------------\n\ntrain_dataset = KneeStudyDataset(\n train_df\n)\n\nval_dataset = KneeStudyDataset(\n val_df\n)\n\n\nprint()\nprint(\n \"Train Dataset:\",\n len(train_dataset)\n)\n\nprint(\n \"Validation Dataset:\",\n len(val_dataset)\n)\n\n\n# ---------------------------------------------------------------\n# Create DataLoaders\n# ---------------------------------------------------------------\n\ntrain_loader = DataLoader(\n train_dataset,\n batch_size=BATCH_SIZE,\n shuffle=True,\n num_workers=NUM_WORKERS,\n pin_memory=False\n)\n\nval_loader = DataLoader(\n val_dataset,\n batch_size=BATCH_SIZE,\n shuffle=False,\n num_workers=NUM_WORKERS,\n pin_memory=False\n)\n\n\nprint()\nprint(\n \"Train batches:\",\n len(train_loader)\n)\n\nprint(\n \"Validation batches:\",\n len(val_loader)\n)\n\n\n# ---------------------------------------------------------------\n# Smoke test\n# ---------------------------------------------------------------\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"DATALOADER SMOKE TEST\"\n)\n\nprint(\n \"=\" * 70\n)\n\n\nbatch = next(\n iter(train_loader)\n)\n\nimages = batch[\n \"image\"\n]\n\nlabels = batch[\n \"label\"\n]\n\nstudy_ids = batch[\n \"study_id\"\n]\n\n\nprint(\n \"Image batch shape:\",\n tuple(images.shape)\n)\n\nprint(\n \"Expected image shape:\",\n (\n BATCH_SIZE,\n 21,\n IMAGE_SIZE,\n IMAGE_SIZE\n )\n)\n\nprint(\n \"Label batch shape:\",\n tuple(labels.shape)\n)\n\nprint(\n \"Expected label shape:\",\n (\n BATCH_SIZE,\n len(TARGET_COLUMNS)\n )\n)\n\nprint(\n \"Image dtype:\",\n images.dtype\n)\n\nprint(\n \"Label dtype:\",\n labels.dtype\n)\n\nprint(\n \"Image min:\",\n float(images.min())\n)\n\nprint(\n \"Image max:\",\n float(images.max())\n)\n\nprint(\n \"Image mean:\",\n float(images.mean())\n)\n\nprint(\n \"Image std:\",\n float(images.std())\n)\n\nprint(\n \"First study IDs:\",\n list(study_ids)\n)\n\n\n# ---------------------------------------------------------------\n# Final shape validation\n# ---------------------------------------------------------------\n\nexpected_image_shape = (\n BATCH_SIZE,\n 21,\n IMAGE_SIZE,\n IMAGE_SIZE\n)\n\nexpected_label_shape = (\n BATCH_SIZE,\n len(TARGET_COLUMNS)\n)\n\n\nif tuple(images.shape) != expected_image_shape:\n\n raise RuntimeError(\n f\"Unexpected image shape: \"\n f\"{tuple(images.shape)}\"\n )\n\n\nif tuple(labels.shape) != expected_label_shape:\n\n raise RuntimeError(\n f\"Unexpected label shape: \"\n f\"{tuple(labels.shape)}\"\n )\n\n\nif (\n float(images.min()) < 0.0\n or float(images.max()) > 1.0\n):\n\n raise RuntimeError(\n \"Images are outside expected [0, 1] range.\"\n )\n\n\n# ---------------------------------------------------------------\n# Memory estimate\n# ---------------------------------------------------------------\n\nbytes_per_study = (\n 21\n * IMAGE_SIZE\n * IMAGE_SIZE\n * 4\n)\n\nmb_per_study = (\n bytes_per_study\n / (1024 ** 2)\n)\n\nbatch_mb = (\n mb_per_study\n * BATCH_SIZE\n)\n\nprint()\nprint(\n \"Approximate image memory per study:\",\n round(mb_per_study, 3),\n \"MB\"\n)\n\nprint(\n \"Approximate image memory per batch:\",\n round(batch_mb, 3),\n \"MB\"\n)\n\n\n# ---------------------------------------------------------------\n# Final\n# ---------------------------------------------------------------\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"CELL 33 COMPLETE\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Hybrid sampling:\",\n SLICES_PER_PLANE,\n \"slices per plane\"\n)\n\nprint(\n \"Total model input:\",\n \"21 x 224 x 224\"\n)\n\nprint(\n \"Dataset loading is lazy.\"\n)\n\nprint(\n \"Full dataset was NOT loaded into memory.\"\n)\n\nprint(\n \"Fold-aware train/validation separation verified.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:55:23.219967Z","iopub.execute_input":"2026-08-11T08:55:23.22032Z","iopub.status.idle":"2026-08-11T08:55:24.955405Z","shell.execute_reply.started":"2026-08-11T08:55:23.220289Z","shell.execute_reply":"2026-08-11T08:55:24.954257Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 34 - MODEL ARCHITECTURE + FORWARD PASS\n# ================================================================\n\nimport os\nimport torch\nimport torch.nn as nn\nimport torchvision\nfrom torchvision.models import (\n resnet18,\n ResNet18_Weights\n)\n\n\nprint(\"=\" * 70)\nprint(\"CELL 34 - MODEL ARCHITECTURE\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Configuration\n# ---------------------------------------------------------------\n\nNUM_INPUT_CHANNELS = 21\nNUM_CLASSES = len(TARGET_COLUMNS)\n\nIMAGE_SIZE = 224\n\nDEVICE = torch.device(\n \"cuda\" if torch.cuda.is_available()\n else \"cpu\"\n)\n\n\nprint()\nprint(\"PyTorch version:\", torch.__version__)\nprint(\"Torchvision version:\", torchvision.__version__)\nprint(\"Device:\", DEVICE)\nprint(\"Input channels:\", NUM_INPUT_CHANNELS)\nprint(\"Output classes:\", NUM_CLASSES)\n\nprint()\nprint(\"Target order:\")\n\nfor i, target in enumerate(\n TARGET_COLUMNS\n):\n\n print(\n f\"{i:2d}: {target}\"\n )\n\n\n# ---------------------------------------------------------------\n# Build pretrained ResNet18\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"BUILDING RESNET18 BACKBONE\")\nprint(\"=\" * 70)\n\n\nweights = None\n\ntry:\n\n weights = ResNet18_Weights.DEFAULT\n\n backbone = resnet18(\n weights=weights\n )\n\n pretrained_available = True\n\n print(\n \"Pretrained ResNet18 weights loaded.\"\n )\n\nexcept Exception as e:\n\n print(\n \"WARNING: Pretrained weights could not be loaded.\"\n )\n\n print(\n \"Reason:\",\n str(e)\n )\n\n print(\n \"Falling back to randomly initialized ResNet18.\"\n )\n\n backbone = resnet18(\n weights=None\n )\n\n pretrained_available = False\n\n\n# ---------------------------------------------------------------\n# Modify first convolution\n#\n# Original:\n# 3 input channels\n#\n# New:\n# 21 input channels\n#\n# The original RGB filters are averaged and repeated\n# across the 21 medical-image channels.\n# ---------------------------------------------------------------\n\nold_conv = backbone.conv1\n\nnew_conv = nn.Conv2d(\n in_channels=NUM_INPUT_CHANNELS,\n out_channels=old_conv.out_channels,\n kernel_size=old_conv.kernel_size,\n stride=old_conv.stride,\n padding=old_conv.padding,\n bias=False\n)\n\n\nwith torch.no_grad():\n\n if pretrained_available:\n\n original_weight = (\n old_conv.weight.data\n )\n\n mean_weight = (\n original_weight.mean(\n dim=1,\n keepdim=True\n )\n )\n\n repeated_weight = (\n mean_weight.repeat(\n 1,\n NUM_INPUT_CHANNELS,\n 1,\n 1\n )\n )\n\n # Scale so the magnitude is comparable\n # to the original convolution.\n\n repeated_weight = (\n repeated_weight\n / NUM_INPUT_CHANNELS\n * 3.0\n )\n\n new_conv.weight.copy_(\n repeated_weight\n )\n\n else:\n\n nn.init.kaiming_normal_(\n new_conv.weight,\n mode=\"fan_out\",\n nonlinearity=\"relu\"\n )\n\n\nbackbone.conv1 = new_conv\n\n\n# ---------------------------------------------------------------\n# Remove original ImageNet classifier\n# ---------------------------------------------------------------\n\nfeature_dim = (\n backbone.fc.in_features\n)\n\nbackbone.fc = nn.Identity()\n\n\n# ---------------------------------------------------------------\n# Multilabel knee abnormality model\n# ---------------------------------------------------------------\n\nclass KneeAbnormalityModel(\n nn.Module\n):\n\n def __init__(\n self,\n backbone,\n feature_dim,\n num_classes\n ):\n\n super().__init__()\n\n self.backbone = backbone\n\n self.dropout = nn.Dropout(\n p=0.30\n )\n\n self.classifier = nn.Sequential(\n\n nn.Linear(\n feature_dim,\n 256\n ),\n\n nn.ReLU(\n inplace=True\n ),\n\n nn.Dropout(\n p=0.30\n ),\n\n nn.Linear(\n 256,\n num_classes\n )\n )\n\n def forward(\n self,\n x\n ):\n\n features = self.backbone(\n x\n )\n\n features = self.dropout(\n features\n )\n\n logits = self.classifier(\n features\n )\n\n return logits\n\n\nmodel = KneeAbnormalityModel(\n backbone=backbone,\n feature_dim=feature_dim,\n num_classes=NUM_CLASSES\n)\n\n\n# ---------------------------------------------------------------\n# Freeze backbone initially\n# ---------------------------------------------------------------\n\nfor parameter in model.backbone.parameters():\n\n parameter.requires_grad = False\n\n\n# Keep the modified first convolution trainable.\n# This allows the network to adapt the ImageNet\n# input representation to the 21-channel MRI input.\n\nfor parameter in (\n model.backbone.conv1.parameters()\n):\n\n parameter.requires_grad = True\n\n\n# ---------------------------------------------------------------\n# Move model to device\n# ---------------------------------------------------------------\n\nmodel = model.to(\n DEVICE\n)\n\n\n# ---------------------------------------------------------------\n# Parameter statistics\n# ---------------------------------------------------------------\n\ntotal_parameters = sum(\n parameter.numel()\n for parameter in model.parameters()\n)\n\ntrainable_parameters = sum(\n parameter.numel()\n for parameter in model.parameters()\n if parameter.requires_grad\n)\n\nfrozen_parameters = (\n total_parameters\n - trainable_parameters\n)\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"MODEL PARAMETER SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n \"Total parameters:\",\n f\"{total_parameters:,}\"\n)\n\nprint(\n \"Trainable parameters:\",\n f\"{trainable_parameters:,}\"\n)\n\nprint(\n \"Frozen parameters:\",\n f\"{frozen_parameters:,}\"\n)\n\nprint(\n \"Trainable percentage:\",\n f\"{100 * trainable_parameters / total_parameters:.2f}%\"\n)\n\n\n# ---------------------------------------------------------------\n# Forward-pass smoke test\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FORWARD PASS TEST\")\nprint(\"=\" * 70)\n\n\nmodel.eval()\n\n\nwith torch.no_grad():\n\n sample_batch = next(\n iter(train_loader)\n )\n\n sample_images = (\n sample_batch[\"image\"]\n .to(DEVICE)\n )\n\n sample_labels = (\n sample_batch[\"label\"]\n .to(DEVICE)\n )\n\n output_logits = model(\n sample_images\n )\n\n\nprint(\n \"Input shape:\",\n tuple(\n sample_images.shape\n )\n)\n\nprint(\n \"Label shape:\",\n tuple(\n sample_labels.shape\n )\n)\n\nprint(\n \"Output logits shape:\",\n tuple(\n output_logits.shape\n )\n)\n\nprint(\n \"Expected output shape:\",\n (\n sample_images.shape[0],\n NUM_CLASSES\n )\n)\n\nprint(\n \"Logit dtype:\",\n output_logits.dtype\n)\n\nprint(\n \"Logit minimum:\",\n float(\n output_logits.min()\n )\n)\n\nprint(\n \"Logit maximum:\",\n float(\n output_logits.max()\n )\n)\n\n\n# ---------------------------------------------------------------\n# Sigmoid probability test\n# ---------------------------------------------------------------\n\nprobabilities = torch.sigmoid(\n output_logits\n)\n\nprint()\nprint(\n \"Probability shape:\",\n tuple(\n probabilities.shape\n )\n)\n\nprint(\n \"Probability minimum:\",\n float(\n probabilities.min()\n )\n)\n\nprint(\n \"Probability maximum:\",\n float(\n probabilities.max()\n )\n)\n\n\n# ---------------------------------------------------------------\n# Output class mapping\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"OUTPUT MAPPING\")\nprint(\"=\" * 70)\n\nfor i, target in enumerate(\n TARGET_COLUMNS\n):\n\n print(\n f\"Output {i:2d} -> {target}\"\n )\n\n\n# ---------------------------------------------------------------\n# Final validation\n# ---------------------------------------------------------------\n\nexpected_output_shape = (\n sample_images.shape[0],\n NUM_CLASSES\n)\n\nif tuple(\n output_logits.shape\n) != expected_output_shape:\n\n raise RuntimeError(\n \"Model output shape is incorrect.\"\n )\n\n\nif not torch.isfinite(\n output_logits\n).all():\n\n raise RuntimeError(\n \"Model produced NaN or infinite logits.\"\n )\n\n\nif not torch.isfinite(\n probabilities\n).all():\n\n raise RuntimeError(\n \"Model produced NaN or infinite probabilities.\"\n )\n\n\n# ---------------------------------------------------------------\n# Final\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 34 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Model construction: PASS\"\n)\n\nprint(\n \"21-channel input: PASS\"\n)\n\nprint(\n \"12-class multilabel output: PASS\"\n)\n\nprint(\n \"Forward pass: PASS\"\n)\n\nprint(\n \"NaN/Inf check: PASS\"\n)\n\nprint(\n \"Backbone initially frozen: PASS\"\n)\n\nprint(\n \"Ready for loss-function and training setup.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:56:41.78035Z","iopub.execute_input":"2026-08-11T08:56:41.780721Z","iopub.status.idle":"2026-08-11T08:56:48.678694Z","shell.execute_reply.started":"2026-08-11T08:56:41.78069Z","shell.execute_reply":"2026-08-11T08:56:48.677432Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 35 - LOSS FUNCTION + CLASS IMBALANCE + OPTIMIZER\n# ================================================================\n\nimport random\nimport numpy as np\nimport torch\nimport torch.nn as nn\n\n\nprint(\"=\" * 70)\nprint(\"CELL 35 - LOSS + CLASS WEIGHTS + OPTIMIZER\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Reproducibility\n# ---------------------------------------------------------------\n\nSEED = 42\n\nrandom.seed(SEED)\nnp.random.seed(SEED)\ntorch.manual_seed(SEED)\n\nif torch.cuda.is_available():\n torch.cuda.manual_seed_all(SEED)\n\n\nprint()\nprint(\"Random seed:\", SEED)\n\n\n# ---------------------------------------------------------------\n# Training configuration\n# ---------------------------------------------------------------\n\nLEARNING_RATE = 1e-3\n\nWEIGHT_DECAY = 1e-4\n\nMAX_GRAD_NORM = 1.0\n\nNUM_EPOCHS = 10\n\nEARLY_STOPPING_PATIENCE = 3\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"TRAINING CONFIGURATION\")\nprint(\"=\" * 70)\n\nprint(\"Learning rate:\", LEARNING_RATE)\nprint(\"Weight decay:\", WEIGHT_DECAY)\nprint(\"Maximum gradient norm:\", MAX_GRAD_NORM)\nprint(\"Maximum epochs:\", NUM_EPOCHS)\nprint(\n \"Early stopping patience:\",\n EARLY_STOPPING_PATIENCE\n)\n\n\n# ---------------------------------------------------------------\n# Calculate class imbalance weights\n#\n# IMPORTANT:\n# Calculate weights ONLY from the training split.\n# Validation labels must never influence training.\n# ---------------------------------------------------------------\n\ntrain_labels = train_df[\n TARGET_COLUMNS\n].astype(\n np.float32\n)\n\n\npositive_counts = (\n train_labels.sum(\n axis=0\n )\n)\n\nnegative_counts = (\n len(train_labels)\n - positive_counts\n)\n\n\n# pos_weight = negative / positive\n#\n# BCEWithLogitsLoss then gives more importance to\n# positive examples for rare abnormalities.\n\npos_weight_values = (\n negative_counts\n / positive_counts\n)\n\n\npos_weight = torch.tensor(\n pos_weight_values.values,\n dtype=torch.float32,\n device=DEVICE\n)\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"CLASS IMBALANCE\")\nprint(\"=\" * 70)\n\nprint(\n f\"{'Target':22s}\"\n f\"{'Positive':>10s}\"\n f\"{'Negative':>10s}\"\n f\"{'Pos Weight':>14s}\"\n)\n\nprint(\"-\" * 60)\n\nfor target, pos, neg, weight in zip(\n TARGET_COLUMNS,\n positive_counts,\n negative_counts,\n pos_weight_values\n):\n\n print(\n f\"{target:22s}\"\n f\"{int(pos):10d}\"\n f\"{int(neg):10d}\"\n f\"{weight:14.4f}\"\n )\n\n\n# ---------------------------------------------------------------\n# Validate class weights\n# ---------------------------------------------------------------\n\nif not torch.isfinite(\n pos_weight\n).all():\n\n raise RuntimeError(\n \"Invalid NaN/Inf values found in pos_weight.\"\n )\n\n\nif (\n pos_weight <= 0\n).any():\n\n raise RuntimeError(\n \"Invalid non-positive class weight found.\"\n )\n\n\n# ---------------------------------------------------------------\n# Loss function\n# ---------------------------------------------------------------\n\ncriterion = nn.BCEWithLogitsLoss(\n pos_weight=pos_weight\n)\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"LOSS FUNCTION\")\nprint(\"=\" * 70)\n\nprint(\n \"Loss:\",\n criterion\n)\n\nprint(\n \"Type:\",\n type(criterion).__name__\n)\n\nprint(\n \"Using per-target positive weights: YES\"\n)\n\n\n# ---------------------------------------------------------------\n# Optimizer\n#\n# Only currently trainable parameters are passed.\n# The frozen ResNet parameters will not be updated.\n# ---------------------------------------------------------------\n\ntrainable_parameters = [\n parameter\n for parameter in model.parameters()\n if parameter.requires_grad\n]\n\n\noptimizer = torch.optim.AdamW(\n trainable_parameters,\n lr=LEARNING_RATE,\n weight_decay=WEIGHT_DECAY\n)\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"OPTIMIZER\")\nprint(\"=\" * 70)\n\nprint(\n \"Optimizer:\",\n type(optimizer).__name__\n)\n\nprint(\n \"Trainable parameter tensors:\",\n len(trainable_parameters)\n)\n\nprint(\n \"Learning rate:\",\n LEARNING_RATE\n)\n\nprint(\n \"Weight decay:\",\n WEIGHT_DECAY\n)\n\n\n# ---------------------------------------------------------------\n# Learning-rate scheduler\n#\n# Reduce LR when validation performance stops improving.\n# The scheduler will be stepped after validation.\n# ---------------------------------------------------------------\n\nscheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(\n optimizer,\n mode=\"min\",\n factor=0.5,\n patience=1\n)\n\n\nprint(\n \"Scheduler:\",\n type(scheduler).__name__\n)\n\n\n# ---------------------------------------------------------------\n# Verify optimizer parameters\n# ---------------------------------------------------------------\n\noptimizer_parameter_count = sum(\n parameter.numel()\n for group in optimizer.param_groups\n for parameter in group[\"params\"]\n)\n\n\nactual_trainable_count = sum(\n parameter.numel()\n for parameter in model.parameters()\n if parameter.requires_grad\n)\n\n\nprint()\nprint(\n \"Optimizer parameter count:\",\n f\"{optimizer_parameter_count:,}\"\n)\n\nprint(\n \"Actual trainable parameter count:\",\n f\"{actual_trainable_count:,}\"\n)\n\n\nif (\n optimizer_parameter_count\n != actual_trainable_count\n):\n\n raise RuntimeError(\n \"Optimizer parameter count does not match \"\n \"the model's trainable parameter count.\"\n )\n\n\n# ---------------------------------------------------------------\n# One-batch loss test\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"LOSS SMOKE TEST\")\nprint(\"=\" * 70)\n\n\nmodel.train()\n\n\nsample_batch = next(\n iter(train_loader)\n)\n\nsample_images = (\n sample_batch[\"image\"]\n .to(DEVICE)\n)\n\nsample_labels = (\n sample_batch[\"label\"]\n .to(DEVICE)\n)\n\n\noptimizer.zero_grad(\n set_to_none=True\n)\n\n\nsample_logits = model(\n sample_images\n)\n\n\nsample_loss = criterion(\n sample_logits,\n sample_labels\n)\n\n\nprint(\n \"Input shape:\",\n tuple(\n sample_images.shape\n )\n)\n\nprint(\n \"Logit shape:\",\n tuple(\n sample_logits.shape\n )\n)\n\nprint(\n \"Label shape:\",\n tuple(\n sample_labels.shape\n )\n)\n\nprint(\n \"Loss value:\",\n float(\n sample_loss\n )\n)\n\n\nif not torch.isfinite(\n sample_loss\n):\n\n raise RuntimeError(\n \"Loss produced NaN or Inf.\"\n )\n\n\n# ---------------------------------------------------------------\n# Backward-pass test\n# ---------------------------------------------------------------\n\nsample_loss.backward()\n\n\n# ---------------------------------------------------------------\n# Gradient statistics\n# ---------------------------------------------------------------\n\ngradient_norms = []\n\nfor parameter in model.parameters():\n\n if (\n parameter.requires_grad\n and parameter.grad is not None\n ):\n\n gradient_norms.append(\n parameter.grad.detach().norm().item()\n )\n\n\nif len(gradient_norms) == 0:\n\n raise RuntimeError(\n \"No gradients were produced.\"\n )\n\n\nmax_gradient = max(\n gradient_norms\n)\n\nmean_gradient = np.mean(\n gradient_norms\n)\n\n\nprint()\nprint(\n \"Gradient tensors:\",\n len(gradient_norms)\n)\n\nprint(\n \"Mean gradient norm:\",\n float(mean_gradient)\n)\n\nprint(\n \"Maximum gradient norm:\",\n float(max_gradient)\n)\n\n\nif not np.isfinite(\n max_gradient\n):\n\n raise RuntimeError(\n \"Invalid gradient detected.\"\n )\n\n\n# ---------------------------------------------------------------\n# Clear gradients after smoke test\n# ---------------------------------------------------------------\n\noptimizer.zero_grad(\n set_to_none=True\n)\n\n\n# ---------------------------------------------------------------\n# Final configuration summary\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 35 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Weighted BCE loss: PASS\"\n)\n\nprint(\n \"Class imbalance handling: PASS\"\n)\n\nprint(\n \"Optimizer: PASS\"\n)\n\nprint(\n \"Scheduler: PASS\"\n)\n\nprint(\n \"Backward pass: PASS\"\n)\n\nprint(\n \"Gradient validity: PASS\"\n)\n\nprint()\nprint(\n \"Ready for one-batch training test.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T08:58:43.817351Z","iopub.execute_input":"2026-08-11T08:58:43.818393Z","iopub.status.idle":"2026-08-11T08:58:47.683366Z","shell.execute_reply.started":"2026-08-11T08:58:43.81835Z","shell.execute_reply":"2026-08-11T08:58:47.682481Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 36 - ONE-BATCH TRAINING + CPU PERFORMANCE TEST\n# ================================================================\n\nimport time\nimport numpy as np\nimport torch\n\n\nprint(\"=\" * 70)\nprint(\"CELL 36 - ONE-BATCH TRAINING TEST\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Configuration\n# ---------------------------------------------------------------\n\nGRADIENT_CLIP_NORM = MAX_GRAD_NORM\n\n\nprint()\nprint(\"Device:\", DEVICE)\nprint(\"Gradient clipping:\", GRADIENT_CLIP_NORM)\n\n\n# ---------------------------------------------------------------\n# Prepare model\n# ---------------------------------------------------------------\n\nmodel.train()\n\noptimizer.zero_grad(\n set_to_none=True\n)\n\n\n# ---------------------------------------------------------------\n# Load one batch\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"LOADING TRAINING BATCH\")\nprint(\"=\" * 70)\n\n\nbatch_start = time.perf_counter()\n\n\nbatch = next(\n iter(train_loader)\n)\n\n\nbatch_load_time = (\n time.perf_counter()\n - batch_start\n)\n\n\nimages = batch[\n \"image\"\n].to(\n DEVICE\n)\n\nlabels = batch[\n \"label\"\n].to(\n DEVICE\n)\n\n\nprint(\n \"Batch loaded in:\",\n f\"{batch_load_time:.3f} seconds\"\n)\n\nprint(\n \"Images:\",\n tuple(images.shape)\n)\n\nprint(\n \"Labels:\",\n tuple(labels.shape)\n)\n\n\n# ---------------------------------------------------------------\n# Forward pass\n# ---------------------------------------------------------------\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"FORWARD + BACKWARD + OPTIMIZER STEP\"\n)\n\nprint(\n \"=\" * 70\n)\n\n\nstart_time = time.perf_counter()\n\n\nlogits = model(\n images\n)\n\n\nforward_time = (\n time.perf_counter()\n - start_time\n)\n\n\n# ---------------------------------------------------------------\n# Loss\n# ---------------------------------------------------------------\n\nloss = criterion(\n logits,\n labels\n)\n\n\nif not torch.isfinite(\n loss\n):\n\n raise RuntimeError(\n \"Training loss is NaN or Inf.\"\n )\n\n\n# ---------------------------------------------------------------\n# Backward\n# ---------------------------------------------------------------\n\nbackward_start = time.perf_counter()\n\n\nloss.backward()\n\n\nbackward_time = (\n time.perf_counter()\n - backward_start\n)\n\n\n# ---------------------------------------------------------------\n# Gradient statistics BEFORE clipping\n# ---------------------------------------------------------------\n\ngradient_norm_before = (\n torch.nn.utils.clip_grad_norm_(\n model.parameters(),\n max_norm=GRADIENT_CLIP_NORM\n )\n)\n\n\ngradient_norm_before_value = float(\n gradient_norm_before\n)\n\n\nprint(\n \"Gradient norm before clipping:\",\n gradient_norm_before_value\n)\n\n\n# ---------------------------------------------------------------\n# Optimizer step\n# ---------------------------------------------------------------\n\noptimizer_start = time.perf_counter()\n\n\noptimizer.step()\n\n\noptimizer_time = (\n time.perf_counter()\n - optimizer_start\n)\n\n\ntotal_training_time = (\n time.perf_counter()\n - start_time\n)\n\n\n# ---------------------------------------------------------------\n# Clear gradients\n# ---------------------------------------------------------------\n\noptimizer.zero_grad(\n set_to_none=True\n)\n\n\n# ---------------------------------------------------------------\n# Check parameters remain valid\n# ---------------------------------------------------------------\n\ninvalid_parameters = 0\n\nfor parameter in model.parameters():\n\n if not torch.isfinite(\n parameter\n ).all():\n\n invalid_parameters += 1\n\n\nif invalid_parameters > 0:\n\n raise RuntimeError(\n \"Invalid NaN/Inf model parameters detected \"\n f\"after optimizer step: {invalid_parameters}\"\n )\n\n\n# ---------------------------------------------------------------\n# Probability check\n# ---------------------------------------------------------------\n\nwith torch.no_grad():\n\n probabilities = torch.sigmoid(\n logits\n )\n\n\nif not torch.isfinite(\n probabilities\n).all():\n\n raise RuntimeError(\n \"Invalid probabilities detected.\"\n )\n\n\n# ---------------------------------------------------------------\n# Timing summary\n# ---------------------------------------------------------------\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"PERFORMANCE SUMMARY\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Batch loading time:\",\n f\"{batch_load_time:.3f} sec\"\n)\n\nprint(\n \"Forward time:\",\n f\"{forward_time:.3f} sec\"\n)\n\nprint(\n \"Backward time:\",\n f\"{backward_time:.3f} sec\"\n)\n\nprint(\n \"Optimizer step:\",\n f\"{optimizer_time:.3f} sec\"\n)\n\nprint(\n \"Total training step:\",\n f\"{total_training_time:.3f} sec\"\n)\n\n\n# ---------------------------------------------------------------\n# Training result\n# ---------------------------------------------------------------\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"TRAINING STEP RESULT\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Loss:\",\n float(loss)\n)\n\nprint(\n \"Logit min:\",\n float(logits.min())\n)\n\nprint(\n \"Logit max:\",\n float(logits.max())\n)\n\nprint(\n \"Probability min:\",\n float(probabilities.min())\n)\n\nprint(\n \"Probability max:\",\n float(probabilities.max())\n)\n\nprint(\n \"Gradient clipping applied:\",\n \"YES\"\n)\n\nprint(\n \"Model parameters valid:\",\n \"YES\"\n)\n\n\n# ---------------------------------------------------------------\n# Estimate Fold-0 epoch time\n# ---------------------------------------------------------------\n\nsteps_per_epoch = len(\n train_loader\n)\n\nestimated_epoch_seconds = (\n total_training_time\n * steps_per_epoch\n)\n\n\nestimated_epoch_minutes = (\n estimated_epoch_seconds\n / 60.0\n)\n\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"TRAINING TIME ESTIMATE\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Training batches per epoch:\",\n steps_per_epoch\n)\n\nprint(\n \"Estimated Fold-0 epoch time:\",\n f\"{estimated_epoch_minutes:.2f} minutes\"\n)\n\nprint(\n \"Estimated 10-epoch Fold-0 time:\",\n f\"{estimated_epoch_minutes * 10:.2f} minutes\"\n)\n\n\n# ---------------------------------------------------------------\n# Final validation\n# ---------------------------------------------------------------\n\nif not np.isfinite(\n float(loss)\n):\n\n raise RuntimeError(\n \"Invalid loss.\"\n )\n\n\nif not np.isfinite(\n gradient_norm_before_value\n):\n\n raise RuntimeError(\n \"Invalid gradient norm.\"\n )\n\n\nif total_training_time <= 0:\n\n raise RuntimeError(\n \"Invalid timing measurement.\"\n )\n\n\nprint()\nprint(\n \"=\" * 70\n)\n\nprint(\n \"CELL 36 COMPLETE\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Forward pass: PASS\"\n)\n\nprint(\n \"Backward pass: PASS\"\n)\n\nprint(\n \"Gradient clipping: PASS\"\n)\n\nprint(\n \"Optimizer update: PASS\"\n)\n\nprint(\n \"Parameter validity: PASS\"\n)\n\nprint(\n \"CPU performance measurement: PASS\"\n)\n\nprint()\nprint(\n \"Ready to decide the full training configuration.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T09:00:20.370003Z","iopub.execute_input":"2026-08-11T09:00:20.370344Z","iopub.status.idle":"2026-08-11T09:00:22.83923Z","shell.execute_reply.started":"2026-08-11T09:00:20.370313Z","shell.execute_reply":"2026-08-11T09:00:22.838115Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 37 - FOLD 0 TRAINING\n# ================================================================\n\nimport os\nimport time\nimport copy\nimport numpy as np\nimport pandas as pd\nimport torch\n\n\nprint(\"=\" * 70)\nprint(\"CELL 37 - FOLD 0 TRAINING\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Configuration\n# ---------------------------------------------------------------\n\nFOLD_ID = 0\n\nNUM_EPOCHS = 10\n\nEARLY_STOPPING_PATIENCE = 3\n\nGRADIENT_CLIP_NORM = 1.0\n\nCHECKPOINT_PATH = os.path.join(\n AUDIT_DIR,\n \"cell37_fold0_best_model.pt\"\n)\n\nHISTORY_PATH = os.path.join(\n AUDIT_DIR,\n \"cell37_fold0_history.csv\"\n)\n\n\nprint()\nprint(\"Fold:\", FOLD_ID)\nprint(\"Maximum epochs:\", NUM_EPOCHS)\nprint(\n \"Early stopping patience:\",\n EARLY_STOPPING_PATIENCE\n)\n\nprint(\n \"Device:\",\n DEVICE\n)\n\n\n# ---------------------------------------------------------------\n# Recreate Fold-0 train/validation datasets\n# ---------------------------------------------------------------\n\nfold_train_df = modeling_df[\n modeling_df[\"fold\"] != FOLD_ID\n].reset_index(\n drop=True\n)\n\nfold_val_df = modeling_df[\n modeling_df[\"fold\"] == FOLD_ID\n].reset_index(\n drop=True\n)\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"FOLD DATA\")\nprint(\"=\" * 70)\n\nprint(\n \"Training studies:\",\n len(fold_train_df)\n)\n\nprint(\n \"Validation studies:\",\n len(fold_val_df)\n)\n\n\ntrain_dataset_fold0 = KneeStudyDataset(\n fold_train_df\n)\n\nval_dataset_fold0 = KneeStudyDataset(\n fold_val_df\n)\n\n\ntrain_loader_fold0 = DataLoader(\n train_dataset_fold0,\n batch_size=BATCH_SIZE,\n shuffle=True,\n num_workers=NUM_WORKERS,\n pin_memory=False\n)\n\nval_loader_fold0 = DataLoader(\n val_dataset_fold0,\n batch_size=BATCH_SIZE,\n shuffle=False,\n num_workers=NUM_WORKERS,\n pin_memory=False\n)\n\n\nprint(\n \"Training batches:\",\n len(train_loader_fold0)\n)\n\nprint(\n \"Validation batches:\",\n len(val_loader_fold0)\n)\n\n\n# ---------------------------------------------------------------\n# Reset model to the original Cell 34 state\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"RESETTING MODEL\")\nprint(\"=\" * 70)\n\n\nweights = ResNet18_Weights.DEFAULT\n\nbackbone = resnet18(\n weights=weights\n)\n\n\nold_conv = backbone.conv1\n\nnew_conv = nn.Conv2d(\n in_channels=NUM_INPUT_CHANNELS,\n out_channels=old_conv.out_channels,\n kernel_size=old_conv.kernel_size,\n stride=old_conv.stride,\n padding=old_conv.padding,\n bias=False\n)\n\n\nwith torch.no_grad():\n\n original_weight = (\n old_conv.weight.data\n )\n\n mean_weight = (\n original_weight.mean(\n dim=1,\n keepdim=True\n )\n )\n\n repeated_weight = (\n mean_weight.repeat(\n 1,\n NUM_INPUT_CHANNELS,\n 1,\n 1\n )\n )\n\n repeated_weight = (\n repeated_weight\n / NUM_INPUT_CHANNELS\n * 3.0\n )\n\n new_conv.weight.copy_(\n repeated_weight\n )\n\n\nbackbone.conv1 = new_conv\n\nfeature_dim = (\n backbone.fc.in_features\n)\n\nbackbone.fc = nn.Identity()\n\n\nmodel = KneeAbnormalityModel(\n backbone=backbone,\n feature_dim=feature_dim,\n num_classes=NUM_CLASSES\n)\n\n\n# Freeze backbone\n\nfor parameter in model.backbone.parameters():\n\n parameter.requires_grad = False\n\n\n# Keep first convolution trainable\n\nfor parameter in (\n model.backbone.conv1.parameters()\n):\n\n parameter.requires_grad = True\n\n\nmodel = model.to(\n DEVICE\n)\n\n\n# ---------------------------------------------------------------\n# Fold-specific class weights\n# ---------------------------------------------------------------\n\nfold_train_labels = (\n fold_train_df[\n TARGET_COLUMNS\n ].astype(\n np.float32\n )\n)\n\n\npositive_counts_fold = (\n fold_train_labels.sum(\n axis=0\n )\n)\n\nnegative_counts_fold = (\n len(fold_train_labels)\n - positive_counts_fold\n)\n\n\npos_weight_fold = torch.tensor(\n (\n negative_counts_fold\n / positive_counts_fold\n ).values,\n dtype=torch.float32,\n device=DEVICE\n)\n\n\ncriterion_fold = nn.BCEWithLogitsLoss(\n pos_weight=pos_weight_fold\n)\n\n\n# ---------------------------------------------------------------\n# Optimizer\n# ---------------------------------------------------------------\n\noptimizer_fold = torch.optim.AdamW(\n [\n parameter\n for parameter in model.parameters()\n if parameter.requires_grad\n ],\n lr=LEARNING_RATE,\n weight_decay=WEIGHT_DECAY\n)\n\n\nscheduler_fold = (\n torch.optim.lr_scheduler.ReduceLROnPlateau(\n optimizer_fold,\n mode=\"min\",\n factor=0.5,\n patience=1\n )\n)\n\n\nprint(\n \"Trainable parameters:\",\n f\"{sum(p.numel() for p in model.parameters() if p.requires_grad):,}\"\n)\n\n\n# ---------------------------------------------------------------\n# Validation function\n# ---------------------------------------------------------------\n\ndef evaluate_fold0(\n model,\n loader,\n criterion,\n device\n):\n\n model.eval()\n\n total_loss = 0.0\n\n total_samples = 0\n\n all_logits = []\n\n all_labels = []\n\n with torch.no_grad():\n\n for batch in loader:\n\n images = batch[\n \"image\"\n ].to(\n device\n )\n\n labels = batch[\n \"label\"\n ].to(\n device\n )\n\n logits = model(\n images\n )\n\n loss = criterion(\n logits,\n labels\n )\n\n batch_size = (\n images.shape[0]\n )\n\n total_loss += (\n float(loss)\n * batch_size\n )\n\n total_samples += (\n batch_size\n )\n\n all_logits.append(\n logits.cpu()\n )\n\n all_labels.append(\n labels.cpu()\n )\n\n\n mean_loss = (\n total_loss\n / total_samples\n )\n\n\n logits_array = torch.cat(\n all_logits,\n dim=0\n ).numpy()\n\n\n labels_array = torch.cat(\n all_labels,\n dim=0\n ).numpy()\n\n\n probabilities = (\n 1.0\n / (\n 1.0\n + np.exp(\n -np.clip(\n logits_array,\n -50,\n 50\n )\n )\n )\n )\n\n\n predictions = (\n probabilities >= 0.5\n ).astype(\n np.int32\n )\n\n\n return (\n mean_loss,\n labels_array,\n probabilities,\n predictions\n )\n\n\n# ---------------------------------------------------------------\n# Metric calculation\n# ---------------------------------------------------------------\n\ndef calculate_metrics(\n labels,\n probabilities,\n predictions\n):\n\n results = []\n\n for index, target in enumerate(\n TARGET_COLUMNS\n ):\n\n y_true = labels[\n :,\n index\n ]\n\n y_prob = probabilities[\n :,\n index\n ]\n\n y_pred = predictions[\n :,\n index\n ]\n\n\n tp = np.sum(\n (\n y_true == 1\n )\n &\n (\n y_pred == 1\n )\n )\n\n tn = np.sum(\n (\n y_true == 0\n )\n &\n (\n y_pred == 0\n )\n )\n\n fp = np.sum(\n (\n y_true == 0\n )\n &\n (\n y_pred == 1\n )\n )\n\n fn = np.sum(\n (\n y_true == 1\n )\n &\n (\n y_pred == 0\n )\n )\n\n\n sensitivity = (\n tp / (tp + fn)\n if (tp + fn) > 0\n else np.nan\n )\n\n specificity = (\n tn / (tn + fp)\n if (tn + fp) > 0\n else np.nan\n )\n\n precision = (\n tp / (tp + fp)\n if (tp + fp) > 0\n else 0.0\n )\n\n f1 = (\n 2 * precision * sensitivity\n / (precision + sensitivity)\n if (\n precision + sensitivity\n ) > 0\n else 0.0\n )\n\n\n results.append(\n {\n \"target\": target,\n \"positive_count\": int(\n np.sum(\n y_true == 1\n )\n ),\n \"negative_count\": int(\n np.sum(\n y_true == 0\n )\n ),\n \"predicted_positive\": int(\n np.sum(\n y_pred == 1\n )\n ),\n \"sensitivity\": sensitivity,\n \"specificity\": specificity,\n \"precision\": precision,\n \"f1\": f1\n }\n )\n\n\n return pd.DataFrame(\n results\n )\n\n\n# ---------------------------------------------------------------\n# Training loop\n# ---------------------------------------------------------------\n\nhistory = []\n\nbest_val_loss = float(\n \"inf\"\n)\n\nbest_epoch = None\n\nbest_state = None\n\nepochs_without_improvement = 0\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"STARTING FOLD 0 TRAINING\")\nprint(\"=\" * 70)\n\n\nfor epoch in range(\n 1,\n NUM_EPOCHS + 1\n):\n\n epoch_start = time.perf_counter()\n\n\n # -----------------------------------------------------------\n # TRAIN\n # -----------------------------------------------------------\n\n model.train()\n\n train_loss_sum = 0.0\n\n train_samples = 0\n\n train_batch_times = []\n\n\n for batch in train_loader_fold0:\n\n batch_start = time.perf_counter()\n\n\n images = batch[\n \"image\"\n ].to(\n DEVICE\n )\n\n labels = batch[\n \"label\"\n ].to(\n DEVICE\n )\n\n\n optimizer_fold.zero_grad(\n set_to_none=True\n )\n\n\n logits = model(\n images\n )\n\n\n loss = criterion_fold(\n logits,\n labels\n )\n\n\n if not torch.isfinite(\n loss\n ):\n\n raise RuntimeError(\n f\"Non-finite training loss \"\n f\"at epoch {epoch}.\"\n )\n\n\n loss.backward()\n\n\n torch.nn.utils.clip_grad_norm_(\n model.parameters(),\n max_norm=GRADIENT_CLIP_NORM\n )\n\n\n optimizer_fold.step()\n\n\n batch_size = (\n images.shape[0]\n )\n\n\n train_loss_sum += (\n float(loss)\n * batch_size\n )\n\n train_samples += (\n batch_size\n )\n\n\n train_batch_times.append(\n time.perf_counter()\n - batch_start\n )\n\n\n train_loss = (\n train_loss_sum\n / train_samples\n )\n\n\n # -----------------------------------------------------------\n # VALIDATION\n # -----------------------------------------------------------\n\n validation_start = (\n time.perf_counter()\n )\n\n\n (\n val_loss,\n val_labels,\n val_probabilities,\n val_predictions\n ) = evaluate_fold0(\n model,\n val_loader_fold0,\n criterion_fold,\n DEVICE\n )\n\n\n validation_time = (\n time.perf_counter()\n - validation_start\n )\n\n\n # -----------------------------------------------------------\n # Metrics\n # -----------------------------------------------------------\n\n metric_df = calculate_metrics(\n val_labels,\n val_probabilities,\n val_predictions\n )\n\n\n mean_f1 = float(\n metric_df[\n \"f1\"\n ].mean()\n )\n\n\n mean_sensitivity = float(\n metric_df[\n \"sensitivity\"\n ].replace(\n [np.inf, -np.inf],\n np.nan\n ).mean()\n )\n\n\n mean_specificity = float(\n metric_df[\n \"specificity\"\n ].replace(\n [np.inf, -np.inf],\n np.nan\n ).mean()\n )\n\n\n mean_precision = float(\n metric_df[\n \"precision\"\n ].mean()\n )\n\n\n # -----------------------------------------------------------\n # Scheduler\n # -----------------------------------------------------------\n\n scheduler_fold.step(\n val_loss\n )\n\n\n current_lr = (\n optimizer_fold\n .param_groups[0][\"lr\"]\n )\n\n\n # -----------------------------------------------------------\n # Epoch timing\n # -----------------------------------------------------------\n\n epoch_time = (\n time.perf_counter()\n - epoch_start\n )\n\n\n average_batch_time = (\n np.mean(\n train_batch_times\n )\n )\n\n\n # -----------------------------------------------------------\n # Improvement check\n # -----------------------------------------------------------\n\n improved = (\n val_loss\n < best_val_loss\n - 1e-5\n )\n\n\n if improved:\n\n best_val_loss = val_loss\n\n best_epoch = epoch\n\n best_state = copy.deepcopy(\n model.state_dict()\n )\n\n torch.save(\n {\n \"epoch\": epoch,\n \"model_state_dict\": best_state,\n \"optimizer_state_dict\":\n optimizer_fold.state_dict(),\n \"scheduler_state_dict\":\n scheduler_fold.state_dict(),\n \"best_val_loss\":\n best_val_loss,\n \"target_columns\":\n TARGET_COLUMNS,\n \"fold\":\n FOLD_ID\n },\n CHECKPOINT_PATH\n )\n\n epochs_without_improvement = 0\n\n else:\n\n epochs_without_improvement += 1\n\n\n # -----------------------------------------------------------\n # Store history\n # -----------------------------------------------------------\n\n history.append(\n {\n \"epoch\": epoch,\n \"train_loss\": train_loss,\n \"val_loss\": val_loss,\n \"mean_f1\": mean_f1,\n \"mean_sensitivity\":\n mean_sensitivity,\n \"mean_specificity\":\n mean_specificity,\n \"mean_precision\":\n mean_precision,\n \"learning_rate\": current_lr,\n \"epoch_time_seconds\":\n epoch_time,\n \"validation_time_seconds\":\n validation_time,\n \"average_batch_time_seconds\":\n average_batch_time\n }\n )\n\n\n # -----------------------------------------------------------\n # Print epoch summary\n # -----------------------------------------------------------\n\n print()\n print(\n f\"Epoch {epoch:02d}/{NUM_EPOCHS}\"\n )\n\n print(\n f\"Train Loss: {train_loss:.5f}\"\n )\n\n print(\n f\"Val Loss: {val_loss:.5f}\"\n )\n\n print(\n f\"Mean F1: {mean_f1:.4f}\"\n )\n\n print(\n f\"Sensitivity: {mean_sensitivity:.4f}\"\n )\n\n print(\n f\"Specificity: {mean_specificity:.4f}\"\n )\n\n print(\n f\"Precision: {mean_precision:.4f}\"\n )\n\n print(\n f\"Learning Rate: {current_lr:.6f}\"\n )\n\n print(\n f\"Epoch Time: {epoch_time:.2f} sec\"\n )\n\n print(\n f\"Avg Train Batch: \"\n f\"{average_batch_time:.2f} sec\"\n )\n\n print(\n f\"Validation Time: \"\n f\"{validation_time:.2f} sec\"\n )\n\n print(\n \"Best Epoch:\",\n best_epoch\n )\n\n print(\n \"Best Val Loss:\",\n f\"{best_val_loss:.5f}\"\n )\n\n\n # -----------------------------------------------------------\n # Early stopping\n # -----------------------------------------------------------\n\n if (\n epochs_without_improvement\n >= EARLY_STOPPING_PATIENCE\n ):\n\n print()\n print(\n \"Early stopping triggered.\"\n )\n\n break\n\n\n# ---------------------------------------------------------------\n# Restore best model\n# ---------------------------------------------------------------\n\nif best_state is not None:\n\n model.load_state_dict(\n best_state\n )\n\n\n# ---------------------------------------------------------------\n# Save history\n# ---------------------------------------------------------------\n\nhistory_df = pd.DataFrame(\n history\n)\n\nhistory_df.to_csv(\n HISTORY_PATH,\n index=False\n)\n\n\n# ---------------------------------------------------------------\n# Final validation using best model\n# ---------------------------------------------------------------\n\n(\n final_val_loss,\n final_val_labels,\n final_val_probabilities,\n final_val_predictions\n) = evaluate_fold0(\n model,\n val_loader_fold0,\n criterion_fold,\n DEVICE\n)\n\n\nfinal_metrics = calculate_metrics(\n final_val_labels,\n final_val_probabilities,\n final_val_predictions\n)\n\n\n# ---------------------------------------------------------------\n# Save validation metrics\n# ---------------------------------------------------------------\n\nMETRICS_PATH = os.path.join(\n AUDIT_DIR,\n \"cell37_fold0_metrics.csv\"\n)\n\nfinal_metrics.to_csv(\n METRICS_PATH,\n index=False\n)\n\n\n# ---------------------------------------------------------------\n# Final summary\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FOLD 0 TRAINING COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Best epoch:\",\n best_epoch\n)\n\nprint(\n \"Best validation loss:\",\n f\"{best_val_loss:.6f}\"\n)\n\nprint(\n \"Final validation loss:\",\n f\"{final_val_loss:.6f}\"\n)\n\nprint()\nprint(\"Per-target validation metrics:\")\n\nprint(\n final_metrics.to_string(\n index=False\n )\n)\n\nprint()\nprint(\n \"Mean F1:\",\n f\"{final_metrics['f1'].mean():.4f}\"\n)\n\nprint(\n \"Mean sensitivity:\",\n f\"{final_metrics['sensitivity'].mean():.4f}\"\n)\n\nprint(\n \"Mean specificity:\",\n f\"{final_metrics['specificity'].mean():.4f}\"\n)\n\nprint(\n \"Mean precision:\",\n f\"{final_metrics['precision'].mean():.4f}\"\n)\n\nprint()\nprint(\n \"Checkpoint saved:\",\n CHECKPOINT_PATH\n)\n\nprint(\n \"History saved:\",\n HISTORY_PATH\n)\n\nprint(\n \"Metrics saved:\",\n METRICS_PATH\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 37 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T09:03:19.08149Z","iopub.execute_input":"2026-08-11T09:03:19.081846Z","iopub.status.idle":"2026-08-11T09:09:05.029324Z","shell.execute_reply.started":"2026-08-11T09:03:19.081814Z","shell.execute_reply":"2026-08-11T09:09:05.028171Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 38 - FOLD 0 BEST MODEL DIAGNOSTIC\n# ================================================================\n\nprint(\"=\" * 70)\nprint(\"CELL 38 - FOLD 0 BEST MODEL DIAGNOSTIC\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Load best checkpoint\n# ---------------------------------------------------------------\n\ncheckpoint = torch.load(\n CHECKPOINT_PATH,\n map_location=DEVICE\n)\n\nmodel.load_state_dict(\n checkpoint[\"model_state_dict\"]\n)\n\nmodel.eval()\n\n\nprint()\nprint(\"Best checkpoint epoch:\")\nprint(\n checkpoint[\"epoch\"]\n)\n\nprint(\n \"Best validation loss:\",\n checkpoint[\"best_val_loss\"]\n)\n\n\n# ---------------------------------------------------------------\n# Run validation\n# ---------------------------------------------------------------\n\n(\n diagnostic_loss,\n diagnostic_labels,\n diagnostic_probabilities,\n diagnostic_predictions\n) = evaluate_fold0(\n model,\n val_loader_fold0,\n criterion_fold,\n DEVICE\n)\n\n\n# ---------------------------------------------------------------\n# Build detailed diagnostics\n# ---------------------------------------------------------------\n\ndiagnostic_rows = []\n\n\nfor index, target in enumerate(\n TARGET_COLUMNS\n):\n\n y_true = diagnostic_labels[\n :,\n index\n ]\n\n y_prob = diagnostic_probabilities[\n :,\n index\n ]\n\n y_pred = diagnostic_predictions[\n :,\n index\n ]\n\n\n actual_positive = int(\n np.sum(\n y_true == 1\n )\n )\n\n actual_negative = int(\n np.sum(\n y_true == 0\n )\n )\n\n predicted_positive = int(\n np.sum(\n y_pred == 1\n )\n )\n\n predicted_negative = int(\n np.sum(\n y_pred == 0\n )\n )\n\n\n tp = int(\n np.sum(\n (y_true == 1)\n &\n (y_pred == 1)\n )\n )\n\n tn = int(\n np.sum(\n (y_true == 0)\n &\n (y_pred == 0)\n )\n )\n\n fp = int(\n np.sum(\n (y_true == 0)\n &\n (y_pred == 1)\n )\n )\n\n fn = int(\n np.sum(\n (y_true == 1)\n &\n (y_pred == 0)\n )\n )\n\n\n sensitivity = (\n tp / (tp + fn)\n if (tp + fn) > 0\n else np.nan\n )\n\n specificity = (\n tn / (tn + fp)\n if (tn + fp) > 0\n else np.nan\n )\n\n precision = (\n tp / (tp + fp)\n if (tp + fp) > 0\n else 0.0\n )\n\n f1 = (\n 2 * precision * sensitivity\n / (precision + sensitivity)\n if (\n precision + sensitivity\n ) > 0\n else 0.0\n )\n\n\n diagnostic_rows.append(\n {\n \"target\": target,\n \"actual_positive\": actual_positive,\n \"actual_negative\": actual_negative,\n \"predicted_positive\": predicted_positive,\n \"predicted_negative\": predicted_negative,\n \"TP\": tp,\n \"TN\": tn,\n \"FP\": fp,\n \"FN\": fn,\n \"mean_probability\": float(\n np.mean(y_prob)\n ),\n \"min_probability\": float(\n np.min(y_prob)\n ),\n \"max_probability\": float(\n np.max(y_prob)\n ),\n \"sensitivity\": sensitivity,\n \"specificity\": specificity,\n \"precision\": precision,\n \"f1\": f1\n }\n )\n\n\ndiagnostic_df = pd.DataFrame(\n diagnostic_rows\n)\n\n\n# ---------------------------------------------------------------\n# Save diagnostic table\n# ---------------------------------------------------------------\n\nDIAGNOSTIC_PATH = os.path.join(\n AUDIT_DIR,\n \"cell38_fold0_best_model_diagnostics.csv\"\n)\n\ndiagnostic_df.to_csv(\n DIAGNOSTIC_PATH,\n index=False\n)\n\n\n# ---------------------------------------------------------------\n# Print diagnostics\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"BEST MODEL VALIDATION DIAGNOSTICS\")\nprint(\"=\" * 70)\n\nprint(\n \"Validation loss:\",\n f\"{diagnostic_loss:.6f}\"\n)\n\nprint()\n\nprint(\n diagnostic_df[\n [\n \"target\",\n \"actual_positive\",\n \"predicted_positive\",\n \"TP\",\n \"TN\",\n \"FP\",\n \"FN\",\n \"sensitivity\",\n \"specificity\",\n \"precision\",\n \"f1\",\n \"mean_probability\"\n ]\n ].to_string(\n index=False\n )\n)\n\n\n# ---------------------------------------------------------------\n# Aggregate diagnostics\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"AGGREGATE RESULTS\")\nprint(\"=\" * 70)\n\nprint(\n \"Mean F1:\",\n f\"{diagnostic_df['f1'].mean():.4f}\"\n)\n\nprint(\n \"Mean sensitivity:\",\n f\"{diagnostic_df['sensitivity'].mean():.4f}\"\n)\n\nprint(\n \"Mean specificity:\",\n f\"{diagnostic_df['specificity'].mean():.4f}\"\n)\n\nprint(\n \"Mean precision:\",\n f\"{diagnostic_df['precision'].mean():.4f}\"\n)\n\nprint(\n \"Total actual positives:\",\n int(\n diagnostic_df[\n \"actual_positive\"\n ].sum()\n )\n)\n\nprint(\n \"Total predicted positives:\",\n int(\n diagnostic_df[\n \"predicted_positive\"\n ].sum()\n )\n)\n\n\n# ---------------------------------------------------------------\n# Prediction-collapse checks\n# ---------------------------------------------------------------\n\nall_negative_targets = diagnostic_df[\n diagnostic_df[\n \"predicted_positive\"\n ] == 0\n][\"target\"].tolist()\n\n\nall_positive_targets = diagnostic_df[\n diagnostic_df[\n \"predicted_positive\"\n ] == len(\n diagnostic_labels\n )\n][\"target\"].tolist()\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"PREDICTION COLLAPSE CHECK\")\nprint(\"=\" * 70)\n\nprint(\n \"All-negative targets:\",\n all_negative_targets\n)\n\nprint(\n \"All-positive targets:\",\n all_positive_targets\n)\n\n\n# ---------------------------------------------------------------\n# Probability distribution\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"PROBABILITY SUMMARY\")\nprint(\"=\" * 70)\n\nfor index, target in enumerate(\n TARGET_COLUMNS\n):\n\n probabilities = (\n diagnostic_probabilities[\n :,\n index\n ]\n )\n\n print(\n f\"{target:20s} \"\n f\"mean={np.mean(probabilities):.4f} \"\n f\"median={np.median(probabilities):.4f} \"\n f\"p10={np.percentile(probabilities, 10):.4f} \"\n f\"p90={np.percentile(probabilities, 90):.4f}\"\n )\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 38 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Saved:\",\n DIAGNOSTIC_PATH\n)\n\nprint()\nprint(\n \"Use this diagnostic before continuing full training.\"\n)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 38 - FOLD 0 BEST MODEL DIAGNOSTIC\n# ================================================================\n\nprint(\"=\" * 70)\nprint(\"CELL 38 - FOLD 0 BEST MODEL DIAGNOSTIC\")\nprint(\"=\" * 70)\n\n\n# ---------------------------------------------------------------\n# Load best checkpoint\n# ---------------------------------------------------------------\n\ncheckpoint = torch.load(\n CHECKPOINT_PATH,\n map_location=DEVICE\n)\n\nmodel.load_state_dict(\n checkpoint[\"model_state_dict\"]\n)\n\nmodel.eval()\n\n\nprint()\nprint(\"Best checkpoint epoch:\")\nprint(\n checkpoint[\"epoch\"]\n)\n\nprint(\n \"Best validation loss:\",\n checkpoint[\"best_val_loss\"]\n)\n\n\n# ---------------------------------------------------------------\n# Run validation\n# ---------------------------------------------------------------\n\n(\n diagnostic_loss,\n diagnostic_labels,\n diagnostic_probabilities,\n diagnostic_predictions\n) = evaluate_fold0(\n model,\n val_loader_fold0,\n criterion_fold,\n DEVICE\n)\n\n\n# ---------------------------------------------------------------\n# Build detailed diagnostics\n# ---------------------------------------------------------------\n\ndiagnostic_rows = []\n\n\nfor index, target in enumerate(\n TARGET_COLUMNS\n):\n\n y_true = diagnostic_labels[\n :,\n index\n ]\n\n y_prob = diagnostic_probabilities[\n :,\n index\n ]\n\n y_pred = diagnostic_predictions[\n :,\n index\n ]\n\n\n actual_positive = int(\n np.sum(\n y_true == 1\n )\n )\n\n actual_negative = int(\n np.sum(\n y_true == 0\n )\n )\n\n predicted_positive = int(\n np.sum(\n y_pred == 1\n )\n )\n\n predicted_negative = int(\n np.sum(\n y_pred == 0\n )\n )\n\n\n tp = int(\n np.sum(\n (y_true == 1)\n &\n (y_pred == 1)\n )\n )\n\n tn = int(\n np.sum(\n (y_true == 0)\n &\n (y_pred == 0)\n )\n )\n\n fp = int(\n np.sum(\n (y_true == 0)\n &\n (y_pred == 1)\n )\n )\n\n fn = int(\n np.sum(\n (y_true == 1)\n &\n (y_pred == 0)\n )\n )\n\n\n sensitivity = (\n tp / (tp + fn)\n if (tp + fn) > 0\n else np.nan\n )\n\n specificity = (\n tn / (tn + fp)\n if (tn + fp) > 0\n else np.nan\n )\n\n precision = (\n tp / (tp + fp)\n if (tp + fp) > 0\n else 0.0\n )\n\n f1 = (\n 2 * precision * sensitivity\n / (precision + sensitivity)\n if (\n precision + sensitivity\n ) > 0\n else 0.0\n )\n\n\n diagnostic_rows.append(\n {\n \"target\": target,\n \"actual_positive\": actual_positive,\n \"actual_negative\": actual_negative,\n \"predicted_positive\": predicted_positive,\n \"predicted_negative\": predicted_negative,\n \"TP\": tp,\n \"TN\": tn,\n \"FP\": fp,\n \"FN\": fn,\n \"mean_probability\": float(\n np.mean(y_prob)\n ),\n \"min_probability\": float(\n np.min(y_prob)\n ),\n \"max_probability\": float(\n np.max(y_prob)\n ),\n \"sensitivity\": sensitivity,\n \"specificity\": specificity,\n \"precision\": precision,\n \"f1\": f1\n }\n )\n\n\ndiagnostic_df = pd.DataFrame(\n diagnostic_rows\n)\n\n\n# ---------------------------------------------------------------\n# Save diagnostic table\n# ---------------------------------------------------------------\n\nDIAGNOSTIC_PATH = os.path.join(\n AUDIT_DIR,\n \"cell38_fold0_best_model_diagnostics.csv\"\n)\n\ndiagnostic_df.to_csv(\n DIAGNOSTIC_PATH,\n index=False\n)\n\n\n# ---------------------------------------------------------------\n# Print diagnostics\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"BEST MODEL VALIDATION DIAGNOSTICS\")\nprint(\"=\" * 70)\n\nprint(\n \"Validation loss:\",\n f\"{diagnostic_loss:.6f}\"\n)\n\nprint()\n\nprint(\n diagnostic_df[\n [\n \"target\",\n \"actual_positive\",\n \"predicted_positive\",\n \"TP\",\n \"TN\",\n \"FP\",\n \"FN\",\n \"sensitivity\",\n \"specificity\",\n \"precision\",\n \"f1\",\n \"mean_probability\"\n ]\n ].to_string(\n index=False\n )\n)\n\n\n# ---------------------------------------------------------------\n# Aggregate diagnostics\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"AGGREGATE RESULTS\")\nprint(\"=\" * 70)\n\nprint(\n \"Mean F1:\",\n f\"{diagnostic_df['f1'].mean():.4f}\"\n)\n\nprint(\n \"Mean sensitivity:\",\n f\"{diagnostic_df['sensitivity'].mean():.4f}\"\n)\n\nprint(\n \"Mean specificity:\",\n f\"{diagnostic_df['specificity'].mean():.4f}\"\n)\n\nprint(\n \"Mean precision:\",\n f\"{diagnostic_df['precision'].mean():.4f}\"\n)\n\nprint(\n \"Total actual positives:\",\n int(\n diagnostic_df[\n \"actual_positive\"\n ].sum()\n )\n)\n\nprint(\n \"Total predicted positives:\",\n int(\n diagnostic_df[\n \"predicted_positive\"\n ].sum()\n )\n)\n\n\n# ---------------------------------------------------------------\n# Prediction-collapse checks\n# ---------------------------------------------------------------\n\nall_negative_targets = diagnostic_df[\n diagnostic_df[\n \"predicted_positive\"\n ] == 0\n][\"target\"].tolist()\n\n\nall_positive_targets = diagnostic_df[\n diagnostic_df[\n \"predicted_positive\"\n ] == len(\n diagnostic_labels\n )\n][\"target\"].tolist()\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"PREDICTION COLLAPSE CHECK\")\nprint(\"=\" * 70)\n\nprint(\n \"All-negative targets:\",\n all_negative_targets\n)\n\nprint(\n \"All-positive targets:\",\n all_positive_targets\n)\n\n\n# ---------------------------------------------------------------\n# Probability distribution\n# ---------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"PROBABILITY SUMMARY\")\nprint(\"=\" * 70)\n\nfor index, target in enumerate(\n TARGET_COLUMNS\n):\n\n probabilities = (\n diagnostic_probabilities[\n :,\n index\n ]\n )\n\n print(\n f\"{target:20s} \"\n f\"mean={np.mean(probabilities):.4f} \"\n f\"median={np.median(probabilities):.4f} \"\n f\"p10={np.percentile(probabilities, 10):.4f} \"\n f\"p90={np.percentile(probabilities, 90):.4f}\"\n )\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 38 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Saved:\",\n DIAGNOSTIC_PATH\n)\n\nprint()\nprint(\n \"Use this diagnostic before continuing full training.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T09:09:36.057689Z","iopub.execute_input":"2026-08-11T09:09:36.058509Z","iopub.status.idle":"2026-08-11T09:09:44.866422Z","shell.execute_reply.started":"2026-08-11T09:09:36.058471Z","shell.execute_reply":"2026-08-11T09:09:44.865314Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 39 - FOLD 0 PROBABILITY SEPARATION AUDIT\n# ================================================================\n\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom sklearn.metrics import roc_auc_score, f1_score\n\nprint(\"=\" * 70)\nprint(\"CELL 39 - FOLD 0 PROBABILITY SEPARATION AUDIT\")\nprint(\"=\" * 70)\n\n# ------------------------------------------------\n# 1. VERIFY EXISTING NOTEBOOK STATE\n# ------------------------------------------------\n\nprint(\"\\nChecking existing notebook objects...\")\n\nprint(\"model:\", \"FOUND\" if \"model\" in globals() else \"MISSING\")\nprint(\"val_loader:\", \"FOUND\" if \"val_loader\" in globals() else \"MISSING\")\nprint(\"TARGETS:\", \"FOUND\" if \"TARGETS\" in globals() else \"MISSING\")\n\nif \"model\" not in globals():\n raise RuntimeError(\"model is missing. Do not retrain. Send me this output.\")\n\nif \"val_loader\" not in globals():\n raise RuntimeError(\"val_loader is missing. Do not retrain. Send me this output.\")\n\nif \"TARGETS\" not in globals():\n raise RuntimeError(\"TARGETS is missing. Do not retrain. Send me this output.\")\n\n# ------------------------------------------------\n# 2. DETERMINE DEVICE FROM THE EXISTING MODEL\n# ------------------------------------------------\n\nDEVICE = next(model.parameters()).device\n\nprint(\"Device detected from model:\", DEVICE)\n\n# ------------------------------------------------\n# 3. LOAD THE BEST FOLD-0 CHECKPOINT\n# ------------------------------------------------\n\ncheckpoint_path = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell37_fold0_best_model.pt\"\n)\n\nprint(\"\\nLoading best Fold-0 checkpoint:\")\nprint(checkpoint_path)\n\ncheckpoint = torch.load(\n checkpoint_path,\n map_location=DEVICE,\n weights_only=False\n)\n\nif isinstance(checkpoint, dict) and \"model_state_dict\" in checkpoint:\n model.load_state_dict(checkpoint[\"model_state_dict\"])\nelse:\n model.load_state_dict(checkpoint)\n\nmodel.to(DEVICE)\nmodel.eval()\n\nprint(\"Best checkpoint loaded: PASS\")\nprint(\"Model set to evaluation mode: PASS\")\n\n# ------------------------------------------------\n# 4. RUN VALIDATION INFERENCE\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"VALIDATION INFERENCE\")\nprint(\"=\" * 70)\n\nall_probabilities = []\nall_labels = []\n\nwith torch.no_grad():\n\n for batch in val_loader:\n\n images = batch[0].to(DEVICE)\n labels = batch[1].to(DEVICE)\n\n logits = model(images)\n\n probs = torch.sigmoid(logits)\n\n all_probabilities.append(\n probs.cpu().numpy()\n )\n\n all_labels.append(\n labels.cpu().numpy()\n )\n\nprobabilities = np.concatenate(\n all_probabilities,\n axis=0\n)\n\nlabels = np.concatenate(\n all_labels,\n axis=0\n)\n\nprint(\"Probability shape:\", probabilities.shape)\nprint(\"Label shape:\", labels.shape)\n\nif probabilities.shape != labels.shape:\n raise RuntimeError(\n f\"Shape mismatch: probabilities={probabilities.shape}, \"\n f\"labels={labels.shape}\"\n )\n\nif probabilities.shape[1] != len(TARGETS):\n raise RuntimeError(\n f\"Expected {len(TARGETS)} targets, \"\n f\"received {probabilities.shape[1]} outputs.\"\n )\n\nif not np.isfinite(probabilities).all():\n raise RuntimeError(\"NaN/Inf detected in probabilities.\")\n\nprint(\"Inference: PASS\")\nprint(\"Shape check: PASS\")\nprint(\"NaN/Inf check: PASS\")\n\n# ------------------------------------------------\n# 5. PROBABILITY SEPARATION\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PROBABILITY SEPARATION\")\nprint(\"=\" * 70)\n\nseparation_results = []\n\nfor i, target in enumerate(TARGETS):\n\n y_true = labels[:, i].astype(int)\n y_prob = probabilities[:, i]\n\n positive_probs = y_prob[y_true == 1]\n negative_probs = y_prob[y_true == 0]\n\n positive_count = len(positive_probs)\n negative_count = len(negative_probs)\n\n positive_mean = np.mean(positive_probs)\n negative_mean = np.mean(negative_probs)\n\n positive_median = np.median(positive_probs)\n negative_median = np.median(negative_probs)\n\n separation = (\n positive_mean - negative_mean\n )\n\n auc = roc_auc_score(\n y_true,\n y_prob\n )\n\n separation_results.append({\n \"target\": target,\n \"positive_count\": positive_count,\n \"negative_count\": negative_count,\n \"positive_mean\": positive_mean,\n \"negative_mean\": negative_mean,\n \"positive_median\": positive_median,\n \"negative_median\": negative_median,\n \"probability_separation\": separation,\n \"roc_auc\": auc\n })\n\nseparation_df = pd.DataFrame(\n separation_results\n)\n\nprint(\n separation_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ------------------------------------------------\n# 6. THRESHOLD AUDIT\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PER-TARGET THRESHOLD AUDIT\")\nprint(\"=\" * 70)\n\nthresholds = np.arange(\n 0.20,\n 0.801,\n 0.025\n)\n\nthreshold_results = []\n\nfor i, target in enumerate(TARGETS):\n\n y_true = labels[:, i].astype(int)\n y_prob = probabilities[:, i]\n\n baseline_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n baseline_f1 = f1_score(\n y_true,\n baseline_pred,\n zero_division=0\n )\n\n best_threshold = 0.50\n best_f1 = baseline_f1\n\n for threshold in thresholds:\n\n predictions = (\n y_prob >= threshold\n ).astype(int)\n\n current_f1 = f1_score(\n y_true,\n predictions,\n zero_division=0\n )\n\n if current_f1 > best_f1:\n\n best_f1 = current_f1\n best_threshold = threshold\n\n threshold_results.append({\n \"target\": target,\n \"f1_at_0.50\": baseline_f1,\n \"best_threshold\": best_threshold,\n \"best_f1\": best_f1,\n \"f1_improvement\": (\n best_f1 - baseline_f1\n )\n })\n\nthreshold_df = pd.DataFrame(\n threshold_results\n)\n\nprint(\n threshold_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ------------------------------------------------\n# 7. SUMMARY\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 39 SUMMARY\")\nprint(\"=\" * 70)\n\nmean_auc = separation_df[\"roc_auc\"].mean()\n\nmean_separation = (\n separation_df[\n \"probability_separation\"\n ].mean()\n)\n\nmean_f1_050 = (\n threshold_df[\"f1_at_0.50\"].mean()\n)\n\nmean_best_f1 = (\n threshold_df[\"best_f1\"].mean()\n)\n\nprint(\n f\"Mean ROC-AUC: {mean_auc:.4f}\"\n)\n\nprint(\n f\"Mean probability separation: {mean_separation:.4f}\"\n)\n\nprint(\n f\"Mean F1 @ 0.50: {mean_f1_050:.4f}\"\n)\n\nprint(\n f\"Mean best-threshold F1: {mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Potential F1 improvement: \"\n f\"{mean_best_f1 - mean_f1_050:.4f}\"\n)\n\nprint(\n \"\\nTargets with ROC-AUC >= 0.65:\",\n int(\n (separation_df[\"roc_auc\"] >= 0.65).sum()\n )\n)\n\nprint(\n \"Targets with negative probability separation:\",\n int(\n (\n separation_df[\n \"probability_separation\"\n ] < 0\n ).sum()\n )\n)\n\nprint(\n \"Targets with >0.05 F1 improvement from threshold:\",\n int(\n (\n threshold_df[\n \"f1_improvement\"\n ] > 0.05\n ).sum()\n )\n)\n\n# ------------------------------------------------\n# 8. SAVE RESULTS\n# ------------------------------------------------\n\noutput_dir = (\n \"/kaggle/working/rsna_knee_audit\"\n)\n\nseparation_path = (\n output_dir +\n \"/cell39_probability_separation.csv\"\n)\n\nthreshold_path = (\n output_dir +\n \"/cell39_threshold_audit.csv\"\n)\n\nseparation_df.to_csv(\n separation_path,\n index=False\n)\n\nthreshold_df.to_csv(\n threshold_path,\n index=False\n)\n\nprint(\"\\nSaved:\")\nprint(separation_path)\nprint(threshold_path)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 39 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T09:41:26.824517Z","iopub.execute_input":"2026-08-11T09:41:26.825132Z","iopub.status.idle":"2026-08-11T09:41:27.950269Z","shell.execute_reply.started":"2026-08-11T09:41:26.825094Z","shell.execute_reply":"2026-08-11T09:41:27.949051Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 39 - PROBABILITY SEPARATION + THRESHOLD AUDIT\n# ================================================================\n\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom sklearn.metrics import roc_auc_score, f1_score\n\nprint(\"=\" * 70)\nprint(\"CELL 39 - PROBABILITY SEPARATION + THRESHOLD AUDIT\")\nprint(\"=\" * 70)\n\n# ------------------------------------------------\n# 1. VERIFY EXISTING NOTEBOOK STATE\n# ------------------------------------------------\n\nprint(\"\\nChecking notebook state...\")\n\nrequired_objects = [\n \"model\",\n \"val_loader\",\n \"TARGETS\"\n]\n\nmissing_objects = [\n name for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n )\n\nprint(\"model : AVAILABLE\")\nprint(\"val_loader : AVAILABLE\")\nprint(\"TARGETS : AVAILABLE\")\n\n# ------------------------------------------------\n# 2. DETERMINE MODEL DEVICE\n# ------------------------------------------------\n\nMODEL_DEVICE = next(\n model.parameters()\n).device\n\nprint(\"Model device:\", MODEL_DEVICE)\n\n# ------------------------------------------------\n# 3. LOAD BEST FOLD-0 CHECKPOINT\n# ------------------------------------------------\n\ncheckpoint_path = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell37_fold0_best_model.pt\"\n)\n\nprint(\"\\nLoading best Fold-0 checkpoint:\")\nprint(checkpoint_path)\n\ncheckpoint = torch.load(\n checkpoint_path,\n map_location=MODEL_DEVICE,\n weights_only=False\n)\n\nif (\n isinstance(checkpoint, dict)\n and \"model_state_dict\" in checkpoint\n):\n model.load_state_dict(\n checkpoint[\"model_state_dict\"]\n )\nelse:\n model.load_state_dict(checkpoint)\n\nmodel.to(MODEL_DEVICE)\nmodel.eval()\n\nprint(\"Best Fold-0 checkpoint: LOADED\")\nprint(\"Model mode: evaluation\")\n\n# ------------------------------------------------\n# 4. INSPECT ACTUAL VAL_LOADER BATCH STRUCTURE\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"INSPECTING VALIDATION BATCH STRUCTURE\")\nprint(\"=\" * 70)\n\nfirst_batch = next(iter(val_loader))\n\nprint(\"Batch type:\", type(first_batch))\n\nif isinstance(first_batch, dict):\n\n print(\"Batch keys:\")\n for key, value in first_batch.items():\n if hasattr(value, \"shape\"):\n print(\n f\" {key}: \"\n f\"type={type(value).__name__}, \"\n f\"shape={tuple(value.shape)}\"\n )\n else:\n print(\n f\" {key}: \"\n f\"type={type(value).__name__}\"\n )\n\nelif isinstance(first_batch, (tuple, list)):\n\n print(\n \"Batch sequence length:\",\n len(first_batch)\n )\n\n for i, value in enumerate(first_batch):\n\n if hasattr(value, \"shape\"):\n print(\n f\" [{i}]: \"\n f\"type={type(value).__name__}, \"\n f\"shape={tuple(value.shape)}\"\n )\n else:\n print(\n f\" [{i}]: \"\n f\"type={type(value).__name__}\"\n )\n\nelse:\n\n raise RuntimeError(\n \"Unsupported validation batch type: \"\n + str(type(first_batch))\n )\n\n# ------------------------------------------------\n# 5. GENERIC BATCH EXTRACTION\n# ------------------------------------------------\n\ndef extract_image_and_label(batch):\n \"\"\"\n Extract image and label tensors from the existing\n Dataset/DataLoader batch without assuming fixed\n dictionary key names or tuple positions.\n \"\"\"\n\n candidates = []\n\n if isinstance(batch, dict):\n\n items = list(batch.items())\n\n for key, value in items:\n\n if torch.is_tensor(value):\n\n candidates.append(\n (str(key), value)\n )\n\n elif isinstance(batch, (tuple, list)):\n\n for idx, value in enumerate(batch):\n\n if torch.is_tensor(value):\n\n candidates.append(\n (str(idx), value)\n )\n\n else:\n\n raise RuntimeError(\n \"Unsupported batch type: \"\n + str(type(batch))\n )\n\n image_candidates = []\n label_candidates = []\n\n for name, tensor in candidates:\n\n if tensor.ndim == 4:\n image_candidates.append(\n (name, tensor)\n )\n\n elif tensor.ndim == 2:\n if tensor.shape[-1] == len(TARGETS):\n label_candidates.append(\n (name, tensor)\n )\n\n if len(image_candidates) == 0:\n\n raise RuntimeError(\n \"Could not identify image tensor. \"\n \"Expected a 4D tensor such as \"\n \"(batch, channels, height, width).\"\n )\n\n if len(label_candidates) == 0:\n\n raise RuntimeError(\n \"Could not identify label tensor. \"\n f\"Expected a 2D tensor with \"\n f\"{len(TARGETS)} targets.\"\n )\n\n image_name, images = image_candidates[0]\n label_name, labels = label_candidates[0]\n\n return (\n images,\n labels,\n image_name,\n label_name\n )\n\nimages_test, labels_test, image_key, label_key = (\n extract_image_and_label(first_batch)\n)\n\nprint(\"\\nIdentified image field:\", image_key)\nprint(\n \"Identified label field:\",\n label_key\n)\n\nprint(\n \"Image batch shape:\",\n tuple(images_test.shape)\n)\n\nprint(\n \"Label batch shape:\",\n tuple(labels_test.shape)\n)\n\n# ------------------------------------------------\n# 6. VALIDATE BATCH STRUCTURE\n# ------------------------------------------------\n\nif images_test.ndim != 4:\n raise RuntimeError(\n \"Image tensor is not 4-dimensional.\"\n )\n\nif labels_test.ndim != 2:\n raise RuntimeError(\n \"Label tensor is not 2-dimensional.\"\n )\n\nif labels_test.shape[1] != len(TARGETS):\n raise RuntimeError(\n f\"Label tensor has {labels_test.shape[1]} \"\n f\"targets, expected {len(TARGETS)}.\"\n )\n\nif images_test.shape[0] != labels_test.shape[0]:\n raise RuntimeError(\n \"Image and label batch sizes do not match.\"\n )\n\nprint(\"Batch structure: PASS\")\n\n# ------------------------------------------------\n# 7. RUN VALIDATION INFERENCE\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"VALIDATION INFERENCE\")\nprint(\"=\" * 70)\n\nall_probabilities = []\nall_labels = []\n\nbatch_count = 0\n\nwith torch.no_grad():\n\n for batch in val_loader:\n\n images, labels, _, _ = (\n extract_image_and_label(batch)\n )\n\n images = images.to(\n MODEL_DEVICE,\n non_blocking=True\n )\n\n labels = labels.to(\n MODEL_DEVICE,\n non_blocking=True\n )\n\n logits = model(images)\n\n probabilities_batch = torch.sigmoid(\n logits\n )\n\n all_probabilities.append(\n probabilities_batch.cpu().numpy()\n )\n\n all_labels.append(\n labels.cpu().numpy()\n )\n\n batch_count += 1\n\nprobabilities = np.concatenate(\n all_probabilities,\n axis=0\n)\n\nlabels = np.concatenate(\n all_labels,\n axis=0\n)\n\nprint(\"Validation batches:\", batch_count)\nprint(\n \"Probability shape:\",\n probabilities.shape\n)\n\nprint(\n \"Label shape:\",\n labels.shape\n)\n\n# ------------------------------------------------\n# 8. FINAL INFERENCE VALIDATION\n# ------------------------------------------------\n\nexpected_shape = (\n probabilities.shape[0],\n len(TARGETS)\n)\n\nif probabilities.shape != expected_shape:\n raise RuntimeError(\n f\"Unexpected probability shape: \"\n f\"{probabilities.shape}\"\n )\n\nif labels.shape != expected_shape:\n raise RuntimeError(\n f\"Unexpected label shape: \"\n f\"{labels.shape}\"\n )\n\nif not np.isfinite(probabilities).all():\n raise RuntimeError(\n \"NaN/Inf detected in probabilities.\"\n )\n\nprint(\"Probability validity: PASS\")\nprint(\"Shape consistency: PASS\")\n\n# ------------------------------------------------\n# 9. PROBABILITY SEPARATION\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PROBABILITY SEPARATION\")\nprint(\"=\" * 70)\n\nseparation_rows = []\n\nfor idx, target in enumerate(TARGETS):\n\n y_true = labels[:, idx].astype(int)\n y_prob = probabilities[:, idx]\n\n positive_probs = y_prob[y_true == 1]\n negative_probs = y_prob[y_true == 0]\n\n positive_count = len(\n positive_probs\n )\n\n negative_count = len(\n negative_probs\n )\n\n if (\n positive_count > 0\n and negative_count > 0\n ):\n\n positive_mean = float(\n positive_probs.mean()\n )\n\n negative_mean = float(\n negative_probs.mean()\n )\n\n positive_median = float(\n np.median(positive_probs)\n )\n\n negative_median = float(\n np.median(negative_probs)\n )\n\n separation = (\n positive_mean\n - negative_mean\n )\n\n auc = float(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n\n else:\n\n positive_mean = np.nan\n negative_mean = np.nan\n positive_median = np.nan\n negative_median = np.nan\n separation = np.nan\n auc = np.nan\n\n separation_rows.append({\n \"target\": target,\n \"positive_count\": positive_count,\n \"negative_count\": negative_count,\n \"positive_mean\": positive_mean,\n \"negative_mean\": negative_mean,\n \"positive_median\": positive_median,\n \"negative_median\": negative_median,\n \"probability_separation\": separation,\n \"roc_auc\": auc\n })\n\nseparation_df = pd.DataFrame(\n separation_rows\n)\n\nprint(\n separation_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ------------------------------------------------\n# 10. THRESHOLD AUDIT\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PER-TARGET THRESHOLD AUDIT\")\nprint(\"=\" * 70)\n\nthreshold_values = np.arange(\n 0.20,\n 0.801,\n 0.025\n)\n\nthreshold_rows = []\n\nfor idx, target in enumerate(TARGETS):\n\n y_true = labels[:, idx].astype(int)\n y_prob = probabilities[:, idx]\n\n f1_at_050 = f1_score(\n y_true,\n (y_prob >= 0.50).astype(int),\n zero_division=0\n )\n\n best_threshold = 0.50\n best_f1 = f1_at_050\n\n for threshold in threshold_values:\n\n predictions = (\n y_prob >= threshold\n ).astype(int)\n\n current_f1 = f1_score(\n y_true,\n predictions,\n zero_division=0\n )\n\n if current_f1 > best_f1:\n\n best_f1 = current_f1\n best_threshold = float(\n threshold\n )\n\n threshold_rows.append({\n \"target\": target,\n \"f1_at_0.50\": float(\n f1_at_050\n ),\n \"best_threshold\": float(\n best_threshold\n ),\n \"best_f1\": float(\n best_f1\n ),\n \"f1_improvement\": float(\n best_f1 - f1_at_050\n )\n })\n\nthreshold_df = pd.DataFrame(\n threshold_rows\n)\n\nprint(\n threshold_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ------------------------------------------------\n# 11. OVERALL SUMMARY\n# ------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 39 SUMMARY\")\nprint(\"=\" * 70)\n\nmean_auc = float(\n separation_df[\"roc_auc\"].mean()\n)\n\nmean_separation = float(\n separation_df[\n \"probability_separation\"\n ].mean()\n)\n\nmean_f1_050 = float(\n threshold_df[\"f1_at_0.50\"].mean()\n)\n\nmean_best_f1 = float(\n threshold_df[\"best_f1\"].mean()\n)\n\nprint(\n f\"Mean ROC-AUC: \"\n f\"{mean_auc:.4f}\"\n)\n\nprint(\n f\"Mean probability separation: \"\n f\"{mean_separation:.4f}\"\n)\n\nprint(\n f\"Mean F1 @ 0.50: \"\n f\"{mean_f1_050:.4f}\"\n)\n\nprint(\n f\"Mean best-threshold F1: \"\n f\"{mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Potential F1 improvement: \"\n f\"{mean_best_f1 - mean_f1_050:.4f}\"\n)\n\nprint(\n \"\\nTargets with ROC-AUC >= 0.65:\",\n int(\n (\n separation_df[\"roc_auc\"]\n >= 0.65\n ).sum()\n )\n)\n\nprint(\n \"Targets with negative probability separation:\",\n int(\n (\n separation_df[\n \"probability_separation\"\n ] < 0\n ).sum()\n )\n)\n\nprint(\n \"Targets with >0.05 F1 improvement:\",\n int(\n (\n threshold_df[\n \"f1_improvement\"\n ] > 0.05\n ).sum()\n )\n)\n\n# ------------------------------------------------\n# 12. SAVE RESULTS\n# ------------------------------------------------\n\noutput_dir = (\n \"/kaggle/working/rsna_knee_audit\"\n)\n\nseparation_path = (\n output_dir +\n \"/cell39_probability_separation.csv\"\n)\n\nthreshold_path = (\n output_dir +\n \"/cell39_threshold_audit.csv\"\n)\n\nseparation_df.to_csv(\n separation_path,\n index=False\n)\n\nthreshold_df.to_csv(\n threshold_path,\n index=False\n)\n\nprint(\"\\nSaved:\")\nprint(separation_path)\nprint(threshold_path)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 39 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T09:44:15.229654Z","iopub.execute_input":"2026-08-11T09:44:15.230439Z","iopub.status.idle":"2026-08-11T09:44:26.175161Z","shell.execute_reply.started":"2026-08-11T09:44:15.230404Z","shell.execute_reply":"2026-08-11T09:44:26.173994Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 40 - FOLD 1 TRAINING\n# ================================================================\n\nimport os\nimport copy\nimport random\nimport numpy as np\nimport pandas as pd\nimport torch\nimport torch.nn as nn\nfrom sklearn.metrics import f1_score\n\nprint(\"=\" * 70)\nprint(\"CELL 40 - FOLD 1 TRAINING\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# CONFIGURATION\n# ----------------------------------------------------------------\n\nSEED = 42\nFOLD = 1\nMAX_EPOCHS = 10\nPATIENCE = 3\nLEARNING_RATE = 0.001\nWEIGHT_DECAY = 0.0001\nMAX_GRAD_NORM = 1.0\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\nCHECKPOINT_PATH = os.path.join(\n AUDIT_DIR, \"cell40_fold1_best_model.pt\"\n)\nHISTORY_PATH = os.path.join(\n AUDIT_DIR, \"cell40_fold1_history.csv\"\n)\nMETRICS_PATH = os.path.join(\n AUDIT_DIR, \"cell40_fold1_metrics.csv\"\n)\n\n# ----------------------------------------------------------------\n# REPRODUCIBILITY\n# ----------------------------------------------------------------\n\nrandom.seed(SEED)\nnp.random.seed(SEED)\ntorch.manual_seed(SEED)\n\nif torch.cuda.is_available():\n torch.cuda.manual_seed_all(SEED)\n\n# ----------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------\n\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\nprint(f\"Fold: {FOLD}\")\nprint(f\"Maximum epochs: {MAX_EPOCHS}\")\nprint(f\"Early stopping patience: {PATIENCE}\")\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------\n# REQUIRED OBJECT CHECK\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"train_loader\",\n \"val_loader\",\n \"TARGETS\",\n]\n\nmissing_objects = [\n name for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# VERIFY MODEL\n# ----------------------------------------------------------------\n\nmodel = model.to(device)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"MODEL STATE\")\nprint(\"=\" * 70)\n\ntotal_params = sum(p.numel() for p in model.parameters())\ntrainable_params = sum(\n p.numel() for p in model.parameters()\n if p.requires_grad\n)\n\nprint(f\"Total parameters: {total_params:,}\")\nprint(f\"Trainable parameters before reset: {trainable_params:,}\")\n\nif total_params != 11_367_372:\n print(\n f\"WARNING: Expected 11,367,372 parameters from the \"\n f\"validated Fold-0 model, found {total_params:,}.\"\n )\n\nif trainable_params != 200_268:\n print(\n f\"WARNING: Expected 200,268 trainable parameters, \"\n f\"found {trainable_params:,}.\"\n )\n\n# ----------------------------------------------------------------\n# CREATE FRESH FOLD-1 MODEL\n#\n# Cell 39 loaded the best Fold-0 checkpoint into `model`.\n# The backbone/frozen parameters were never updated during Fold 0.\n#\n# We therefore copy the validated architecture and reset ONLY\n# parameters belonging to trainable modules.\n# ----------------------------------------------------------------\n\nfold1_model = copy.deepcopy(model)\n\nreset_parameter_count = 0\nreset_module_names = []\n\nfor module_name, module in fold1_model.named_modules():\n\n if not any(\n parameter.requires_grad\n for parameter in module.parameters(recurse=False)\n ):\n continue\n\n if hasattr(module, \"reset_parameters\"):\n module.reset_parameters()\n\n for parameter in module.parameters(recurse=False):\n if parameter.requires_grad:\n reset_parameter_count += parameter.numel()\n\n reset_module_names.append(module_name)\n\nfold1_model = fold1_model.to(device)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"RESETTING MODEL FOR FOLD 1\")\nprint(\"=\" * 70)\n\nprint(\n f\"Trainable parameters after reset: \"\n f\"{sum(p.numel() for p in fold1_model.parameters() if p.requires_grad):,}\"\n)\n\nprint(f\"Trainable modules reset: {len(reset_module_names)}\")\nprint(f\"Parameters reset: {reset_parameter_count:,}\")\n\n# ----------------------------------------------------------------\n# VERIFY TRAINABLE PARAMETER COUNT\n# ----------------------------------------------------------------\n\nfold1_trainable_params = sum(\n p.numel()\n for p in fold1_model.parameters()\n if p.requires_grad\n)\n\nif fold1_trainable_params != trainable_params:\n raise RuntimeError(\n \"Trainable parameter count changed during reset: \"\n f\"{trainable_params:,} -> {fold1_trainable_params:,}\"\n )\n\nprint(\"Trainable parameter count: PASS\")\n\n# ----------------------------------------------------------------\n# CLASS WEIGHTS\n#\n# Recalculate weights strictly from the Fold-1 training studies.\n# This prevents validation labels from influencing the loss.\n# ----------------------------------------------------------------\n\nif \"modeling_df\" in globals():\n fold_table = modeling_df.copy()\nelif \"modeling_table\" in globals():\n fold_table = modeling_table.copy()\nelse:\n fold_table = pd.read_csv(\n os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n )\n )\n\nif \"fold_table\" not in globals():\n raise RuntimeError(\"Unable to load modeling table.\")\n\nif \"fold\" not in fold_table.columns:\n raise RuntimeError(\n \"Modeling table does not contain the required 'fold' column.\"\n )\n\ntrain_fold_df = fold_table[\n fold_table[\"fold\"] != FOLD\n].copy()\n\nval_fold_df = fold_table[\n fold_table[\"fold\"] == FOLD\n].copy()\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD DATA\")\nprint(\"=\" * 70)\n\nprint(f\"Training studies: {len(train_fold_df)}\")\nprint(f\"Validation studies: {len(val_fold_df)}\")\n\nif len(train_fold_df) == 0 or len(val_fold_df) == 0:\n raise RuntimeError(\"Fold 1 has empty training or validation data.\")\n\n# ----------------------------------------------------------------\n# POSITIVE WEIGHTS\n# ----------------------------------------------------------------\n\npositive_weights = []\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 1 CLASS WEIGHTS\")\nprint(\"=\" * 70)\n\nfor target in TARGETS:\n\n if target not in train_fold_df.columns:\n raise RuntimeError(\n f\"Target column missing from modeling table: {target}\"\n )\n\n positives = int(train_fold_df[target].sum())\n negatives = int(len(train_fold_df) - positives)\n\n if positives == 0:\n raise RuntimeError(\n f\"Fold 1 training set has zero positive samples for {target}.\"\n )\n\n if negatives == 0:\n raise RuntimeError(\n f\"Fold 1 training set has zero negative samples for {target}.\"\n )\n\n pos_weight = negatives / positives\n positive_weights.append(pos_weight)\n\n print(\n f\"{target:22s} \"\n f\"positive={positives:2d} \"\n f\"negative={negatives:2d} \"\n f\"pos_weight={pos_weight:.4f}\"\n )\n\npos_weight_tensor = torch.tensor(\n positive_weights,\n dtype=torch.float32,\n device=device\n)\n\n# ----------------------------------------------------------------\n# LOSS\n# ----------------------------------------------------------------\n\ncriterion = nn.BCEWithLogitsLoss(\n pos_weight=pos_weight_tensor\n)\n\n# ----------------------------------------------------------------\n# OPTIMIZER\n# ----------------------------------------------------------------\n\noptimizer = torch.optim.AdamW(\n [\n parameter\n for parameter in fold1_model.parameters()\n if parameter.requires_grad\n ],\n lr=LEARNING_RATE,\n weight_decay=WEIGHT_DECAY\n)\n\nscheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(\n optimizer,\n mode=\"min\",\n factor=0.5,\n patience=1\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TRAINING CONFIGURATION\")\nprint(\"=\" * 70)\n\nprint(f\"Loss: BCEWithLogitsLoss\")\nprint(f\"Optimizer: AdamW\")\nprint(f\"Learning rate: {LEARNING_RATE}\")\nprint(f\"Weight decay: {WEIGHT_DECAY}\")\nprint(f\"Gradient clipping: {MAX_GRAD_NORM}\")\nprint(f\"Scheduler: ReduceLROnPlateau\")\n\n# ----------------------------------------------------------------\n# TRAINING / VALIDATION FUNCTIONS\n# ----------------------------------------------------------------\n\ndef run_training_epoch(model, loader, criterion, optimizer, device):\n model.train()\n\n running_loss = 0.0\n batches = 0\n\n for batch in loader:\n\n images = batch[\"image\"].to(device, non_blocking=True)\n labels = batch[\"label\"].to(device, non_blocking=True)\n\n optimizer.zero_grad(set_to_none=True)\n\n logits = model(images)\n\n loss = criterion(logits, labels)\n\n if not torch.isfinite(loss):\n raise RuntimeError(\n \"Non-finite training loss detected.\"\n )\n\n loss.backward()\n\n torch.nn.utils.clip_grad_norm_(\n model.parameters(),\n MAX_GRAD_NORM\n )\n\n optimizer.step()\n\n running_loss += loss.item()\n batches += 1\n\n if batches == 0:\n raise RuntimeError(\"Training loader produced zero batches.\")\n\n return running_loss / batches\n\n\n@torch.no_grad()\ndef run_validation_epoch(model, loader, criterion, device):\n model.eval()\n\n running_loss = 0.0\n batches = 0\n\n all_probabilities = []\n all_labels = []\n\n for batch in loader:\n\n images = batch[\"image\"].to(device, non_blocking=True)\n labels = batch[\"label\"].to(device, non_blocking=True)\n\n logits = model(images)\n\n loss = criterion(logits, labels)\n\n if not torch.isfinite(loss):\n raise RuntimeError(\n \"Non-finite validation loss detected.\"\n )\n\n probabilities = torch.sigmoid(logits)\n\n running_loss += loss.item()\n batches += 1\n\n all_probabilities.append(\n probabilities.cpu()\n )\n all_labels.append(\n labels.cpu()\n )\n\n if batches == 0:\n raise RuntimeError(\"Validation loader produced zero batches.\")\n\n probabilities = torch.cat(all_probabilities, dim=0).numpy()\n labels = torch.cat(all_labels, dim=0).numpy()\n\n predictions = (probabilities >= 0.5).astype(np.int32)\n\n per_target_f1 = []\n\n for target_index in range(len(TARGETS)):\n f1 = f1_score(\n labels[:, target_index],\n predictions[:, target_index],\n zero_division=0\n )\n per_target_f1.append(f1)\n\n mean_f1 = float(np.mean(per_target_f1))\n\n return (\n running_loss / batches,\n mean_f1,\n probabilities,\n labels\n )\n\n# ----------------------------------------------------------------\n# TRAINING\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"STARTING FOLD 1 TRAINING\")\nprint(\"=\" * 70)\n\nhistory = []\n\nbest_val_loss = float(\"inf\")\nbest_epoch = 0\nepochs_without_improvement = 0\n\nfor epoch in range(1, MAX_EPOCHS + 1):\n\n train_loss = run_training_epoch(\n fold1_model,\n train_loader,\n criterion,\n optimizer,\n device\n )\n\n (\n val_loss,\n val_f1,\n probabilities,\n labels\n ) = run_validation_epoch(\n fold1_model,\n val_loader,\n criterion,\n device\n )\n\n scheduler.step(val_loss)\n\n current_lr = optimizer.param_groups[0][\"lr\"]\n\n history.append({\n \"epoch\": epoch,\n \"train_loss\": train_loss,\n \"val_loss\": val_loss,\n \"mean_f1\": val_f1,\n \"learning_rate\": current_lr\n })\n\n improved = val_loss < best_val_loss\n\n if improved:\n\n best_val_loss = val_loss\n best_epoch = epoch\n epochs_without_improvement = 0\n\n torch.save(\n {\n \"model_state_dict\": fold1_model.state_dict(),\n \"fold\": FOLD,\n \"epoch\": epoch,\n \"val_loss\": val_loss,\n \"targets\": list(TARGETS),\n \"input_channels\": 21,\n \"image_size\": 224,\n \"slices_per_plane\": 7,\n \"planes\": [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\"\n ]\n },\n CHECKPOINT_PATH\n )\n\n else:\n epochs_without_improvement += 1\n\n print(f\"\\nEpoch {epoch:02d}/{MAX_EPOCHS}\")\n print(f\"Train Loss: {train_loss:.5f}\")\n print(f\"Val Loss: {val_loss:.5f}\")\n print(f\"Mean F1: {val_f1:.4f}\")\n print(f\"Learning Rate: {current_lr:.6f}\")\n print(f\"Best Epoch: {best_epoch}\")\n print(f\"Best Val Loss: {best_val_loss:.5f}\")\n\n if epochs_without_improvement >= PATIENCE:\n print(\n f\"\\nEarly stopping triggered after \"\n f\"{PATIENCE} epochs without improvement.\"\n )\n break\n\n# ----------------------------------------------------------------\n# RESTORE BEST CHECKPOINT\n# ----------------------------------------------------------------\n\nif not os.path.exists(CHECKPOINT_PATH):\n raise RuntimeError(\n \"Best Fold-1 checkpoint was not created.\"\n )\n\ncheckpoint = torch.load(\n CHECKPOINT_PATH,\n map_location=device\n)\n\nfold1_model.load_state_dict(\n checkpoint[\"model_state_dict\"]\n)\n\nfold1_model.eval()\n\n# ----------------------------------------------------------------\n# FINAL BEST-MODEL VALIDATION\n# ----------------------------------------------------------------\n\n(\n final_val_loss,\n final_mean_f1,\n final_probabilities,\n final_labels\n) = run_validation_epoch(\n fold1_model,\n val_loader,\n criterion,\n device\n)\n\n# ----------------------------------------------------------------\n# PER-TARGET METRICS\n# ----------------------------------------------------------------\n\nmetrics_rows = []\n\nfor target_index, target in enumerate(TARGETS):\n\n y_true = final_labels[:, target_index]\n y_prob = final_probabilities[:, target_index]\n y_pred = (y_prob >= 0.5).astype(np.int32)\n\n tp = int(((y_true == 1) & (y_pred == 1)).sum())\n tn = int(((y_true == 0) & (y_pred == 0)).sum())\n fp = int(((y_true == 0) & (y_pred == 1)).sum())\n fn = int(((y_true == 1) & (y_pred == 0)).sum())\n\n sensitivity = (\n tp / (tp + fn)\n if (tp + fn) > 0\n else 0.0\n )\n\n specificity = (\n tn / (tn + fp)\n if (tn + fp) > 0\n else 0.0\n )\n\n precision = (\n tp / (tp + fp)\n if (tp + fp) > 0\n else 0.0\n )\n\n f1 = (\n 2 * precision * sensitivity /\n (precision + sensitivity)\n if (precision + sensitivity) > 0\n else 0.0\n )\n\n metrics_rows.append({\n \"target\": target,\n \"positive_count\": int(y_true.sum()),\n \"negative_count\": int(len(y_true) - y_true.sum()),\n \"predicted_positive\": int(y_pred.sum()),\n \"TP\": tp,\n \"TN\": tn,\n \"FP\": fp,\n \"FN\": fn,\n \"sensitivity\": sensitivity,\n \"specificity\": specificity,\n \"precision\": precision,\n \"f1\": f1,\n \"mean_probability\": float(y_prob.mean())\n })\n\nmetrics_df = pd.DataFrame(metrics_rows)\n\n# ----------------------------------------------------------------\n# SAVE HISTORY AND METRICS\n# ----------------------------------------------------------------\n\nhistory_df = pd.DataFrame(history)\n\nhistory_df.to_csv(\n HISTORY_PATH,\n index=False\n)\n\nmetrics_df.to_csv(\n METRICS_PATH,\n index=False\n)\n\n# ----------------------------------------------------------------\n# FINAL OUTPUT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 1 TRAINING COMPLETE\")\nprint(\"=\" * 70)\n\nprint(f\"Best epoch: {best_epoch}\")\nprint(f\"Best validation loss: {best_val_loss:.6f}\")\nprint(f\"Final validation loss: {final_val_loss:.6f}\")\n\nprint(\"\\nPer-target validation metrics:\")\nprint(\n metrics_df[\n [\n \"target\",\n \"positive_count\",\n \"negative_count\",\n \"predicted_positive\",\n \"sensitivity\",\n \"specificity\",\n \"precision\",\n \"f1\"\n ]\n ].to_string(index=False)\n)\n\nprint(f\"\\nMean F1: {metrics_df['f1'].mean():.4f}\")\nprint(f\"Mean sensitivity: {metrics_df['sensitivity'].mean():.4f}\")\nprint(f\"Mean specificity: {metrics_df['specificity'].mean():.4f}\")\nprint(f\"Mean precision: {metrics_df['precision'].mean():.4f}\")\n\nprint(f\"\\nCheckpoint saved: {CHECKPOINT_PATH}\")\nprint(f\"History saved: {HISTORY_PATH}\")\nprint(f\"Metrics saved: {METRICS_PATH}\")\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 40 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T09:50:09.17935Z","iopub.execute_input":"2026-08-11T09:50:09.179673Z","iopub.status.idle":"2026-08-11T09:52:20.831026Z","shell.execute_reply.started":"2026-08-11T09:50:09.179647Z","shell.execute_reply":"2026-08-11T09:52:20.829974Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 41 - FOLD 1 PROBABILITY SEPARATION + THRESHOLD AUDIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom sklearn.metrics import roc_auc_score, f1_score\n\nprint(\"=\" * 70)\nprint(\"CELL 41 - FOLD 1 PROBABILITY SEPARATION + THRESHOLD AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# CONFIGURATION\n# ----------------------------------------------------------------\n\nFOLD = 1\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nCHECKPOINT_PATH = os.path.join(\n AUDIT_DIR,\n \"cell40_fold1_best_model.pt\"\n)\n\nSEPARATION_PATH = os.path.join(\n AUDIT_DIR,\n \"cell41_fold1_probability_separation.csv\"\n)\n\nTHRESHOLD_PATH = os.path.join(\n AUDIT_DIR,\n \"cell41_fold1_threshold_audit.csv\"\n)\n\n# ----------------------------------------------------------------\n# CHECK REQUIRED OBJECTS\n# ----------------------------------------------------------------\n\nprint(\"\\nChecking notebook state...\")\n\nrequired_objects = [\n \"fold1_model\",\n \"val_loader\",\n \"TARGETS\"\n]\n\nmissing_objects = [\n name for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"fold1_model : AVAILABLE\")\nprint(\"val_loader : AVAILABLE\")\nprint(\"TARGETS : AVAILABLE\")\n\n# ----------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------\n\ndevice = next(fold1_model.parameters()).device\n\nprint(f\"Model device: {device}\")\n\n# ----------------------------------------------------------------\n# CHECKPOINT\n# ----------------------------------------------------------------\n\nif not os.path.exists(CHECKPOINT_PATH):\n raise RuntimeError(\n f\"Fold-1 checkpoint not found: {CHECKPOINT_PATH}\"\n )\n\nprint(\"\\nLoading best Fold-1 checkpoint:\")\nprint(CHECKPOINT_PATH)\n\ncheckpoint = torch.load(\n CHECKPOINT_PATH,\n map_location=device\n)\n\nfold1_model.load_state_dict(\n checkpoint[\"model_state_dict\"]\n)\n\nfold1_model.eval()\n\nprint(\"Best Fold-1 checkpoint: LOADED\")\nprint(\"Model mode: evaluation\")\n\n# ----------------------------------------------------------------\n# COLLECT VALIDATION PREDICTIONS\n# ----------------------------------------------------------------\n\nall_probabilities = []\nall_labels = []\n\nbatch_count = 0\n\nwith torch.no_grad():\n\n for batch in val_loader:\n\n if not isinstance(batch, dict):\n raise RuntimeError(\n f\"Unexpected batch type: {type(batch)}\"\n )\n\n if \"image\" not in batch or \"label\" not in batch:\n raise RuntimeError(\n \"Validation batch must contain 'image' and 'label'.\"\n )\n\n images = batch[\"image\"].to(device)\n labels = batch[\"label\"].to(device)\n\n logits = fold1_model(images)\n probabilities = torch.sigmoid(logits)\n\n if not torch.isfinite(probabilities).all():\n raise RuntimeError(\n \"NaN/Inf detected in validation probabilities.\"\n )\n\n all_probabilities.append(\n probabilities.cpu()\n )\n\n all_labels.append(\n labels.cpu()\n )\n\n batch_count += 1\n\nif batch_count == 0:\n raise RuntimeError(\n \"Validation loader produced zero batches.\"\n )\n\nprobabilities = torch.cat(\n all_probabilities,\n dim=0\n).numpy()\n\nlabels = torch.cat(\n all_labels,\n dim=0\n).numpy()\n\nprint(\"\\nValidation batches:\", batch_count)\nprint(\"Probability shape:\", probabilities.shape)\nprint(\"Label shape:\", labels.shape)\n\nif probabilities.shape != labels.shape:\n raise RuntimeError(\n \"Probability and label shapes do not match.\"\n )\n\nif probabilities.shape[1] != len(TARGETS):\n raise RuntimeError(\n \"Number of prediction columns does not match TARGETS.\"\n )\n\nprint(\"Probability validity: PASS\")\nprint(\"Shape consistency: PASS\")\n\n# ----------------------------------------------------------------\n# PROBABILITY SEPARATION + ROC-AUC\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PROBABILITY SEPARATION\")\nprint(\"=\" * 70)\n\nseparation_rows = []\n\nfor index, target in enumerate(TARGETS):\n\n y_true = labels[:, index]\n y_prob = probabilities[:, index]\n\n positive_probabilities = y_prob[y_true == 1]\n negative_probabilities = y_prob[y_true == 0]\n\n positive_count = len(positive_probabilities)\n negative_count = len(negative_probabilities)\n\n if positive_count == 0 or negative_count == 0:\n roc_auc = np.nan\n probability_separation = np.nan\n positive_mean = np.nan\n negative_mean = np.nan\n positive_median = np.nan\n negative_median = np.nan\n else:\n\n positive_mean = float(\n positive_probabilities.mean()\n )\n\n negative_mean = float(\n negative_probabilities.mean()\n )\n\n positive_median = float(\n np.median(positive_probabilities)\n )\n\n negative_median = float(\n np.median(negative_probabilities)\n )\n\n probability_separation = (\n positive_mean - negative_mean\n )\n\n roc_auc = float(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n\n separation_rows.append({\n \"target\": target,\n \"positive_count\": positive_count,\n \"negative_count\": negative_count,\n \"positive_mean\": positive_mean,\n \"negative_mean\": negative_mean,\n \"positive_median\": positive_median,\n \"negative_median\": negative_median,\n \"probability_separation\": probability_separation,\n \"roc_auc\": roc_auc\n })\n\nseparation_df = pd.DataFrame(\n separation_rows\n)\n\nprint(\n separation_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# THRESHOLD AUDIT\n#\n# This is diagnostic only.\n# We DO NOT permanently adopt these thresholds.\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"THRESHOLD AUDIT\")\nprint(\"=\" * 70)\n\nthresholds = np.arange(\n 0.20,\n 0.801,\n 0.025\n)\n\nthreshold_rows = []\n\nfor index, target in enumerate(TARGETS):\n\n y_true = labels[:, index]\n y_prob = probabilities[:, index]\n\n f1_at_050 = f1_score(\n y_true,\n (y_prob >= 0.50).astype(np.int32),\n zero_division=0\n )\n\n best_f1 = -1.0\n best_threshold = 0.50\n\n for threshold in thresholds:\n\n y_pred = (\n y_prob >= threshold\n ).astype(np.int32)\n\n score = f1_score(\n y_true,\n y_pred,\n zero_division=0\n )\n\n if score > best_f1:\n best_f1 = float(score)\n best_threshold = float(threshold)\n\n threshold_rows.append({\n \"target\": target,\n \"f1_at_0.50\": float(f1_at_050),\n \"best_threshold\": best_threshold,\n \"best_f1\": best_f1,\n \"f1_improvement\": best_f1 - float(f1_at_050)\n })\n\nthreshold_df = pd.DataFrame(\n threshold_rows\n)\n\nprint(\n threshold_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# SUMMARY\n# ----------------------------------------------------------------\n\nmean_roc_auc = float(\n separation_df[\"roc_auc\"].mean()\n)\n\nmean_separation = float(\n separation_df[\n \"probability_separation\"\n ].mean()\n)\n\nmean_f1_050 = float(\n threshold_df[\"f1_at_0.50\"].mean()\n)\n\nmean_best_f1 = float(\n threshold_df[\"best_f1\"].mean()\n)\n\npotential_improvement = (\n mean_best_f1 - mean_f1_050\n)\n\nnegative_separation_targets = (\n separation_df[\n separation_df[\"probability_separation\"] < 0\n ][\"target\"].tolist()\n)\n\nhigh_auc_targets = (\n separation_df[\n separation_df[\"roc_auc\"] >= 0.65\n ][\"target\"].tolist()\n)\n\nlarge_threshold_improvement_targets = (\n threshold_df[\n threshold_df[\"f1_improvement\"] > 0.05\n ][\"target\"].tolist()\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 1 DIAGNOSTIC SUMMARY\")\nprint(\"=\" * 70)\n\nprint(f\"Mean ROC-AUC: {mean_roc_auc:.4f}\")\nprint(\n f\"Mean probability separation: \"\n f\"{mean_separation:.4f}\"\n)\nprint(f\"Mean F1 @ 0.50: {mean_f1_050:.4f}\")\nprint(\n f\"Mean best-threshold F1: \"\n f\"{mean_best_f1:.4f}\"\n)\nprint(\n f\"Potential F1 improvement: \"\n f\"{potential_improvement:.4f}\"\n)\n\nprint(\n f\"\\nTargets with ROC-AUC >= 0.65: \"\n f\"{len(high_auc_targets)}\"\n)\n\nif high_auc_targets:\n print(high_auc_targets)\n\nprint(\n f\"\\nTargets with negative probability separation: \"\n f\"{len(negative_separation_targets)}\"\n)\n\nif negative_separation_targets:\n print(negative_separation_targets)\n\nprint(\n f\"\\nTargets with >0.05 F1 improvement: \"\n f\"{len(large_threshold_improvement_targets)}\"\n)\n\nif large_threshold_improvement_targets:\n print(large_threshold_improvement_targets)\n\n# ----------------------------------------------------------------\n# SAVE RESULTS\n# ----------------------------------------------------------------\n\nseparation_df.to_csv(\n SEPARATION_PATH,\n index=False\n)\n\nthreshold_df.to_csv(\n THRESHOLD_PATH,\n index=False\n)\n\nprint(\"\\nSaved:\")\nprint(SEPARATION_PATH)\nprint(THRESHOLD_PATH)\n\n# ----------------------------------------------------------------\n# COMPLETION\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 41 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:01:26.936473Z","iopub.execute_input":"2026-08-11T10:01:26.936871Z","iopub.status.idle":"2026-08-11T10:01:38.318813Z","shell.execute_reply.started":"2026-08-11T10:01:26.936838Z","shell.execute_reply":"2026-08-11T10:01:38.317858Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 42 - FOLD 2 DATA PREPARATION\n# ================================================================\n\nimport os\nimport pandas as pd\nimport numpy as np\nimport torch\nfrom torch.utils.data import DataLoader\n\nprint(\"=\" * 70)\nprint(\"CELL 42 - FOLD 2 DATA PREPARATION\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# CONFIGURATION\n# ----------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nMODELING_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nFOLDS_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_study_folds.csv\"\n)\n\nFOLD = 2\nBATCH_SIZE = 2\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"KneeStudyDataset\",\n \"TARGETS\",\n \"model\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required dataset/model objects: PASS\")\n\n# ----------------------------------------------------------------\n# LOAD EXACT SAVED TABLES\n# ----------------------------------------------------------------\n\nif not os.path.exists(MODELING_PATH):\n raise RuntimeError(\n f\"Missing modeling table: {MODELING_PATH}\"\n )\n\nif not os.path.exists(FOLDS_PATH):\n raise RuntimeError(\n f\"Missing fold table: {FOLDS_PATH}\"\n )\n\nmodeling_table = pd.read_csv(\n MODELING_PATH\n)\n\nstudy_folds = pd.read_csv(\n FOLDS_PATH\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"RESTORED MODELING DATA\")\nprint(\"=\" * 70)\n\nprint(\n f\"Modeling table shape: \"\n f\"{modeling_table.shape}\"\n)\n\nprint(\n f\"Fold table shape: \"\n f\"{study_folds.shape}\"\n)\n\n# ----------------------------------------------------------------\n# EXACT SCHEMA INSPECTION\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 29 MODELING TABLE COLUMNS\")\nprint(\"=\" * 70)\n\nfor index, column in enumerate(\n modeling_table.columns\n):\n print(\n f\"{index:02d}: {column}\"\n )\n\n# ----------------------------------------------------------------\n# TARGET VALIDATION\n# ----------------------------------------------------------------\n\nmissing_targets = [\n target\n for target in TARGETS\n if target not in modeling_table.columns\n]\n\nif missing_targets:\n raise RuntimeError(\n \"Missing target columns: \"\n + \", \".join(missing_targets)\n )\n\nprint(\"\\nTarget columns: PASS\")\n\n# ----------------------------------------------------------------\n# STUDY UID VALIDATION\n# ----------------------------------------------------------------\n\nif \"StudyInstanceUID\" not in modeling_table.columns:\n raise RuntimeError(\n \"StudyInstanceUID is missing from modeling table.\"\n )\n\nif \"StudyInstanceUID\" not in study_folds.columns:\n raise RuntimeError(\n \"StudyInstanceUID is missing from fold table.\"\n )\n\nif \"fold\" not in study_folds.columns:\n raise RuntimeError(\n \"fold column is missing from fold table.\"\n )\n\n# ----------------------------------------------------------------\n# RESTORE FOLD ASSIGNMENT\n# ----------------------------------------------------------------\n\nstudy_folds[\"fold\"] = (\n study_folds[\"fold\"]\n .astype(int)\n)\n\nfold2_validation_ids = set(\n study_folds.loc[\n study_folds[\"fold\"] == FOLD,\n \"StudyInstanceUID\"\n ]\n)\n\nfold2_training_ids = set(\n study_folds.loc[\n study_folds[\"fold\"] != FOLD,\n \"StudyInstanceUID\"\n ]\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 2 STUDY SPLIT\")\nprint(\"=\" * 70)\n\nprint(\n f\"Training studies: \"\n f\"{len(fold2_training_ids)}\"\n)\n\nprint(\n f\"Validation studies: \"\n f\"{len(fold2_validation_ids)}\"\n)\n\noverlap = (\n fold2_training_ids\n .intersection(\n fold2_validation_ids\n )\n)\n\nprint(\n f\"Train/validation overlap: \"\n f\"{len(overlap)}\"\n)\n\nif overlap:\n raise RuntimeError(\n \"Fold-2 train/validation study leakage detected.\"\n )\n\nif len(fold2_training_ids) != 39:\n raise RuntimeError(\n \"Expected 39 Fold-2 training studies.\"\n )\n\nif len(fold2_validation_ids) != 19:\n raise RuntimeError(\n \"Expected 19 Fold-2 validation studies.\"\n )\n\nprint(\"Fold-2 split: PASS\")\n\n# ----------------------------------------------------------------\n# BUILD EXACT FOLD TABLES\n# ----------------------------------------------------------------\n\nfold2_train_table = modeling_table[\n modeling_table[\"StudyInstanceUID\"].isin(\n fold2_training_ids\n )\n].copy()\n\nfold2_val_table = modeling_table[\n modeling_table[\"StudyInstanceUID\"].isin(\n fold2_validation_ids\n )\n].copy()\n\nif len(fold2_train_table) != 39:\n raise RuntimeError(\n \"Fold-2 training table does not contain 39 rows.\"\n )\n\nif len(fold2_val_table) != 19:\n raise RuntimeError(\n \"Fold-2 validation table does not contain 19 rows.\"\n )\n\n# ----------------------------------------------------------------\n# IMPORTANT:\n# DO NOT GUESS PRIMARY SERIES COLUMN NAMES.\n#\n# Instead, identify the actual columns from Cell 29.\n# ----------------------------------------------------------------\n\nseries_uid_columns = [\n column\n for column in modeling_table.columns\n if (\n \"SeriesInstanceUID\" in column\n or \"series\" in column.lower()\n and \"uid\" in column.lower()\n )\n]\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"SERIES UID COLUMNS FOUND\")\nprint(\"=\" * 70)\n\nfor column in series_uid_columns:\n print(column)\n\nif len(series_uid_columns) == 0:\n print(\n \"No series UID columns detected by name.\"\n )\n\n# ----------------------------------------------------------------\n# EXISTING DATASET CLASS CHECK\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"EXISTING DATASET CLASS\")\nprint(\"=\" * 70)\n\nprint(\n f\"KneeStudyDataset: \"\n f\"{KneeStudyDataset}\"\n)\n\n# ----------------------------------------------------------------\n# DO NOT CREATE DATASETS YET.\n#\n# We first verify the exact Cell-29 schema. This prevents another\n# fabricated column-name error and preserves the Cell-33 pipeline.\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 42 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Modeling table restored: PASS\"\n)\n\nprint(\n \"Fold table restored: PASS\"\n)\n\nprint(\n \"Fold-2 split: PASS\"\n)\n\nprint(\n \"Exact modeling-table schema printed: PASS\"\n)\n\nprint(\n \"No fabricated primary-series columns used.\"\n)\n\nprint(\n \"No training performed.\"\n)\n\nprint(\n \"No dataset pipeline modified.\"\n)\n\nprint(\"\\nNext step requires the actual column names printed above.\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:12:12.899139Z","iopub.execute_input":"2026-08-11T10:12:12.899518Z","iopub.status.idle":"2026-08-11T10:12:12.933144Z","shell.execute_reply.started":"2026-08-11T10:12:12.899485Z","shell.execute_reply":"2026-08-11T10:12:12.931788Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 43 - FOLD 2 DATASET + DATALOADER\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom torch.utils.data import DataLoader\n\nprint(\"=\" * 70)\nprint(\"CELL 43 - FOLD 2 DATASET + DATALOADER\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# CONFIGURATION\n# ----------------------------------------------------------------\n\nFOLD = 2\nBATCH_SIZE = 2\n\nIMAGE_SIZE = 224\nSLICES_PER_PLANE = 7\n\nEXPECTED_CHANNELS = 21\n\nPLANE_COLUMNS = {\n \"Sagittal\": \"Sagittal_SeriesInstanceUID\",\n \"Coronal\": \"Coronal_SeriesInstanceUID\",\n \"Axial\": \"Axial_SeriesInstanceUID\"\n}\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"modeling_table\",\n \"study_folds\",\n \"KneeStudyDataset\",\n \"TARGETS\",\n \"model\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# VERIFY EXACT MODELING TABLE SCHEMA\n# ----------------------------------------------------------------\n\nrequired_columns = [\n \"StudyInstanceUID\",\n \"Sagittal_SeriesInstanceUID\",\n \"Coronal_SeriesInstanceUID\",\n \"Axial_SeriesInstanceUID\",\n \"fold\"\n] + list(TARGETS)\n\nmissing_columns = [\n column\n for column in required_columns\n if column not in modeling_table.columns\n]\n\nif missing_columns:\n raise RuntimeError(\n \"Missing required modeling-table columns: \"\n + \", \".join(missing_columns)\n )\n\nprint(\"Modeling-table schema: PASS\")\n\n# ----------------------------------------------------------------\n# VERIFY PRIMARY SERIES COMPLETENESS\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PRIMARY SERIES VALIDATION\")\nprint(\"=\" * 70)\n\nfor plane, column in PLANE_COLUMNS.items():\n\n missing_count = (\n modeling_table[column]\n .isna()\n .sum()\n )\n\n empty_count = (\n modeling_table[column]\n .astype(str)\n .str.strip()\n .eq(\"\")\n .sum()\n )\n\n print(\n f\"{plane:8s}: \"\n f\"missing={missing_count}, \"\n f\"empty={empty_count}, \"\n f\"unique={modeling_table[column].nunique()}\"\n )\n\n if missing_count > 0 or empty_count > 0:\n raise RuntimeError(\n f\"{plane} primary series contains missing values.\"\n )\n\nprint(\"All three primary planes: PASS\")\n\n# ----------------------------------------------------------------\n# RESTORE FOLD-2 STUDY IDS\n# ----------------------------------------------------------------\n\nfold2_validation_ids = set(\n study_folds.loc[\n study_folds[\"fold\"].astype(int) == FOLD,\n \"StudyInstanceUID\"\n ]\n)\n\nfold2_training_ids = set(\n study_folds.loc[\n study_folds[\"fold\"].astype(int) != FOLD,\n \"StudyInstanceUID\"\n ]\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 2 STUDY SPLIT\")\nprint(\"=\" * 70)\n\nprint(\n f\"Training studies: \"\n f\"{len(fold2_training_ids)}\"\n)\n\nprint(\n f\"Validation studies: \"\n f\"{len(fold2_validation_ids)}\"\n)\n\noverlap = (\n fold2_training_ids\n .intersection(fold2_validation_ids)\n)\n\nprint(\n f\"Train/validation overlap: \"\n f\"{len(overlap)}\"\n)\n\nif len(fold2_training_ids) != 39:\n raise RuntimeError(\n f\"Expected 39 training studies, \"\n f\"found {len(fold2_training_ids)}.\"\n )\n\nif len(fold2_validation_ids) != 19:\n raise RuntimeError(\n f\"Expected 19 validation studies, \"\n f\"found {len(fold2_validation_ids)}.\"\n )\n\nif overlap:\n raise RuntimeError(\n \"Train/validation study overlap detected.\"\n )\n\nprint(\"Fold-2 split: PASS\")\n\n# ----------------------------------------------------------------\n# CREATE FOLD-2 TABLES\n# ----------------------------------------------------------------\n\nfold2_train_table = modeling_table[\n modeling_table[\"StudyInstanceUID\"].isin(\n fold2_training_ids\n )\n].copy()\n\nfold2_val_table = modeling_table[\n modeling_table[\"StudyInstanceUID\"].isin(\n fold2_validation_ids\n )\n].copy()\n\n# Preserve deterministic ordering\nfold2_train_table = (\n fold2_train_table\n .sort_values(\"StudyInstanceUID\")\n .reset_index(drop=True)\n)\n\nfold2_val_table = (\n fold2_val_table\n .sort_values(\"StudyInstanceUID\")\n .reset_index(drop=True)\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 2 TABLES\")\nprint(\"=\" * 70)\n\nprint(\n f\"Training table: \"\n f\"{fold2_train_table.shape}\"\n)\n\nprint(\n f\"Validation table: \"\n f\"{fold2_val_table.shape}\"\n)\n\nif len(fold2_train_table) != 39:\n raise RuntimeError(\n \"Fold-2 training table must contain 39 studies.\"\n )\n\nif len(fold2_val_table) != 19:\n raise RuntimeError(\n \"Fold-2 validation table must contain 19 studies.\"\n )\n\n# ----------------------------------------------------------------\n# VERIFY TARGET VALUES\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TARGET VALIDATION\")\nprint(\"=\" * 70)\n\nfor target in TARGETS:\n\n train_values = set(\n fold2_train_table[target]\n .astype(float)\n .unique()\n )\n\n val_values = set(\n fold2_val_table[target]\n .astype(float)\n .unique()\n )\n\n if not train_values.issubset({0.0, 1.0}):\n raise RuntimeError(\n f\"Invalid training values for {target}: \"\n f\"{train_values}\"\n )\n\n if not val_values.issubset({0.0, 1.0}):\n raise RuntimeError(\n f\"Invalid validation values for {target}: \"\n f\"{val_values}\"\n )\n\n print(\n f\"{target:20s} \"\n f\"train={sorted(train_values)} \"\n f\"val={sorted(val_values)}\"\n )\n\nprint(\"Target values: PASS\")\n\n# ----------------------------------------------------------------\n# CONSTRUCT DATASETS USING THE EXISTING CELL-33 CLASS\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CREATING FOLD 2 DATASETS\")\nprint(\"=\" * 70)\n\ntrain_dataset = KneeStudyDataset(\n fold2_train_table,\n targets=TARGETS\n)\n\nval_dataset = KneeStudyDataset(\n fold2_val_table,\n targets=TARGETS\n)\n\nprint(\n f\"Train Dataset: \"\n f\"{len(train_dataset)}\"\n)\n\nprint(\n f\"Validation Dataset: \"\n f\"{len(val_dataset)}\"\n)\n\nif len(train_dataset) != 39:\n raise RuntimeError(\n \"Fold-2 training dataset length is not 39.\"\n )\n\nif len(val_dataset) != 19:\n raise RuntimeError(\n \"Fold-2 validation dataset length is not 19.\"\n )\n\n# ----------------------------------------------------------------\n# CREATE DATALOADERS\n# ----------------------------------------------------------------\n\ntrain_loader = DataLoader(\n train_dataset,\n batch_size=BATCH_SIZE,\n shuffle=True,\n num_workers=0,\n pin_memory=False\n)\n\nval_loader = DataLoader(\n val_dataset,\n batch_size=BATCH_SIZE,\n shuffle=False,\n num_workers=0,\n pin_memory=False\n)\n\nprint(\n f\"Train batches: \"\n f\"{len(train_loader)}\"\n)\n\nprint(\n f\"Validation batches: \"\n f\"{len(val_loader)}\"\n)\n\n# ----------------------------------------------------------------\n# DATALOADER SMOKE TEST\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 2 DATALOADER SMOKE TEST\")\nprint(\"=\" * 70)\n\nbatch = next(\n iter(train_loader)\n)\n\nif not isinstance(batch, dict):\n raise RuntimeError(\n f\"Expected dictionary batch, \"\n f\"received {type(batch)}.\"\n )\n\nprint(\n \"Batch keys:\",\n list(batch.keys())\n)\n\nif \"image\" not in batch:\n raise RuntimeError(\n \"Batch is missing image field.\"\n )\n\nif \"label\" not in batch:\n raise RuntimeError(\n \"Batch is missing label field.\"\n )\n\nimages = batch[\"image\"]\nlabels = batch[\"label\"]\n\nprint(\n f\"Image batch shape: \"\n f\"{tuple(images.shape)}\"\n)\n\nprint(\n f\"Label batch shape: \"\n f\"{tuple(labels.shape)}\"\n)\n\nexpected_image_shape = (\n BATCH_SIZE,\n EXPECTED_CHANNELS,\n IMAGE_SIZE,\n IMAGE_SIZE\n)\n\nexpected_label_shape = (\n BATCH_SIZE,\n len(TARGETS)\n)\n\nif tuple(images.shape) != expected_image_shape:\n raise RuntimeError(\n f\"Unexpected image shape. \"\n f\"Expected {expected_image_shape}, \"\n f\"got {tuple(images.shape)}.\"\n )\n\nif tuple(labels.shape) != expected_label_shape:\n raise RuntimeError(\n f\"Unexpected label shape. \"\n f\"Expected {expected_label_shape}, \"\n f\"got {tuple(labels.shape)}.\"\n )\n\nif images.dtype != torch.float32:\n raise RuntimeError(\n f\"Expected float32 images, \"\n f\"got {images.dtype}.\"\n )\n\nif labels.dtype != torch.float32:\n raise RuntimeError(\n f\"Expected float32 labels, \"\n f\"got {labels.dtype}.\"\n )\n\nif not torch.isfinite(images).all():\n raise RuntimeError(\n \"Image batch contains NaN or Inf.\"\n )\n\nif not torch.isfinite(labels).all():\n raise RuntimeError(\n \"Label batch contains NaN or Inf.\"\n )\n\n# ----------------------------------------------------------------\n# VALUE RANGE CHECK\n# ----------------------------------------------------------------\n\nprint(\n f\"Image min: \"\n f\"{images.min().item():.6f}\"\n)\n\nprint(\n f\"Image max: \"\n f\"{images.max().item():.6f}\"\n)\n\nprint(\n f\"Image mean: \"\n f\"{images.mean().item():.6f}\"\n)\n\nprint(\n f\"Image std: \"\n f\"{images.std().item():.6f}\"\n)\n\nif images.min().item() < -1e-6:\n raise RuntimeError(\n \"Image values below expected [0, 1] range.\"\n )\n\nif images.max().item() > 1.000001:\n raise RuntimeError(\n \"Image values above expected [0, 1] range.\"\n )\n\n# ----------------------------------------------------------------\n# STUDY IDS\n# ----------------------------------------------------------------\n\nif \"study_id\" in batch:\n print(\n \"First study IDs:\",\n batch[\"study_id\"]\n )\n\n# ----------------------------------------------------------------\n# MEMORY ESTIMATE\n# ----------------------------------------------------------------\n\nbytes_per_study = (\n EXPECTED_CHANNELS\n * IMAGE_SIZE\n * IMAGE_SIZE\n * 4\n)\n\nmb_per_study = (\n bytes_per_study / (1024 ** 2)\n)\n\nmb_per_batch = (\n mb_per_study * BATCH_SIZE\n)\n\nprint(\n f\"Approximate image memory per study: \"\n f\"{mb_per_study:.2f} MB\"\n)\n\nprint(\n f\"Approximate image memory per batch: \"\n f\"{mb_per_batch:.2f} MB\"\n)\n\n# ----------------------------------------------------------------\n# FINAL VERDICT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 43 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"Fold-2 study split: PASS\")\nprint(\"Fold-2 datasets: PASS\")\nprint(\"Fold-2 DataLoaders: PASS\")\nprint(\"21-channel input: PASS\")\nprint(\"12-target labels: PASS\")\nprint(\"Image dtype/range: PASS\")\nprint(\"NaN/Inf check: PASS\")\nprint(\"Lazy loading preserved: PASS\")\nprint(\"No model training performed.\")\nprint(\"Ready for Fold-2 model reset and training.\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:13:35.881115Z","iopub.execute_input":"2026-08-11T10:13:35.881438Z","iopub.status.idle":"2026-08-11T10:13:36.714517Z","shell.execute_reply.started":"2026-08-11T10:13:35.881412Z","shell.execute_reply":"2026-08-11T10:13:36.713196Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 44 - FOLD 2 TRAINING\n# ================================================================\n\nimport os\nimport random\nimport numpy as np\nimport pandas as pd\nimport torch\nimport torch.nn as nn\nfrom torch.optim import AdamW\nfrom torch.optim.lr_scheduler import ReduceLROnPlateau\nfrom sklearn.metrics import f1_score\n\nprint(\"=\" * 70)\nprint(\"CELL 44 - FOLD 2 TRAINING\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# CONFIGURATION\n# ----------------------------------------------------------------\n\nSEED = 42\nFOLD = 2\n\nMAX_EPOCHS = 10\nEARLY_STOPPING_PATIENCE = 3\n\nLEARNING_RATE = 0.001\nWEIGHT_DECAY = 0.0001\nMAX_GRAD_NORM = 1.0\n\n# CPU, because previous cells confirmed CUDA is unavailable\nDEVICE = torch.device(\"cpu\")\n\nCHECKPOINT_PATH = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell44_fold2_best_model.pt\"\n)\n\nHISTORY_PATH = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell44_fold2_history.csv\"\n)\n\nMETRICS_PATH = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell44_fold2_metrics.csv\"\n)\n\n# ----------------------------------------------------------------\n# RANDOM SEEDS\n# ----------------------------------------------------------------\n\nrandom.seed(SEED)\nnp.random.seed(SEED)\ntorch.manual_seed(SEED)\n\nprint(f\"Fold: {FOLD}\")\nprint(f\"Maximum epochs: {MAX_EPOCHS}\")\nprint(f\"Early stopping patience: {EARLY_STOPPING_PATIENCE}\")\nprint(f\"Device: {DEVICE}\")\n\n# ----------------------------------------------------------------\n# REQUIRED OBJECT CHECK\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"train_loader\",\n \"val_loader\",\n \"TARGETS\",\n \"fold2_train_table\",\n \"fold2_val_table\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# VERIFY FOLD-2 DATA\n# ----------------------------------------------------------------\n\nif len(fold2_train_table) != 39:\n raise RuntimeError(\n f\"Expected 39 Fold-2 training studies, \"\n f\"found {len(fold2_train_table)}.\"\n )\n\nif len(fold2_val_table) != 19:\n raise RuntimeError(\n f\"Expected 19 Fold-2 validation studies, \"\n f\"found {len(fold2_val_table)}.\"\n )\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD DATA\")\nprint(\"=\" * 70)\n\nprint(\n f\"Training studies: {len(fold2_train_table)}\"\n)\n\nprint(\n f\"Validation studies: {len(fold2_val_table)}\"\n)\n\nprint(\n f\"Training batches: {len(train_loader)}\"\n)\n\nprint(\n f\"Validation batches: {len(val_loader)}\"\n)\n\n# ----------------------------------------------------------------\n# RESET TRAINABLE MODEL PARAMETERS\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"RESETTING MODEL FOR FOLD 2\")\nprint(\"=\" * 70)\n\n# The Cell-35 architecture has a frozen ResNet18 backbone\n# and trainable final classification layer.\n#\n# We reset ONLY parameters that were trainable before training.\n# Frozen pretrained backbone weights are preserved.\n\ntrainable_parameter_count = 0\nreset_parameter_count = 0\nreset_tensor_count = 0\n\nfor module in model.modules():\n\n # Reset standard trainable layers where a reset_parameters\n # method exists.\n if hasattr(module, \"reset_parameters\"):\n\n module_trainable = any(\n parameter.requires_grad\n for parameter in module.parameters(recurse=False)\n )\n\n if module_trainable:\n before = sum(\n parameter.numel()\n for parameter in module.parameters(recurse=False)\n if parameter.requires_grad\n )\n\n if before > 0:\n module.reset_parameters()\n\n reset_tensor_count += 1\n reset_parameter_count += before\n\n# Count trainable parameters after reset\nfor parameter in model.parameters():\n if parameter.requires_grad:\n trainable_parameter_count += parameter.numel()\n\nprint(\n f\"Trainable parameters after reset: \"\n f\"{trainable_parameter_count:,}\"\n)\n\nprint(\n f\"Parameters reset: \"\n f\"{reset_parameter_count:,}\"\n)\n\nprint(\n f\"Trainable modules reset: \"\n f\"{reset_tensor_count}\"\n)\n\nif trainable_parameter_count != 200268:\n raise RuntimeError(\n \"Unexpected trainable parameter count. \"\n f\"Expected 200,268, found \"\n f\"{trainable_parameter_count}.\"\n )\n\nif reset_parameter_count != 200268:\n raise RuntimeError(\n \"Unexpected reset parameter count. \"\n f\"Expected 200,268, found \"\n f\"{reset_parameter_count}.\"\n )\n\nprint(\"Trainable parameter count: PASS\")\nprint(\"Trainable parameter reset: PASS\")\n\n# ----------------------------------------------------------------\n# MOVE MODEL TO CPU\n# ----------------------------------------------------------------\n\nmodel = model.to(DEVICE)\n\n# ----------------------------------------------------------------\n# COMPUTE FOLD-2 CLASS WEIGHTS\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 2 CLASS WEIGHTS\")\nprint(\"=\" * 70)\n\npositive_counts = []\nnegative_counts = []\npositive_weights = []\n\nfor target in TARGETS:\n\n positives = int(\n fold2_train_table[target]\n .astype(float)\n .sum()\n )\n\n negatives = int(\n len(fold2_train_table) - positives\n )\n\n if positives <= 0:\n raise RuntimeError(\n f\"No positive samples for target: {target}\"\n )\n\n if negatives <= 0:\n raise RuntimeError(\n f\"No negative samples for target: {target}\"\n )\n\n pos_weight = (\n negatives / positives\n )\n\n positive_counts.append(positives)\n negative_counts.append(negatives)\n positive_weights.append(pos_weight)\n\n print(\n f\"{target:20s} \"\n f\"positive={positives:2d} \"\n f\"negative={negatives:2d} \"\n f\"pos_weight={pos_weight:.4f}\"\n )\n\npos_weight_tensor = torch.tensor(\n positive_weights,\n dtype=torch.float32,\n device=DEVICE\n)\n\n# ----------------------------------------------------------------\n# LOSS\n# ----------------------------------------------------------------\n\ncriterion = nn.BCEWithLogitsLoss(\n pos_weight=pos_weight_tensor\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TRAINING CONFIGURATION\")\nprint(\"=\" * 70)\n\nprint(\"Loss: BCEWithLogitsLoss\")\nprint(\"Optimizer: AdamW\")\nprint(f\"Learning rate: {LEARNING_RATE}\")\nprint(f\"Weight decay: {WEIGHT_DECAY}\")\nprint(f\"Gradient clipping: {MAX_GRAD_NORM}\")\nprint(\"Scheduler: ReduceLROnPlateau\")\n\n# ----------------------------------------------------------------\n# OPTIMIZER\n# ----------------------------------------------------------------\n\ntrainable_parameters = [\n parameter\n for parameter in model.parameters()\n if parameter.requires_grad\n]\n\noptimizer = AdamW(\n trainable_parameters,\n lr=LEARNING_RATE,\n weight_decay=WEIGHT_DECAY\n)\n\nscheduler = ReduceLROnPlateau(\n optimizer,\n mode=\"min\",\n factor=0.5,\n patience=2\n)\n\n# ----------------------------------------------------------------\n# TRAINING HISTORY\n# ----------------------------------------------------------------\n\nhistory = []\n\nbest_val_loss = float(\"inf\")\nbest_epoch = 0\n\nepochs_without_improvement = 0\n\n# ----------------------------------------------------------------\n# TRAINING LOOP\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"STARTING FOLD 2 TRAINING\")\nprint(\"=\" * 70)\n\nfor epoch in range(1, MAX_EPOCHS + 1):\n\n model.train()\n\n train_losses = []\n\n for batch in train_loader:\n\n images = batch[\"image\"].to(\n DEVICE,\n non_blocking=False\n )\n\n labels = batch[\"label\"].to(\n DEVICE,\n non_blocking=False\n )\n\n optimizer.zero_grad(\n set_to_none=True\n )\n\n logits = model(images)\n\n if logits.shape != labels.shape:\n raise RuntimeError(\n \"Logit/label shape mismatch: \"\n f\"{tuple(logits.shape)} vs \"\n f\"{tuple(labels.shape)}\"\n )\n\n loss = criterion(\n logits,\n labels\n )\n\n if not torch.isfinite(loss):\n raise RuntimeError(\n \"Non-finite training loss detected.\"\n )\n\n loss.backward()\n\n torch.nn.utils.clip_grad_norm_(\n trainable_parameters,\n MAX_GRAD_NORM\n )\n\n optimizer.step()\n\n train_losses.append(\n loss.detach().item()\n )\n\n # ------------------------------------------------------------\n # VALIDATION\n # ------------------------------------------------------------\n\n model.eval()\n\n val_losses = []\n\n all_probabilities = []\n all_labels = []\n\n with torch.no_grad():\n\n for batch in val_loader:\n\n images = batch[\"image\"].to(\n DEVICE,\n non_blocking=False\n )\n\n labels = batch[\"label\"].to(\n DEVICE,\n non_blocking=False\n )\n\n logits = model(images)\n\n loss = criterion(\n logits,\n labels\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n if not torch.isfinite(loss):\n raise RuntimeError(\n \"Non-finite validation loss detected.\"\n )\n\n if not torch.isfinite(probabilities).all():\n raise RuntimeError(\n \"Non-finite validation probabilities detected.\"\n )\n\n val_losses.append(\n loss.item()\n )\n\n all_probabilities.append(\n probabilities.cpu().numpy()\n )\n\n all_labels.append(\n labels.cpu().numpy()\n )\n\n train_loss = float(\n np.mean(train_losses)\n )\n\n val_loss = float(\n np.mean(val_losses)\n )\n\n probabilities_np = np.concatenate(\n all_probabilities,\n axis=0\n )\n\n labels_np = np.concatenate(\n all_labels,\n axis=0\n )\n\n predictions_np = (\n probabilities_np >= 0.50\n ).astype(int)\n\n target_f1 = []\n\n for target_index in range(len(TARGETS)):\n\n target_f1.append(\n f1_score(\n labels_np[:, target_index],\n predictions_np[:, target_index],\n zero_division=0\n )\n )\n\n mean_f1 = float(\n np.mean(target_f1)\n )\n\n # ------------------------------------------------------------\n # SCHEDULER\n # ------------------------------------------------------------\n\n scheduler.step(val_loss)\n\n current_lr = optimizer.param_groups[0][\"lr\"]\n\n # ------------------------------------------------------------\n # HISTORY\n # ------------------------------------------------------------\n\n history.append(\n {\n \"epoch\": epoch,\n \"train_loss\": train_loss,\n \"val_loss\": val_loss,\n \"mean_f1\": mean_f1,\n \"learning_rate\": current_lr\n }\n )\n\n # ------------------------------------------------------------\n # BEST CHECKPOINT\n # ------------------------------------------------------------\n\n if val_loss < best_val_loss:\n\n best_val_loss = val_loss\n best_epoch = epoch\n epochs_without_improvement = 0\n\n torch.save(\n {\n \"model_state_dict\": model.state_dict(),\n \"epoch\": epoch,\n \"val_loss\": val_loss,\n \"targets\": TARGETS,\n \"fold\": FOLD,\n \"seed\": SEED\n },\n CHECKPOINT_PATH\n )\n\n else:\n\n epochs_without_improvement += 1\n\n print(f\"\\nEpoch {epoch:02d}/{MAX_EPOCHS}\")\n print(f\"Train Loss: {train_loss:.5f}\")\n print(f\"Val Loss: {val_loss:.5f}\")\n print(f\"Mean F1: {mean_f1:.4f}\")\n print(f\"Learning Rate: {current_lr:.6f}\")\n print(f\"Best Epoch: {best_epoch}\")\n print(f\"Best Val Loss: {best_val_loss:.5f}\")\n\n # ------------------------------------------------------------\n # EARLY STOPPING\n # ------------------------------------------------------------\n\n if (\n epochs_without_improvement\n >= EARLY_STOPPING_PATIENCE\n ):\n\n print(\n \"\\nEarly stopping triggered \"\n f\"after {EARLY_STOPPING_PATIENCE} \"\n \"epochs without improvement.\"\n )\n\n break\n\n# ----------------------------------------------------------------\n# RESTORE BEST CHECKPOINT\n# ----------------------------------------------------------------\n\nif not os.path.exists(CHECKPOINT_PATH):\n raise RuntimeError(\n \"Best Fold-2 checkpoint was not created.\"\n )\n\ncheckpoint = torch.load(\n CHECKPOINT_PATH,\n map_location=DEVICE\n)\n\nmodel.load_state_dict(\n checkpoint[\"model_state_dict\"]\n)\n\nmodel.eval()\n\n# ----------------------------------------------------------------\n# FINAL BEST-CHECKPOINT VALIDATION\n# ----------------------------------------------------------------\n\nfinal_val_losses = []\nfinal_probabilities = []\nfinal_labels = []\n\nwith torch.no_grad():\n\n for batch in val_loader:\n\n images = batch[\"image\"].to(\n DEVICE\n )\n\n labels = batch[\"label\"].to(\n DEVICE\n )\n\n logits = model(images)\n\n loss = criterion(\n logits,\n labels\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n final_val_losses.append(\n loss.item()\n )\n\n final_probabilities.append(\n probabilities.cpu().numpy()\n )\n\n final_labels.append(\n labels.cpu().numpy()\n )\n\nfinal_val_loss = float(\n np.mean(final_val_losses)\n)\n\nfinal_probabilities = np.concatenate(\n final_probabilities,\n axis=0\n)\n\nfinal_labels = np.concatenate(\n final_labels,\n axis=0\n)\n\nfinal_predictions = (\n final_probabilities >= 0.50\n).astype(int)\n\n# ----------------------------------------------------------------\n# PER-TARGET METRICS\n# ----------------------------------------------------------------\n\nmetrics = []\n\nfor target_index, target in enumerate(TARGETS):\n\n y_true = final_labels[:, target_index]\n y_pred = final_predictions[:, target_index]\n\n tp = int(\n np.sum(\n (y_true == 1) &\n (y_pred == 1)\n )\n )\n\n tn = int(\n np.sum(\n (y_true == 0) &\n (y_pred == 0)\n )\n )\n\n fp = int(\n np.sum(\n (y_true == 0) &\n (y_pred == 1)\n )\n )\n\n fn = int(\n np.sum(\n (y_true == 1) &\n (y_pred == 0)\n )\n )\n\n actual_positive = tp + fn\n actual_negative = tn + fp\n\n sensitivity = (\n tp / actual_positive\n if actual_positive > 0\n else 0.0\n )\n\n specificity = (\n tn / actual_negative\n if actual_negative > 0\n else 0.0\n )\n\n precision = (\n tp / (tp + fp)\n if (tp + fp) > 0\n else 0.0\n )\n\n f1 = (\n 2 * precision * sensitivity\n / (precision + sensitivity)\n if (precision + sensitivity) > 0\n else 0.0\n )\n\n metrics.append(\n {\n \"target\": target,\n \"positive_count\": int(np.sum(y_true)),\n \"negative_count\": int(np.sum(y_true == 0)),\n \"predicted_positive\": int(np.sum(y_pred)),\n \"TP\": tp,\n \"TN\": tn,\n \"FP\": fp,\n \"FN\": fn,\n \"sensitivity\": sensitivity,\n \"specificity\": specificity,\n \"precision\": precision,\n \"f1\": f1,\n \"mean_probability\": float(\n np.mean(\n final_probabilities[:, target_index]\n )\n )\n }\n )\n\nmetrics_df = pd.DataFrame(metrics)\n\n# ----------------------------------------------------------------\n# SAVE RESULTS\n# ----------------------------------------------------------------\n\nhistory_df = pd.DataFrame(history)\n\nhistory_df.to_csv(\n HISTORY_PATH,\n index=False\n)\n\nmetrics_df.to_csv(\n METRICS_PATH,\n index=False\n)\n\n# ----------------------------------------------------------------\n# FINAL SUMMARY\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 2 TRAINING COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n f\"Best epoch: {best_epoch}\"\n)\n\nprint(\n f\"Best validation loss: \"\n f\"{best_val_loss:.6f}\"\n)\n\nprint(\n f\"Final validation loss: \"\n f\"{final_val_loss:.6f}\"\n)\n\nprint(\"\\nPer-target validation metrics:\")\nprint(\n metrics_df[\n [\n \"target\",\n \"positive_count\",\n \"negative_count\",\n \"predicted_positive\",\n \"sensitivity\",\n \"specificity\",\n \"precision\",\n \"f1\"\n ]\n ].to_string(index=False)\n)\n\nprint(\n f\"\\nMean F1: \"\n f\"{metrics_df['f1'].mean():.4f}\"\n)\n\nprint(\n f\"Mean sensitivity: \"\n f\"{metrics_df['sensitivity'].mean():.4f}\"\n)\n\nprint(\n f\"Mean specificity: \"\n f\"{metrics_df['specificity'].mean():.4f}\"\n)\n\nprint(\n f\"Mean precision: \"\n f\"{metrics_df['precision'].mean():.4f}\"\n)\n\nprint(\n f\"\\nCheckpoint saved: \"\n f\"{CHECKPOINT_PATH}\"\n)\n\nprint(\n f\"History saved: \"\n f\"{HISTORY_PATH}\"\n)\n\nprint(\n f\"Metrics saved: \"\n f\"{METRICS_PATH}\"\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 44 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:14:56.53014Z","iopub.execute_input":"2026-08-11T10:14:56.530522Z","iopub.status.idle":"2026-08-11T10:18:43.637763Z","shell.execute_reply.started":"2026-08-11T10:14:56.530488Z","shell.execute_reply":"2026-08-11T10:18:43.636776Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 45 - FOLD 2 PROBABILITY + THRESHOLD AUDIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom sklearn.metrics import roc_auc_score, f1_score\n\nprint(\"=\" * 70)\nprint(\"CELL 45 - FOLD 2 PROBABILITY + THRESHOLD AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"val_loader\",\n \"TARGETS\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------\n\nDEVICE = torch.device(\"cpu\")\n\n# ----------------------------------------------------------------\n# LOAD BEST FOLD-2 CHECKPOINT\n# ----------------------------------------------------------------\n\ncheckpoint_path = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell44_fold2_best_model.pt\"\n)\n\nif not os.path.exists(checkpoint_path):\n raise RuntimeError(\n \"Missing Fold-2 checkpoint: \"\n + checkpoint_path\n )\n\nprint(\"\\nLoading best Fold-2 checkpoint:\")\nprint(checkpoint_path)\n\ncheckpoint = torch.load(\n checkpoint_path,\n map_location=DEVICE\n)\n\nmodel.load_state_dict(\n checkpoint[\"model_state_dict\"]\n)\n\nmodel = model.to(DEVICE)\nmodel.eval()\n\nprint(\"Best Fold-2 checkpoint: LOADED\")\nprint(\"Model mode: evaluation\")\n\n# ----------------------------------------------------------------\n# COLLECT VALIDATION PREDICTIONS\n# ----------------------------------------------------------------\n\nall_probabilities = []\nall_labels = []\n\nwith torch.no_grad():\n\n for batch in val_loader:\n\n images = batch[\"image\"].to(DEVICE)\n labels = batch[\"label\"].to(DEVICE)\n\n logits = model(images)\n\n probabilities = torch.sigmoid(logits)\n\n if not torch.isfinite(probabilities).all():\n raise RuntimeError(\n \"NaN/Inf detected in probabilities.\"\n )\n\n all_probabilities.append(\n probabilities.cpu().numpy()\n )\n\n all_labels.append(\n labels.cpu().numpy()\n )\n\nprobabilities = np.concatenate(\n all_probabilities,\n axis=0\n)\n\nlabels = np.concatenate(\n all_labels,\n axis=0\n)\n\nprint(\n f\"\\nProbability shape: {probabilities.shape}\"\n)\n\nprint(\n f\"Label shape: {labels.shape}\"\n)\n\nif probabilities.shape != labels.shape:\n raise RuntimeError(\n \"Probability and label shapes do not match.\"\n )\n\nif probabilities.shape[1] != len(TARGETS):\n raise RuntimeError(\n \"Unexpected number of target columns.\"\n )\n\nprint(\"Probability validity: PASS\")\nprint(\"Shape consistency: PASS\")\n\n# ----------------------------------------------------------------\n# PROBABILITY SEPARATION + ROC-AUC\n# ----------------------------------------------------------------\n\nseparation_rows = []\n\nfor index, target in enumerate(TARGETS):\n\n y_true = labels[:, index]\n y_prob = probabilities[:, index]\n\n positive_probs = y_prob[y_true == 1]\n negative_probs = y_prob[y_true == 0]\n\n positive_mean = float(\n np.mean(positive_probs)\n )\n\n negative_mean = float(\n np.mean(negative_probs)\n )\n\n positive_median = float(\n np.median(positive_probs)\n )\n\n negative_median = float(\n np.median(negative_probs)\n )\n\n separation = (\n positive_mean -\n negative_mean\n )\n\n if (\n len(np.unique(y_true)) == 2\n ):\n auc = float(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n else:\n auc = np.nan\n\n separation_rows.append(\n {\n \"target\": target,\n \"positive_count\": int(np.sum(y_true == 1)),\n \"negative_count\": int(np.sum(y_true == 0)),\n \"positive_mean\": positive_mean,\n \"negative_mean\": negative_mean,\n \"positive_median\": positive_median,\n \"negative_median\": negative_median,\n \"probability_separation\": separation,\n \"roc_auc\": auc\n }\n )\n\nseparation_df = pd.DataFrame(\n separation_rows\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PROBABILITY SEPARATION\")\nprint(\"=\" * 70)\n\nprint(\n separation_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# THRESHOLD AUDIT\n# ----------------------------------------------------------------\n\nthresholds = np.arange(\n 0.20,\n 0.651,\n 0.025\n)\n\nthreshold_rows = []\n\nfor index, target in enumerate(TARGETS):\n\n y_true = labels[:, index]\n y_prob = probabilities[:, index]\n\n f1_at_050 = f1_score(\n y_true,\n (y_prob >= 0.50).astype(int),\n zero_division=0\n )\n\n best_threshold = 0.50\n best_f1 = f1_at_050\n\n for threshold in thresholds:\n\n y_pred = (\n y_prob >= threshold\n ).astype(int)\n\n current_f1 = f1_score(\n y_true,\n y_pred,\n zero_division=0\n )\n\n if current_f1 > best_f1:\n\n best_f1 = current_f1\n best_threshold = float(\n threshold\n )\n\n threshold_rows.append(\n {\n \"target\": target,\n \"f1_at_0.50\": float(f1_at_050),\n \"best_threshold\": float(best_threshold),\n \"best_f1\": float(best_f1),\n \"f1_improvement\": float(\n best_f1 - f1_at_050\n )\n }\n )\n\nthreshold_df = pd.DataFrame(\n threshold_rows\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"THRESHOLD AUDIT\")\nprint(\"=\" * 70)\n\nprint(\n threshold_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# SUMMARY\n# ----------------------------------------------------------------\n\nmean_auc = float(\n separation_df[\"roc_auc\"].mean()\n)\n\nmean_separation = float(\n separation_df[\n \"probability_separation\"\n ].mean()\n)\n\nmean_f1_050 = float(\n threshold_df[\"f1_at_0.50\"].mean()\n)\n\nmean_best_f1 = float(\n threshold_df[\"best_f1\"].mean()\n)\n\nf1_improvement = (\n mean_best_f1 -\n mean_f1_050\n)\n\nauc_good_targets = separation_df[\n separation_df[\"roc_auc\"] >= 0.65\n][\"target\"].tolist()\n\nnegative_separation_targets = separation_df[\n separation_df[\"probability_separation\"] < 0\n][\"target\"].tolist()\n\nimprovement_targets = threshold_df[\n threshold_df[\"f1_improvement\"] > 0.05\n][\"target\"].tolist()\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Mean ROC-AUC: \"\n f\"{mean_auc:.4f}\"\n)\n\nprint(\n f\"Mean probability separation: \"\n f\"{mean_separation:.4f}\"\n)\n\nprint(\n f\"Mean F1 @ 0.50: \"\n f\"{mean_f1_050:.4f}\"\n)\n\nprint(\n f\"Mean best-threshold F1: \"\n f\"{mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Potential F1 improvement: \"\n f\"{f1_improvement:.4f}\"\n)\n\nprint(\n f\"\\nTargets with ROC-AUC >= 0.65: \"\n f\"{len(auc_good_targets)}\"\n)\n\nprint(\n auc_good_targets\n)\n\nprint(\n f\"\\nTargets with negative probability separation: \"\n f\"{len(negative_separation_targets)}\"\n)\n\nprint(\n negative_separation_targets\n)\n\nprint(\n f\"\\nTargets with >0.05 F1 improvement: \"\n f\"{len(improvement_targets)}\"\n)\n\nprint(\n improvement_targets\n)\n\n# ----------------------------------------------------------------\n# SAVE\n# ----------------------------------------------------------------\n\nseparation_path = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell45_fold2_probability_separation.csv\"\n)\n\nthreshold_path = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell45_fold2_threshold_audit.csv\"\n)\n\nseparation_df.to_csv(\n separation_path,\n index=False\n)\n\nthreshold_df.to_csv(\n threshold_path,\n index=False\n)\n\nprint(\"\\nSaved:\")\nprint(separation_path)\nprint(threshold_path)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 45 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:19:38.270991Z","iopub.execute_input":"2026-08-11T10:19:38.271353Z","iopub.status.idle":"2026-08-11T10:19:47.982643Z","shell.execute_reply.started":"2026-08-11T10:19:38.271322Z","shell.execute_reply":"2026-08-11T10:19:47.981618Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 46 - 3-FOLD OUT-OF-FOLD PREDICTION CONSOLIDATION\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom sklearn.metrics import roc_auc_score, f1_score\n\nprint(\"=\" * 70)\nprint(\"CELL 46 - 3-FOLD OUT-OF-FOLD PREDICTION CONSOLIDATION\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# CONSTANTS\n# ----------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nMODELING_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_modeling_studies.csv\"\n)\n\nFOLD_PATH = os.path.join(\n AUDIT_DIR,\n \"cell29_study_folds.csv\"\n)\n\nCHECKPOINTS = {\n 0: os.path.join(\n AUDIT_DIR,\n \"cell37_fold0_best_model.pt\"\n ),\n 1: os.path.join(\n AUDIT_DIR,\n \"cell40_fold1_best_model.pt\"\n ),\n 2: os.path.join(\n AUDIT_DIR,\n \"cell44_fold2_best_model.pt\"\n )\n}\n\nTARGETS_EXPECTED = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n]\n\nDEVICE = torch.device(\"cpu\")\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"KneeStudyDataset\",\n \"model\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# LOAD MODELING TABLE AND FOLD TABLE FROM SAVED FILES\n# ----------------------------------------------------------------\n\nif not os.path.exists(MODELING_PATH):\n raise RuntimeError(\n \"Missing modeling table: \"\n + MODELING_PATH\n )\n\nif not os.path.exists(FOLD_PATH):\n raise RuntimeError(\n \"Missing fold table: \"\n + FOLD_PATH\n )\n\nmodeling_table = pd.read_csv(\n MODELING_PATH\n)\n\nstudy_folds = pd.read_csv(\n FOLD_PATH\n)\n\nprint(\"\\nModeling table:\")\nprint(modeling_table.shape)\n\nprint(\"Fold table:\")\nprint(study_folds.shape)\n\n# ----------------------------------------------------------------\n# BASIC TABLE VALIDATION\n# ----------------------------------------------------------------\n\nif \"StudyInstanceUID\" not in modeling_table.columns:\n raise RuntimeError(\n \"StudyInstanceUID missing from modeling table.\"\n )\n\nif \"StudyInstanceUID\" not in study_folds.columns:\n raise RuntimeError(\n \"StudyInstanceUID missing from fold table.\"\n )\n\nif \"fold\" not in study_folds.columns:\n raise RuntimeError(\n \"fold column missing from fold table.\"\n )\n\nmissing_targets = [\n target\n for target in TARGETS_EXPECTED\n if target not in modeling_table.columns\n]\n\nif missing_targets:\n raise RuntimeError(\n \"Missing target columns: \"\n + \", \".join(missing_targets)\n )\n\n# ----------------------------------------------------------------\n# PRIMARY SERIES COLUMN VALIDATION\n# ----------------------------------------------------------------\n\nprimary_columns = [\n \"Sagittal_SeriesInstanceUID\",\n \"Coronal_SeriesInstanceUID\",\n \"Axial_SeriesInstanceUID\"\n]\n\nmissing_primary_columns = [\n column\n for column in primary_columns\n if column not in modeling_table.columns\n]\n\nif missing_primary_columns:\n raise RuntimeError(\n \"Missing primary-series columns: \"\n + \", \".join(missing_primary_columns)\n )\n\nprint(\"Target columns: PASS\")\nprint(\"Primary-series columns: PASS\")\n\n# ----------------------------------------------------------------\n# MERGE FOLD ASSIGNMENTS\n# ----------------------------------------------------------------\n\nfold_columns = [\n \"StudyInstanceUID\",\n \"fold\"\n]\n\nfold_assignment = study_folds[\n fold_columns\n].copy()\n\nif len(\n fold_assignment[\"StudyInstanceUID\"].unique()\n) != len(fold_assignment):\n\n raise RuntimeError(\n \"Duplicate StudyInstanceUID values \"\n \"found in fold table.\"\n )\n\nmodeling_table = modeling_table.drop(\n columns=[\"fold\"],\n errors=\"ignore\"\n)\n\nmodeling_table = modeling_table.merge(\n fold_assignment,\n on=\"StudyInstanceUID\",\n how=\"left\",\n validate=\"one_to_one\"\n)\n\nif modeling_table[\"fold\"].isna().any():\n raise RuntimeError(\n \"Some modeling studies do not have \"\n \"a fold assignment.\"\n )\n\nprint(\"Fold assignment merge: PASS\")\n\n# ----------------------------------------------------------------\n# GLOBAL STUDY VALIDATION\n# ----------------------------------------------------------------\n\nif len(modeling_table) != 58:\n raise RuntimeError(\n f\"Expected 58 modeling studies, \"\n f\"found {len(modeling_table)}.\"\n )\n\nif (\n modeling_table[\"StudyInstanceUID\"]\n .nunique()\n != 58\n):\n raise RuntimeError(\n \"Modeling table does not contain \"\n \"58 unique studies.\"\n )\n\navailable_folds = sorted(\n modeling_table[\"fold\"].astype(int).unique()\n)\n\nif available_folds != [0, 1, 2]:\n raise RuntimeError(\n \"Expected folds [0, 1, 2], \"\n f\"found {available_folds}\"\n )\n\nprint(\"Total modeling studies: 58\")\nprint(\"Unique studies: 58\")\nprint(\"Available folds: [0, 1, 2]\")\nprint(\"Three-fold configuration: PASS\")\n\n# ----------------------------------------------------------------\n# CHECK CHECKPOINTS\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CHECKPOINT VALIDATION\")\nprint(\"=\" * 70)\n\nfor fold, path in CHECKPOINTS.items():\n\n if not os.path.exists(path):\n raise RuntimeError(\n f\"Missing Fold-{fold} checkpoint: {path}\"\n )\n\n print(\n f\"Fold {fold} checkpoint: EXISTS\"\n )\n\nprint(\"All three checkpoints: PASS\")\n\n# ----------------------------------------------------------------\n# FUNCTION TO CREATE VALIDATION LOADER\n# ----------------------------------------------------------------\n\ndef create_validation_loader(\n fold_number,\n dataframe\n):\n\n validation_table = dataframe[\n dataframe[\"fold\"].astype(int)\n == fold_number\n ].copy()\n\n if validation_table.empty:\n raise RuntimeError(\n f\"Fold {fold_number} validation table is empty.\"\n )\n\n dataset = KneeStudyDataset(\n validation_table,\n targets=TARGETS_EXPECTED\n )\n\n loader = torch.utils.data.DataLoader(\n dataset,\n batch_size=2,\n shuffle=False,\n num_workers=0,\n pin_memory=False\n )\n\n return validation_table, dataset, loader\n\n# ----------------------------------------------------------------\n# OOF STORAGE\n# ----------------------------------------------------------------\n\noof_rows = []\n\n# ----------------------------------------------------------------\n# PROCESS EACH FOLD\n# ----------------------------------------------------------------\n\nfor fold_number in [0, 1, 2]:\n\n print(\"\\n\" + \"=\" * 70)\n print(\n f\"PROCESSING FOLD {fold_number} VALIDATION\"\n )\n print(\"=\" * 70)\n\n validation_table, validation_dataset, validation_loader = (\n create_validation_loader(\n fold_number,\n modeling_table\n )\n )\n\n expected_studies = (\n validation_table[\n \"StudyInstanceUID\"\n ]\n .astype(str)\n .tolist()\n )\n\n print(\n f\"Validation studies: \"\n f\"{len(expected_studies)}\"\n )\n\n print(\n f\"Validation dataset length: \"\n f\"{len(validation_dataset)}\"\n )\n\n # ------------------------------------------------------------\n # LOAD CHECKPOINT\n # ------------------------------------------------------------\n\n checkpoint_path = CHECKPOINTS[\n fold_number\n ]\n\n checkpoint = torch.load(\n checkpoint_path,\n map_location=DEVICE\n )\n\n model.load_state_dict(\n checkpoint[\"model_state_dict\"]\n )\n\n model = model.to(DEVICE)\n model.eval()\n\n print(\n f\"Fold-{fold_number} checkpoint: LOADED\"\n )\n\n # ------------------------------------------------------------\n # PREDICTION\n # ------------------------------------------------------------\n\n fold_study_ids = []\n\n with torch.no_grad():\n\n for batch in validation_loader:\n\n images = batch[\"image\"].to(DEVICE)\n\n labels = batch[\"label\"].cpu().numpy()\n\n study_ids = [\n str(study_id)\n for study_id in batch[\"study_id\"]\n ]\n\n logits = model(images)\n\n probabilities = torch.sigmoid(\n logits\n ).cpu().numpy()\n\n if not np.isfinite(\n probabilities\n ).all():\n\n raise RuntimeError(\n f\"NaN/Inf detected in \"\n f\"Fold {fold_number} predictions.\"\n )\n\n if probabilities.shape[1] != 12:\n raise RuntimeError(\n f\"Unexpected output shape \"\n f\"in Fold {fold_number}: \"\n f\"{probabilities.shape}\"\n )\n\n for row_index, study_id in enumerate(\n study_ids\n ):\n\n fold_study_ids.append(\n study_id\n )\n\n row = {\n \"StudyInstanceUID\": study_id,\n \"fold\": fold_number\n }\n\n for target_index, target in enumerate(\n TARGETS_EXPECTED\n ):\n\n row[\n f\"{target}_label\"\n ] = float(\n labels[\n row_index,\n target_index\n ]\n )\n\n row[\n f\"{target}_prob\"\n ] = float(\n probabilities[\n row_index,\n target_index\n ]\n )\n\n oof_rows.append(row)\n\n # ------------------------------------------------------------\n # FOLD COVERAGE VALIDATION\n # ------------------------------------------------------------\n\n if len(fold_study_ids) != len(\n expected_studies\n ):\n\n raise RuntimeError(\n f\"Fold {fold_number}: expected \"\n f\"{len(expected_studies)} predictions, \"\n f\"got {len(fold_study_ids)}.\"\n )\n\n if len(set(fold_study_ids)) != len(\n fold_study_ids\n ):\n\n raise RuntimeError(\n f\"Fold {fold_number}: duplicate \"\n f\"study predictions detected.\"\n )\n\n if set(fold_study_ids) != set(\n expected_studies\n ):\n\n raise RuntimeError(\n f\"Fold {fold_number}: predicted study \"\n f\"IDs do not match validation studies.\"\n )\n\n print(\n f\"Fold {fold_number} predictions: \"\n f\"{len(fold_study_ids)}\"\n )\n\n print(\n f\"Fold {fold_number} coverage: PASS\"\n )\n\n# ----------------------------------------------------------------\n# BUILD OOF DATAFRAME\n# ----------------------------------------------------------------\n\noof_df = pd.DataFrame(\n oof_rows\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OOF DATASET VALIDATION\")\nprint(\"=\" * 70)\n\nprint(\n \"OOF shape:\",\n oof_df.shape\n)\n\nif len(oof_df) != 58:\n raise RuntimeError(\n f\"Expected 58 OOF rows, \"\n f\"found {len(oof_df)}.\"\n )\n\nif (\n oof_df[\"StudyInstanceUID\"]\n .nunique()\n != 58\n):\n\n raise RuntimeError(\n \"OOF predictions do not contain \"\n \"exactly 58 unique studies.\"\n )\n\nif (\n oof_df.groupby(\n \"StudyInstanceUID\"\n ).size() != 1\n).any():\n\n raise RuntimeError(\n \"Some studies appear more than once \"\n \"in the OOF prediction table.\"\n )\n\nprint(\"OOF rows: 58\")\nprint(\"Unique OOF studies: 58\")\nprint(\"One prediction per study: PASS\")\n\n# ----------------------------------------------------------------\n# FOLD COVERAGE\n# ----------------------------------------------------------------\n\nprint(\"\\nOOF fold distribution:\")\n\nprint(\n oof_df[\"fold\"]\n .value_counts()\n .sort_index()\n)\n\nexpected_fold_sizes = {\n 0: 20,\n 1: 19,\n 2: 19\n}\n\nfor fold_number, expected_size in (\n expected_fold_sizes.items()\n):\n\n actual_size = int(\n (oof_df[\"fold\"] == fold_number).sum()\n )\n\n if actual_size != expected_size:\n\n raise RuntimeError(\n f\"Fold {fold_number}: expected \"\n f\"{expected_size} OOF rows, \"\n f\"found {actual_size}.\"\n )\n\nprint(\"Fold coverage: PASS\")\n\n# ----------------------------------------------------------------\n# CALCULATE OOF METRICS\n# ----------------------------------------------------------------\n\nmetric_rows = []\n\nfor target in TARGETS_EXPECTED:\n\n y_true = oof_df[\n f\"{target}_label\"\n ].values\n\n y_prob = oof_df[\n f\"{target}_prob\"\n ].values\n\n y_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n auc = roc_auc_score(\n y_true,\n y_prob\n )\n\n f1 = f1_score(\n y_true,\n y_pred,\n zero_division=0\n )\n\n metric_rows.append(\n {\n \"target\": target,\n \"positive_count\": int(\n y_true.sum()\n ),\n \"negative_count\": int(\n len(y_true) - y_true.sum()\n ),\n \"roc_auc\": float(auc),\n \"f1_at_0.50\": float(f1)\n }\n )\n\noof_metrics = pd.DataFrame(\n metric_rows\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OOF BASELINE METRICS\")\nprint(\"=\" * 70)\n\nprint(\n oof_metrics.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\nmean_oof_auc = float(\n oof_metrics[\"roc_auc\"].mean()\n)\n\nmean_oof_f1 = float(\n oof_metrics[\"f1_at_0.50\"].mean()\n)\n\nprint(\n f\"\\nMean OOF ROC-AUC: \"\n f\"{mean_oof_auc:.4f}\"\n)\n\nprint(\n f\"Mean OOF F1 @ 0.50: \"\n f\"{mean_oof_f1:.4f}\"\n)\n\n# ----------------------------------------------------------------\n# SAVE OOF RESULTS\n# ----------------------------------------------------------------\n\noof_path = os.path.join(\n AUDIT_DIR,\n \"cell46_oof_predictions.csv\"\n)\n\nmetrics_path = os.path.join(\n AUDIT_DIR,\n \"cell46_oof_metrics.csv\"\n)\n\noof_df.to_csv(\n oof_path,\n index=False\n)\n\noof_metrics.to_csv(\n metrics_path,\n index=False\n)\n\nprint(\"\\nSaved:\")\nprint(oof_path)\nprint(metrics_path)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 46 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"All 58 labeled studies have exactly one \"\n \"out-of-fold prediction.\"\n)\n\nprint(\n \"No training performed in this cell.\"\n)\n\nprint(\n \"No dataset pipeline modified.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:21:33.969385Z","iopub.execute_input":"2026-08-11T10:21:33.969772Z","iopub.status.idle":"2026-08-11T10:22:04.344042Z","shell.execute_reply.started":"2026-08-11T10:21:33.969741Z","shell.execute_reply":"2026-08-11T10:22:04.342912Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 47 - OOF THRESHOLD + CALIBRATION AUDIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nfrom sklearn.metrics import (\n roc_auc_score,\n average_precision_score,\n f1_score,\n precision_score,\n recall_score,\n confusion_matrix\n)\n\nprint(\"=\" * 70)\nprint(\"CELL 47 - OOF THRESHOLD + CALIBRATION AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# CONSTANTS\n# ----------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nOOF_PATH = os.path.join(\n AUDIT_DIR,\n \"cell46_oof_predictions.csv\"\n)\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n]\n\n# ----------------------------------------------------------------\n# LOAD OOF PREDICTIONS\n# ----------------------------------------------------------------\n\nprint(\"\\nLoading OOF predictions...\")\n\nif not os.path.exists(OOF_PATH):\n raise RuntimeError(\n \"Missing OOF prediction file: \"\n + OOF_PATH\n )\n\noof_df = pd.read_csv(OOF_PATH)\n\nprint(\n \"OOF table shape:\",\n oof_df.shape\n)\n\n# ----------------------------------------------------------------\n# BASIC VALIDATION\n# ----------------------------------------------------------------\n\nrequired_columns = [\n \"StudyInstanceUID\",\n \"fold\"\n]\n\nfor target in TARGETS:\n required_columns.append(\n f\"{target}_label\"\n )\n required_columns.append(\n f\"{target}_prob\"\n )\n\nmissing_columns = [\n column\n for column in required_columns\n if column not in oof_df.columns\n]\n\nif missing_columns:\n raise RuntimeError(\n \"Missing OOF columns: \"\n + \", \".join(missing_columns)\n )\n\nif len(oof_df) != 58:\n raise RuntimeError(\n f\"Expected 58 OOF rows, found {len(oof_df)}.\"\n )\n\nif (\n oof_df[\"StudyInstanceUID\"].nunique()\n != 58\n):\n raise RuntimeError(\n \"OOF table does not contain exactly \"\n \"58 unique studies.\"\n )\n\nif (\n oof_df.groupby(\"StudyInstanceUID\")\n .size()\n .max()\n != 1\n):\n raise RuntimeError(\n \"Duplicate OOF study predictions detected.\"\n )\n\nprint(\"OOF row count: PASS\")\nprint(\"Unique study count: PASS\")\nprint(\"One prediction per study: PASS\")\n\n# ----------------------------------------------------------------\n# FOLD VALIDATION\n# ----------------------------------------------------------------\n\nfold_counts = (\n oof_df[\"fold\"]\n .value_counts()\n .sort_index()\n)\n\nprint(\"\\nOOF fold distribution:\")\nprint(fold_counts)\n\nexpected_fold_counts = {\n 0: 20,\n 1: 19,\n 2: 19\n}\n\nfor fold, expected_count in expected_fold_counts.items():\n\n actual_count = int(\n fold_counts.get(fold, 0)\n )\n\n if actual_count != expected_count:\n raise RuntimeError(\n f\"Fold {fold}: expected \"\n f\"{expected_count}, found {actual_count}.\"\n )\n\nprint(\"OOF fold coverage: PASS\")\n\n# ----------------------------------------------------------------\n# PROBABILITY VALIDATION\n# ----------------------------------------------------------------\n\nfor target in TARGETS:\n\n probabilities = oof_df[\n f\"{target}_prob\"\n ].values\n\n labels = oof_df[\n f\"{target}_label\"\n ].values\n\n if not np.isfinite(probabilities).all():\n raise RuntimeError(\n f\"NaN/Inf probabilities detected: {target}\"\n )\n\n if not np.isfinite(labels).all():\n raise RuntimeError(\n f\"NaN/Inf labels detected: {target}\"\n )\n\n if (\n probabilities.min() < 0.0\n or probabilities.max() > 1.0\n ):\n raise RuntimeError(\n f\"Invalid probability range: {target}\"\n )\n\n unique_labels = set(\n np.unique(labels).tolist()\n )\n\n if not unique_labels.issubset({0.0, 1.0}):\n raise RuntimeError(\n f\"Invalid labels for target: {target}\"\n )\n\nprint(\"Probability validity: PASS\")\nprint(\"Label validity: PASS\")\n\n# ----------------------------------------------------------------\n# THRESHOLD GRID\n# ----------------------------------------------------------------\n\nthresholds = np.round(\n np.arange(\n 0.10,\n 0.901,\n 0.025\n ),\n 3\n)\n\nprint(\n \"\\nThreshold grid:\",\n f\"{thresholds[0]:.3f}\",\n \"to\",\n f\"{thresholds[-1]:.3f}\",\n f\"({len(thresholds)} thresholds)\"\n)\n\n# ----------------------------------------------------------------\n# TARGET-WISE OOF ANALYSIS\n# ----------------------------------------------------------------\n\nsummary_rows = []\nthreshold_rows = []\n\nfor target in TARGETS:\n\n y_true = oof_df[\n f\"{target}_label\"\n ].astype(int).values\n\n y_prob = oof_df[\n f\"{target}_prob\"\n ].astype(float).values\n\n positive_count = int(\n y_true.sum()\n )\n\n negative_count = int(\n len(y_true) - positive_count\n )\n\n # ------------------------------------------------------------\n # ROC-AUC\n # ------------------------------------------------------------\n\n if (\n len(np.unique(y_true))\n == 2\n ):\n roc_auc = float(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n\n pr_auc = float(\n average_precision_score(\n y_true,\n y_prob\n )\n )\n\n else:\n roc_auc = np.nan\n pr_auc = np.nan\n\n # ------------------------------------------------------------\n # BASELINE THRESHOLD\n # ------------------------------------------------------------\n\n baseline_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n baseline_f1 = float(\n f1_score(\n y_true,\n baseline_pred,\n zero_division=0\n )\n )\n\n baseline_precision = float(\n precision_score(\n y_true,\n baseline_pred,\n zero_division=0\n )\n )\n\n baseline_recall = float(\n recall_score(\n y_true,\n baseline_pred,\n zero_division=0\n )\n )\n\n # ------------------------------------------------------------\n # BEST F1 THRESHOLD\n # ------------------------------------------------------------\n\n best_threshold = 0.50\n best_f1 = -1.0\n best_precision = 0.0\n best_recall = 0.0\n\n for threshold in thresholds:\n\n prediction = (\n y_prob >= threshold\n ).astype(int)\n\n current_f1 = float(\n f1_score(\n y_true,\n prediction,\n zero_division=0\n )\n )\n\n current_precision = float(\n precision_score(\n y_true,\n prediction,\n zero_division=0\n )\n )\n\n current_recall = float(\n recall_score(\n y_true,\n prediction,\n zero_division=0\n )\n )\n\n threshold_rows.append(\n {\n \"target\": target,\n \"threshold\": float(threshold),\n \"f1\": current_f1,\n \"precision\": current_precision,\n \"recall\": current_recall,\n \"predicted_positive\": int(\n prediction.sum()\n )\n }\n )\n\n if current_f1 > best_f1:\n\n best_f1 = current_f1\n best_threshold = float(threshold)\n best_precision = current_precision\n best_recall = current_recall\n\n # ------------------------------------------------------------\n # PROBABILITY DISTRIBUTION\n # ------------------------------------------------------------\n\n positive_probabilities = y_prob[\n y_true == 1\n ]\n\n negative_probabilities = y_prob[\n y_true == 0\n ]\n\n positive_mean = float(\n positive_probabilities.mean()\n )\n\n negative_mean = float(\n negative_probabilities.mean()\n )\n\n probability_separation = (\n positive_mean\n - negative_mean\n )\n\n # ------------------------------------------------------------\n # BRIER SCORE\n # ------------------------------------------------------------\n\n brier_score = float(\n np.mean(\n (\n y_prob\n - y_true\n ) ** 2\n )\n )\n\n # ------------------------------------------------------------\n # SUMMARY\n # ------------------------------------------------------------\n\n summary_rows.append(\n {\n \"target\": target,\n \"positive_count\": positive_count,\n \"negative_count\": negative_count,\n \"roc_auc\": roc_auc,\n \"pr_auc\": pr_auc,\n \"brier_score\": brier_score,\n \"f1_at_0.50\": baseline_f1,\n \"best_threshold\": best_threshold,\n \"best_f1\": best_f1,\n \"f1_improvement\": (\n best_f1\n - baseline_f1\n ),\n \"best_precision\": best_precision,\n \"best_recall\": best_recall,\n \"positive_mean_probability\": positive_mean,\n \"negative_mean_probability\": negative_mean,\n \"probability_separation\": probability_separation\n }\n )\n\n# ----------------------------------------------------------------\n# DATAFRAMES\n# ----------------------------------------------------------------\n\noof_threshold_summary = pd.DataFrame(\n summary_rows\n)\n\noof_threshold_grid = pd.DataFrame(\n threshold_rows\n)\n\n# ----------------------------------------------------------------\n# PRINT TARGET SUMMARY\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OOF TARGET THRESHOLD SUMMARY\")\nprint(\"=\" * 70)\n\ndisplay_columns = [\n \"target\",\n \"positive_count\",\n \"negative_count\",\n \"roc_auc\",\n \"pr_auc\",\n \"brier_score\",\n \"f1_at_0.50\",\n \"best_threshold\",\n \"best_f1\",\n \"f1_improvement\",\n \"probability_separation\"\n]\n\nprint(\n oof_threshold_summary[\n display_columns\n ].to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# OVERALL SUMMARY\n# ----------------------------------------------------------------\n\nmean_roc_auc = float(\n oof_threshold_summary[\n \"roc_auc\"\n ].mean()\n)\n\nmean_pr_auc = float(\n oof_threshold_summary[\n \"pr_auc\"\n ].mean()\n)\n\nmean_brier = float(\n oof_threshold_summary[\n \"brier_score\"\n ].mean()\n)\n\nmean_f1_050 = float(\n oof_threshold_summary[\n \"f1_at_0.50\"\n ].mean()\n)\n\nmean_best_f1 = float(\n oof_threshold_summary[\n \"best_f1\"\n ].mean()\n)\n\nmean_f1_improvement = float(\n oof_threshold_summary[\n \"f1_improvement\"\n ].mean()\n)\n\nmean_separation = float(\n oof_threshold_summary[\n \"probability_separation\"\n ].mean()\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OVERALL OOF THRESHOLD SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Mean ROC-AUC: {mean_roc_auc:.4f}\"\n)\n\nprint(\n f\"Mean PR-AUC: {mean_pr_auc:.4f}\"\n)\n\nprint(\n f\"Mean Brier score: {mean_brier:.4f}\"\n)\n\nprint(\n f\"Mean F1 @ 0.50: {mean_f1_050:.4f}\"\n)\n\nprint(\n f\"Mean best-threshold F1: {mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Potential F1 improvement: {mean_f1_improvement:.4f}\"\n)\n\nprint(\n f\"Mean probability separation: {mean_separation:.4f}\"\n)\n\n# ----------------------------------------------------------------\n# IDENTIFY STRONG / WEAK TARGETS\n# ----------------------------------------------------------------\n\nauc_strong = oof_threshold_summary[\n oof_threshold_summary[\"roc_auc\"] >= 0.65\n][\n \"target\"\n].tolist()\n\nauc_weak = oof_threshold_summary[\n oof_threshold_summary[\"roc_auc\"] < 0.50\n][\n \"target\"\n].tolist()\n\nnegative_separation = oof_threshold_summary[\n oof_threshold_summary[\n \"probability_separation\"\n ] < 0\n][\n \"target\"\n].tolist()\n\nlarge_threshold_gain = oof_threshold_summary[\n oof_threshold_summary[\n \"f1_improvement\"\n ] >= 0.10\n][\n \"target\"\n].tolist()\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OOF TARGET FLAGS\")\nprint(\"=\" * 70)\n\nprint(\n \"Targets with ROC-AUC >= 0.65:\",\n len(auc_strong)\n)\n\nprint(\n auc_strong\n)\n\nprint(\n \"\\nTargets with ROC-AUC < 0.50:\",\n len(auc_weak)\n)\n\nprint(\n auc_weak\n)\n\nprint(\n \"\\nTargets with negative probability separation:\",\n len(negative_separation)\n)\n\nprint(\n negative_separation\n)\n\nprint(\n \"\\nTargets with >= 0.10 F1 improvement:\",\n len(large_threshold_gain)\n)\n\nprint(\n large_threshold_gain\n)\n\n# ----------------------------------------------------------------\n# OOF CONFUSION MATRICES AT 0.50\n# ----------------------------------------------------------------\n\nconfusion_rows = []\n\nfor target in TARGETS:\n\n y_true = oof_df[\n f\"{target}_label\"\n ].astype(int).values\n\n y_prob = oof_df[\n f\"{target}_prob\"\n ].astype(float).values\n\n y_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n tn, fp, fn, tp = confusion_matrix(\n y_true,\n y_pred,\n labels=[0, 1]\n ).ravel()\n\n sensitivity = (\n tp / (tp + fn)\n if (tp + fn) > 0\n else 0.0\n )\n\n specificity = (\n tn / (tn + fp)\n if (tn + fp) > 0\n else 0.0\n )\n\n confusion_rows.append(\n {\n \"target\": target,\n \"TP\": int(tp),\n \"TN\": int(tn),\n \"FP\": int(fp),\n \"FN\": int(fn),\n \"sensitivity\": sensitivity,\n \"specificity\": specificity\n }\n )\n\noof_confusion = pd.DataFrame(\n confusion_rows\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OOF CONFUSION MATRICES @ 0.50\")\nprint(\"=\" * 70)\n\nprint(\n oof_confusion.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# SAVE RESULTS\n# ----------------------------------------------------------------\n\nsummary_path = os.path.join(\n AUDIT_DIR,\n \"cell47_oof_threshold_summary.csv\"\n)\n\ngrid_path = os.path.join(\n AUDIT_DIR,\n \"cell47_oof_threshold_grid.csv\"\n)\n\nconfusion_path = os.path.join(\n AUDIT_DIR,\n \"cell47_oof_confusion.csv\"\n)\n\noof_threshold_summary.to_csv(\n summary_path,\n index=False\n)\n\noof_threshold_grid.to_csv(\n grid_path,\n index=False\n)\n\noof_confusion.to_csv(\n confusion_path,\n index=False\n)\n\n# ----------------------------------------------------------------\n# FINAL CHECKS\n# ----------------------------------------------------------------\n\nif len(oof_threshold_summary) != 12:\n raise RuntimeError(\n \"Expected 12 target summary rows.\"\n )\n\nif len(oof_threshold_grid) != (\n 12 * len(thresholds)\n):\n raise RuntimeError(\n \"Unexpected threshold-grid size.\"\n )\n\nif len(oof_confusion) != 12:\n raise RuntimeError(\n \"Expected 12 confusion-matrix rows.\"\n )\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 47 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"OOF threshold analysis: PASS\")\nprint(\"OOF calibration/error statistics: PASS\")\nprint(\"No training performed.\")\nprint(\"No model modified.\")\nprint(\"No test data used.\")\n\nprint(\"\\nSaved:\")\nprint(summary_path)\nprint(grid_path)\nprint(confusion_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:27:05.428572Z","iopub.execute_input":"2026-08-11T10:27:05.428947Z","iopub.status.idle":"2026-08-11T10:27:07.631098Z","shell.execute_reply.started":"2026-08-11T10:27:05.428905Z","shell.execute_reply":"2026-08-11T10:27:07.629968Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 48 - OOF ERROR + FOLD GENERALIZATION AUDIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nfrom sklearn.metrics import (\n roc_auc_score,\n f1_score,\n precision_score,\n recall_score\n)\n\nprint(\"=\" * 70)\nprint(\"CELL 48 - OOF ERROR + FOLD GENERALIZATION AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# CONSTANTS\n# ----------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nOOF_PATH = os.path.join(\n AUDIT_DIR,\n \"cell46_oof_predictions.csv\"\n)\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\"\n]\n\n# ----------------------------------------------------------------\n# LOAD OOF DATA\n# ----------------------------------------------------------------\n\nprint(\"\\nLoading OOF predictions...\")\n\nif not os.path.exists(OOF_PATH):\n raise RuntimeError(\n \"Missing OOF prediction file: \"\n + OOF_PATH\n )\n\noof_df = pd.read_csv(\n OOF_PATH\n)\n\nprint(\n \"OOF shape:\",\n oof_df.shape\n)\n\n# ----------------------------------------------------------------\n# VALIDATE OOF DATA\n# ----------------------------------------------------------------\n\nrequired_columns = [\n \"StudyInstanceUID\",\n \"fold\"\n]\n\nfor target in TARGETS:\n required_columns.extend([\n f\"{target}_label\",\n f\"{target}_prob\"\n ])\n\nmissing_columns = [\n column\n for column in required_columns\n if column not in oof_df.columns\n]\n\nif missing_columns:\n raise RuntimeError(\n \"Missing OOF columns: \"\n + \", \".join(missing_columns)\n )\n\nif len(oof_df) != 58:\n raise RuntimeError(\n f\"Expected 58 OOF rows, found {len(oof_df)}.\"\n )\n\nif (\n oof_df[\"StudyInstanceUID\"].nunique()\n != 58\n):\n raise RuntimeError(\n \"Expected 58 unique studies.\"\n )\n\nif (\n oof_df.groupby(\"StudyInstanceUID\")\n .size()\n .max()\n != 1\n):\n raise RuntimeError(\n \"Duplicate StudyInstanceUID detected.\"\n )\n\nif set(\n oof_df[\"fold\"].astype(int).unique()\n) != {0, 1, 2}:\n\n raise RuntimeError(\n \"Expected folds 0, 1, and 2.\"\n )\n\nprint(\"OOF structure: PASS\")\nprint(\"Study uniqueness: PASS\")\nprint(\"Fold structure: PASS\")\n\n# ----------------------------------------------------------------\n# TARGET-WISE ERROR ANALYSIS\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TARGET-WISE OOF ERROR ANALYSIS\")\nprint(\"=\" * 70)\n\ntarget_error_rows = []\n\nfor target in TARGETS:\n\n y_true = oof_df[\n f\"{target}_label\"\n ].astype(int).values\n\n y_prob = oof_df[\n f\"{target}_prob\"\n ].astype(float).values\n\n y_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n tp = int(\n ((y_true == 1) & (y_pred == 1)).sum()\n )\n\n tn = int(\n ((y_true == 0) & (y_pred == 0)).sum()\n )\n\n fp = int(\n ((y_true == 0) & (y_pred == 1)).sum()\n )\n\n fn = int(\n ((y_true == 1) & (y_pred == 0)).sum()\n )\n\n error_count = fp + fn\n\n error_rate = (\n error_count / len(y_true)\n )\n\n f1 = float(\n f1_score(\n y_true,\n y_pred,\n zero_division=0\n )\n )\n\n precision = float(\n precision_score(\n y_true,\n y_pred,\n zero_division=0\n )\n )\n\n recall = float(\n recall_score(\n y_true,\n y_pred,\n zero_division=0\n )\n )\n\n auc = float(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n\n target_error_rows.append({\n \"target\": target,\n \"TP\": tp,\n \"TN\": tn,\n \"FP\": fp,\n \"FN\": fn,\n \"total_errors\": error_count,\n \"error_rate\": error_rate,\n \"precision\": precision,\n \"recall\": recall,\n \"f1\": f1,\n \"roc_auc\": auc\n })\n\ntarget_error_df = pd.DataFrame(\n target_error_rows\n)\n\nprint(\n target_error_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# FOLD-WISE PERFORMANCE\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD-WISE OOF PERFORMANCE\")\nprint(\"=\" * 70)\n\nfold_rows = []\n\nfor fold in [0, 1, 2]:\n\n fold_df = oof_df[\n oof_df[\"fold\"].astype(int)\n == fold\n ].copy()\n\n fold_auc_values = []\n fold_f1_values = []\n\n for target in TARGETS:\n\n y_true = fold_df[\n f\"{target}_label\"\n ].astype(int).values\n\n y_prob = fold_df[\n f\"{target}_prob\"\n ].astype(float).values\n\n y_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n # Every fold was explicitly checked for\n # both classes, but retain a safety check.\n if len(\n np.unique(y_true)\n ) < 2:\n continue\n\n fold_auc_values.append(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n\n fold_f1_values.append(\n f1_score(\n y_true,\n y_pred,\n zero_division=0\n )\n )\n\n fold_rows.append({\n \"fold\": fold,\n \"studies\": len(fold_df),\n \"mean_roc_auc\": float(\n np.mean(fold_auc_values)\n ),\n \"mean_f1\": float(\n np.mean(fold_f1_values)\n )\n })\n\nfold_performance_df = pd.DataFrame(\n fold_rows\n)\n\nprint(\n fold_performance_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# TARGET x FOLD ROC-AUC MATRIX\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TARGET x FOLD ROC-AUC\")\nprint(\"=\" * 70)\n\nauc_matrix = []\n\nfor target in TARGETS:\n\n row = {\n \"target\": target\n }\n\n for fold in [0, 1, 2]:\n\n fold_df = oof_df[\n oof_df[\"fold\"].astype(int)\n == fold\n ]\n\n y_true = fold_df[\n f\"{target}_label\"\n ].astype(int).values\n\n y_prob = fold_df[\n f\"{target}_prob\"\n ].astype(float).values\n\n if len(\n np.unique(y_true)\n ) < 2:\n\n row[\n f\"fold_{fold}_auc\"\n ] = np.nan\n\n else:\n\n row[\n f\"fold_{fold}_auc\"\n ] = float(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n\n auc_matrix.append(row)\n\nauc_matrix_df = pd.DataFrame(\n auc_matrix\n)\n\nprint(\n auc_matrix_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# TARGET x FOLD F1 MATRIX\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TARGET x FOLD F1 @ 0.50\")\nprint(\"=\" * 70)\n\nf1_matrix = []\n\nfor target in TARGETS:\n\n row = {\n \"target\": target\n }\n\n for fold in [0, 1, 2]:\n\n fold_df = oof_df[\n oof_df[\"fold\"].astype(int)\n == fold\n ]\n\n y_true = fold_df[\n f\"{target}_label\"\n ].astype(int).values\n\n y_prob = fold_df[\n f\"{target}_prob\"\n ].astype(float).values\n\n y_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n row[\n f\"fold_{fold}_f1\"\n ] = float(\n f1_score(\n y_true,\n y_pred,\n zero_division=0\n )\n )\n\n f1_matrix.append(row)\n\nf1_matrix_df = pd.DataFrame(\n f1_matrix\n)\n\nprint(\n f1_matrix_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# CROSS-FOLD STABILITY\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CROSS-FOLD TARGET STABILITY\")\nprint(\"=\" * 70)\n\nstability_rows = []\n\nfor target in TARGETS:\n\n auc_values = (\n auc_matrix_df[\n auc_matrix_df[\"target\"]\n == target\n ][\n [\n \"fold_0_auc\",\n \"fold_1_auc\",\n \"fold_2_auc\"\n ]\n ]\n .values\n .flatten()\n )\n\n f1_values = (\n f1_matrix_df[\n f1_matrix_df[\"target\"]\n == target\n ][\n [\n \"fold_0_f1\",\n \"fold_1_f1\",\n \"fold_2_f1\"\n ]\n ]\n .values\n .flatten()\n )\n\n stability_rows.append({\n \"target\": target,\n \"auc_mean\": float(\n np.nanmean(auc_values)\n ),\n \"auc_std\": float(\n np.nanstd(auc_values)\n ),\n \"auc_min\": float(\n np.nanmin(auc_values)\n ),\n \"auc_max\": float(\n np.nanmax(auc_values)\n ),\n \"f1_mean\": float(\n np.mean(f1_values)\n ),\n \"f1_std\": float(\n np.std(f1_values)\n )\n })\n\nstability_df = pd.DataFrame(\n stability_rows\n)\n\nprint(\n stability_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# STUDY-LEVEL ERROR BURDEN\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"STUDY-LEVEL ERROR BURDEN\")\nprint(\"=\" * 70)\n\nstudy_error_rows = []\n\nfor _, row in oof_df.iterrows():\n\n study_id = row[\n \"StudyInstanceUID\"\n ]\n\n fold = int(\n row[\"fold\"]\n )\n\n total_errors = 0\n total_positive_targets = 0\n total_predicted_positive = 0\n total_high_confidence_errors = 0\n\n for target in TARGETS:\n\n label = int(\n row[f\"{target}_label\"]\n )\n\n probability = float(\n row[f\"{target}_prob\"]\n )\n\n prediction = int(\n probability >= 0.50\n )\n\n if label == 1:\n total_positive_targets += 1\n\n if prediction == 1:\n total_predicted_positive += 1\n\n if prediction != label:\n total_errors += 1\n\n # High-confidence error.\n if (\n probability >= 0.75\n or probability <= 0.25\n ):\n total_high_confidence_errors += 1\n\n study_error_rows.append({\n \"StudyInstanceUID\": study_id,\n \"fold\": fold,\n \"total_target_errors\": total_errors,\n \"positive_targets\": total_positive_targets,\n \"predicted_positive_targets\": total_predicted_positive,\n \"high_confidence_errors\": (\n total_high_confidence_errors\n )\n })\n\nstudy_error_df = pd.DataFrame(\n study_error_rows\n)\n\nstudy_error_df = study_error_df.sort_values(\n [\n \"total_target_errors\",\n \"high_confidence_errors\"\n ],\n ascending=False\n).reset_index(drop=True)\n\nprint(\n \"\\nMost difficult studies by total target errors:\"\n)\n\nprint(\n study_error_df.head(\n 15\n ).to_string(\n index=False\n )\n)\n\n# ----------------------------------------------------------------\n# HIGH-CONFIDENCE ERROR SUMMARY\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"HIGH-CONFIDENCE ERROR ANALYSIS\")\nprint(\"=\" * 70)\n\nhigh_confidence_error_rows = []\n\nfor target in TARGETS:\n\n y_true = oof_df[\n f\"{target}_label\"\n ].astype(int).values\n\n y_prob = oof_df[\n f\"{target}_prob\"\n ].astype(float).values\n\n y_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n high_confidence_mask = (\n (\n y_pred != y_true\n )\n &\n (\n (y_prob >= 0.75)\n |\n (y_prob <= 0.25)\n )\n )\n\n high_confidence_error_rows.append({\n \"target\": target,\n \"high_confidence_errors\": int(\n high_confidence_mask.sum()\n ),\n \"total_errors\": int(\n (y_pred != y_true).sum()\n ),\n \"high_confidence_error_rate\": float(\n high_confidence_mask.mean()\n )\n })\n\nhigh_confidence_df = pd.DataFrame(\n high_confidence_error_rows\n)\n\nprint(\n high_confidence_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# ERROR DIRECTION\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FALSE POSITIVE / FALSE NEGATIVE BURDEN\")\nprint(\"=\" * 70)\n\nerror_direction_rows = []\n\nfor target in TARGETS:\n\n y_true = oof_df[\n f\"{target}_label\"\n ].astype(int).values\n\n y_prob = oof_df[\n f\"{target}_prob\"\n ].astype(float).values\n\n y_pred = (\n y_prob >= 0.50\n ).astype(int)\n\n fp = int(\n (\n (y_true == 0)\n &\n (y_pred == 1)\n ).sum()\n )\n\n fn = int(\n (\n (y_true == 1)\n &\n (y_pred == 0)\n ).sum()\n )\n\n if fp > fn:\n dominant_error = \"False Positive\"\n elif fn > fp:\n dominant_error = \"False Negative\"\n else:\n dominant_error = \"Balanced\"\n\n error_direction_rows.append({\n \"target\": target,\n \"false_positives\": fp,\n \"false_negatives\": fn,\n \"FP_minus_FN\": fp - fn,\n \"dominant_error\": dominant_error\n })\n\nerror_direction_df = pd.DataFrame(\n error_direction_rows\n)\n\nprint(\n error_direction_df.to_string(\n index=False\n )\n)\n\n# ----------------------------------------------------------------\n# OVERALL MODEL DIAGNOSTIC\n# ----------------------------------------------------------------\n\nmean_auc = float(\n target_error_df[\n \"roc_auc\"\n ].mean()\n)\n\nmean_f1 = float(\n target_error_df[\n \"f1\"\n ].mean()\n)\n\nmean_error_rate = float(\n target_error_df[\n \"error_rate\"\n ].mean()\n)\n\ntotal_fp = int(\n error_direction_df[\n \"false_positives\"\n ].sum()\n)\n\ntotal_fn = int(\n error_direction_df[\n \"false_negatives\"\n ].sum()\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OVERALL ERROR DIAGNOSTIC\")\nprint(\"=\" * 70)\n\nprint(\n f\"Mean OOF ROC-AUC: {mean_auc:.4f}\"\n)\n\nprint(\n f\"Mean OOF F1 @ 0.50: {mean_f1:.4f}\"\n)\n\nprint(\n f\"Mean target error rate: {mean_error_rate:.4f}\"\n)\n\nprint(\n f\"Total false positives: {total_fp}\"\n)\n\nprint(\n f\"Total false negatives: {total_fn}\"\n)\n\nprint(\n f\"FP/FN ratio: \"\n f\"{total_fp / max(total_fn, 1):.4f}\"\n)\n\n# ----------------------------------------------------------------\n# SAVE RESULTS\n# ----------------------------------------------------------------\n\ntarget_error_path = os.path.join(\n AUDIT_DIR,\n \"cell48_oof_target_errors.csv\"\n)\n\nfold_performance_path = os.path.join(\n AUDIT_DIR,\n \"cell48_oof_fold_performance.csv\"\n)\n\nauc_matrix_path = os.path.join(\n AUDIT_DIR,\n \"cell48_target_fold_auc.csv\"\n)\n\nf1_matrix_path = os.path.join(\n AUDIT_DIR,\n \"cell48_target_fold_f1.csv\"\n)\n\nstability_path = os.path.join(\n AUDIT_DIR,\n \"cell48_target_stability.csv\"\n)\n\nstudy_error_path = os.path.join(\n AUDIT_DIR,\n \"cell48_study_error_burden.csv\"\n)\n\nhigh_confidence_path = os.path.join(\n AUDIT_DIR,\n \"cell48_high_confidence_errors.csv\"\n)\n\nerror_direction_path = os.path.join(\n AUDIT_DIR,\n \"cell48_error_direction.csv\"\n)\n\ntarget_error_df.to_csv(\n target_error_path,\n index=False\n)\n\nfold_performance_df.to_csv(\n fold_performance_path,\n index=False\n)\n\nauc_matrix_df.to_csv(\n auc_matrix_path,\n index=False\n)\n\nf1_matrix_df.to_csv(\n f1_matrix_path,\n index=False\n)\n\nstability_df.to_csv(\n stability_path,\n index=False\n)\n\nstudy_error_df.to_csv(\n study_error_path,\n index=False\n)\n\nhigh_confidence_df.to_csv(\n high_confidence_path,\n index=False\n)\n\nerror_direction_df.to_csv(\n error_direction_path,\n index=False\n)\n\n# ----------------------------------------------------------------\n# FINAL VALIDATION\n# ----------------------------------------------------------------\n\nif len(target_error_df) != 12:\n raise RuntimeError(\n \"Target error table must contain 12 targets.\"\n )\n\nif len(fold_performance_df) != 3:\n raise RuntimeError(\n \"Fold performance table must contain 3 folds.\"\n )\n\nif len(auc_matrix_df) != 12:\n raise RuntimeError(\n \"AUC matrix must contain 12 targets.\"\n )\n\nif len(f1_matrix_df) != 12:\n raise RuntimeError(\n \"F1 matrix must contain 12 targets.\"\n )\n\nif len(stability_df) != 12:\n raise RuntimeError(\n \"Stability table must contain 12 targets.\"\n )\n\nif len(study_error_df) != 58:\n raise RuntimeError(\n \"Study error table must contain 58 studies.\"\n )\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 48 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\"Target-wise error analysis: PASS\")\nprint(\"Fold-wise generalization analysis: PASS\")\nprint(\"Study-level error analysis: PASS\")\nprint(\"High-confidence error analysis: PASS\")\nprint(\"No training performed.\")\nprint(\"No model modified.\")\nprint(\"No test data used.\")\n\nprint(\"\\nSaved:\")\nprint(target_error_path)\nprint(fold_performance_path)\nprint(auc_matrix_path)\nprint(f1_matrix_path)\nprint(stability_path)\nprint(study_error_path)\nprint(high_confidence_path)\nprint(error_direction_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:29:14.684195Z","iopub.execute_input":"2026-08-11T10:29:14.684533Z","iopub.status.idle":"2026-08-11T10:29:15.210617Z","shell.execute_reply.started":"2026-08-11T10:29:14.684505Z","shell.execute_reply":"2026-08-11T10:29:15.209625Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 49 - MODEL CAPACITY + TRAINABLE LAYER AUDIT\n# ================================================================\n\nimport os\nimport torch\nimport torch.nn as nn\n\nprint(\"=\" * 70)\nprint(\"CELL 49 - MODEL CAPACITY + TRAINABLE LAYER AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECT CHECK\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"TARGETS\"\n]\n\nmissing_objects = []\n\nfor name in required_objects:\n if name not in globals():\n missing_objects.append(name)\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# DEVICE CHECK\n# ----------------------------------------------------------------\n\nmodel_device = next(\n model.parameters()\n).device\n\nprint(\n \"Model device:\",\n model_device\n)\n\n# ----------------------------------------------------------------\n# BASIC MODEL INFORMATION\n# ----------------------------------------------------------------\n\ntotal_parameters = sum(\n parameter.numel()\n for parameter in model.parameters()\n)\n\ntrainable_parameters = sum(\n parameter.numel()\n for parameter in model.parameters()\n if parameter.requires_grad\n)\n\nfrozen_parameters = (\n total_parameters\n - trainable_parameters\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"MODEL PARAMETER SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Total parameters: {total_parameters:,}\"\n)\n\nprint(\n f\"Trainable parameters: {trainable_parameters:,}\"\n)\n\nprint(\n f\"Frozen parameters: {frozen_parameters:,}\"\n)\n\nprint(\n f\"Trainable percentage: \"\n f\"{100.0 * trainable_parameters / total_parameters:.2f}%\"\n)\n\n# ----------------------------------------------------------------\n# MODEL STRUCTURE\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"MODEL STRUCTURE\")\nprint(\"=\" * 70)\n\nprint(model)\n\n# ----------------------------------------------------------------\n# TRAINABLE PARAMETER BREAKDOWN\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TRAINABLE PARAMETER BREAKDOWN\")\nprint(\"=\" * 70)\n\ntrainable_rows = []\n\nfor name, parameter in model.named_parameters():\n\n if parameter.requires_grad:\n\n trainable_rows.append({\n \"parameter\": name,\n \"shape\": tuple(\n parameter.shape\n ),\n \"num_parameters\": parameter.numel()\n })\n\nif len(trainable_rows) == 0:\n raise RuntimeError(\n \"No trainable parameters found.\"\n )\n\nfor row in trainable_rows:\n\n print(\n f\"{row['parameter']:<50}\"\n f\"shape={str(row['shape']):<25}\"\n f\"params={row['num_parameters']:,}\"\n )\n\ntrainable_df = None\n\n# ----------------------------------------------------------------\n# TRAINABLE PARAMETERS BY TOP-LEVEL MODULE\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TRAINABLE PARAMETERS BY TOP-LEVEL MODULE\")\nprint(\"=\" * 70)\n\nmodule_parameter_counts = {}\n\nfor name, parameter in model.named_parameters():\n\n if not parameter.requires_grad:\n continue\n\n top_level = name.split(\".\")[0]\n\n if top_level not in module_parameter_counts:\n module_parameter_counts[\n top_level\n ] = 0\n\n module_parameter_counts[\n top_level\n ] += parameter.numel()\n\nfor module_name, count in sorted(\n module_parameter_counts.items()\n):\n\n print(\n f\"{module_name:<25}\"\n f\"{count:,}\"\n )\n\n# ----------------------------------------------------------------\n# ALL PARAMETER MODULE STATUS\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"MODULE TRAINABILITY STATUS\")\nprint(\"=\" * 70)\n\nmodule_rows = []\n\nfor name, module in model.named_modules():\n\n if name == \"\":\n continue\n\n parameters = list(\n module.parameters(\n recurse=False\n )\n )\n\n if not parameters:\n continue\n\n module_total = sum(\n p.numel()\n for p in parameters\n )\n\n module_trainable = sum(\n p.numel()\n for p in parameters\n if p.requires_grad\n )\n\n if module_trainable > 0:\n status = \"TRAINABLE\"\n else:\n status = \"FROZEN\"\n\n module_rows.append({\n \"module\": name,\n \"status\": status,\n \"total_parameters\": module_total,\n \"trainable_parameters\": module_trainable\n })\n\n print(\n f\"{name:<35}\"\n f\"{status:<12}\"\n f\"total={module_total:,} \"\n f\"trainable={module_trainable:,}\"\n )\n\n# ----------------------------------------------------------------\n# INPUT CONVOLUTION AUDIT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"INPUT CHANNEL AUDIT\")\nprint(\"=\" * 70)\n\nconv_candidates = []\n\nfor name, module in model.named_modules():\n\n if isinstance(\n module,\n nn.Conv2d\n ):\n\n conv_candidates.append(\n (\n name,\n module\n )\n )\n\nif len(conv_candidates) == 0:\n raise RuntimeError(\n \"No Conv2d layers found.\"\n )\n\nfirst_conv_name, first_conv = (\n conv_candidates[0]\n)\n\nprint(\n \"First Conv2d:\",\n first_conv_name\n)\n\nprint(\n \"Input channels:\",\n first_conv.in_channels\n)\n\nprint(\n \"Output channels:\",\n first_conv.out_channels\n)\n\nprint(\n \"Kernel:\",\n first_conv.kernel_size\n)\n\nif first_conv.in_channels != 21:\n raise RuntimeError(\n \"Expected first convolution to accept \"\n \"21 input channels, found \"\n f\"{first_conv.in_channels}.\"\n )\n\nprint(\n \"21-channel input: PASS\"\n)\n\n# ----------------------------------------------------------------\n# OUTPUT HEAD AUDIT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OUTPUT HEAD AUDIT\")\nprint(\"=\" * 70)\n\nlinear_candidates = []\n\nfor name, module in model.named_modules():\n\n if isinstance(\n module,\n nn.Linear\n ):\n\n linear_candidates.append(\n (\n name,\n module\n )\n )\n\nif len(linear_candidates) == 0:\n raise RuntimeError(\n \"No Linear layers found.\"\n )\n\nfor name, module in linear_candidates:\n\n print(\n f\"{name}: \"\n f\"in_features={module.in_features}, \"\n f\"out_features={module.out_features}, \"\n f\"trainable={any(p.requires_grad for p in module.parameters())}\"\n )\n\nfinal_linear_name, final_linear = (\n linear_candidates[-1]\n)\n\nif final_linear.out_features != len(TARGETS):\n raise RuntimeError(\n \"Output head mismatch: expected \"\n f\"{len(TARGETS)} outputs, found \"\n f\"{final_linear.out_features}.\"\n )\n\nprint(\n f\"Output classes: {final_linear.out_features}\"\n)\n\nprint(\n \"12-target output head: PASS\"\n)\n\n# ----------------------------------------------------------------\n# TRAINABLE PARAMETER COUNT CONSISTENCY\n# ----------------------------------------------------------------\n\nrecomputed_trainable = sum(\n row[\"num_parameters\"]\n for row in trainable_rows\n)\n\nif (\n recomputed_trainable\n != trainable_parameters\n):\n raise RuntimeError(\n \"Trainable parameter count mismatch.\"\n )\n\nprint(\n \"\\nTrainable parameter count consistency: PASS\"\n)\n\n# ----------------------------------------------------------------\n# MODEL FINITE PARAMETER CHECK\n# ----------------------------------------------------------------\n\ninvalid_parameters = []\n\nfor name, parameter in model.named_parameters():\n\n if not torch.isfinite(\n parameter.detach()\n ).all():\n\n invalid_parameters.append(\n name\n )\n\nif invalid_parameters:\n\n raise RuntimeError(\n \"NaN/Inf detected in model parameters: \"\n + \", \".join(\n invalid_parameters\n )\n )\n\nprint(\n \"Model parameter NaN/Inf check: PASS\"\n)\n\n# ----------------------------------------------------------------\n# FINAL VERDICT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 49 VERDICT\")\nprint(\"=\" * 70)\n\nprint(\n \"Model structure inspection: PASS\"\n)\n\nprint(\n \"Trainable/frozen parameter audit: PASS\"\n)\n\nprint(\n \"21-channel input configuration: PASS\"\n)\n\nprint(\n \"12-target multilabel output: PASS\"\n)\n\nprint(\n \"Model parameter validity: PASS\"\n)\n\nprint(\n \"No training performed.\"\n)\n\nprint(\n \"No model parameters modified.\"\n)\n\nprint(\n \"No dataset modified.\"\n)\n\nprint(\n \"No validation/test data modified.\"\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 49 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"\\nUse this output to determine the next \"\n \"controlled fine-tuning configuration.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:31:14.891427Z","iopub.execute_input":"2026-08-11T10:31:14.891782Z","iopub.status.idle":"2026-08-11T10:31:14.967068Z","shell.execute_reply.started":"2026-08-11T10:31:14.891753Z","shell.execute_reply":"2026-08-11T10:31:14.965955Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 50 - CONTROLLED FINE-TUNING CONFIGURATION AUDIT\n# ================================================================\n\nimport torch\nimport torch.nn as nn\n\nprint(\"=\" * 70)\nprint(\"CELL 50 - CONTROLLED FINE-TUNING CONFIGURATION AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED OBJECT CHECK\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"TARGETS\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------\n\ndevice = next(\n model.parameters()\n).device\n\nprint(\n \"Device:\",\n device\n)\n\n# ----------------------------------------------------------------\n# VERIFY MODEL STRUCTURE\n# ----------------------------------------------------------------\n\nif not hasattr(model, \"backbone\"):\n raise RuntimeError(\n \"Model does not contain 'backbone'.\"\n )\n\nif not hasattr(model, \"classifier\"):\n raise RuntimeError(\n \"Model does not contain 'classifier'.\"\n )\n\nif not hasattr(model.backbone, \"layer4\"):\n raise RuntimeError(\n \"Backbone does not contain 'layer4'.\"\n )\n\nprint(\n \"Backbone structure: PASS\"\n)\n\nprint(\n \"Layer4 availability: PASS\"\n)\n\nprint(\n \"Classifier availability: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY INPUT CHANNELS\n# ----------------------------------------------------------------\n\nfirst_conv = model.backbone.conv1\n\nif first_conv.in_channels != 21:\n raise RuntimeError(\n \"Expected 21 input channels, found \"\n f\"{first_conv.in_channels}.\"\n )\n\nprint(\n \"21-channel input: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY OUTPUT CLASSES\n# ----------------------------------------------------------------\n\nlinear_layers = [\n module\n for module in model.classifier.modules()\n if isinstance(module, nn.Linear)\n]\n\nif not linear_layers:\n raise RuntimeError(\n \"No Linear layers found in classifier.\"\n )\n\nfinal_classifier = linear_layers[-1]\n\nif final_classifier.out_features != len(TARGETS):\n raise RuntimeError(\n \"Classifier output mismatch. Expected \"\n f\"{len(TARGETS)}, found \"\n f\"{final_classifier.out_features}.\"\n )\n\nprint(\n f\"12-target output: PASS\"\n)\n\n# ----------------------------------------------------------------\n# RESET TRAINABILITY\n#\n# Important:\n# We only change requires_grad flags here.\n# No parameter values are modified.\n# ----------------------------------------------------------------\n\nfor parameter in model.parameters():\n parameter.requires_grad = False\n\n# Keep the 21-channel input adaptation trainable.\nfor parameter in model.backbone.conv1.parameters():\n parameter.requires_grad = True\n\n# Unfreeze only the deepest ResNet block.\nfor parameter in model.backbone.layer4.parameters():\n parameter.requires_grad = True\n\n# Keep the classification head trainable.\nfor parameter in model.classifier.parameters():\n parameter.requires_grad = True\n\nprint(\n \"\\nTrainability configuration applied.\"\n)\n\n# ----------------------------------------------------------------\n# TRAINABLE MODULE VERIFICATION\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TRAINABILITY VERIFICATION\")\nprint(\"=\" * 70)\n\nexpected_trainable_prefixes = [\n \"backbone.conv1\",\n \"backbone.layer4\",\n \"classifier\"\n]\n\nunexpected_trainable = []\n\nfor name, parameter in model.named_parameters():\n\n should_train = any(\n name.startswith(prefix)\n for prefix in expected_trainable_prefixes\n )\n\n if parameter.requires_grad != should_train:\n unexpected_trainable.append(\n name\n )\n\nif unexpected_trainable:\n raise RuntimeError(\n \"Unexpected trainability state detected:\\n\"\n + \"\\n\".join(unexpected_trainable)\n )\n\nprint(\n \"Expected trainable modules: PASS\"\n)\n\n# ----------------------------------------------------------------\n# PARAMETER COUNTS\n# ----------------------------------------------------------------\n\ntotal_parameters = sum(\n parameter.numel()\n for parameter in model.parameters()\n)\n\ntrainable_parameters = sum(\n parameter.numel()\n for parameter in model.parameters()\n if parameter.requires_grad\n)\n\nfrozen_parameters = (\n total_parameters\n - trainable_parameters\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FINE-TUNING PARAMETER SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Total parameters: {total_parameters:,}\"\n)\n\nprint(\n f\"Trainable parameters: {trainable_parameters:,}\"\n)\n\nprint(\n f\"Frozen parameters: {frozen_parameters:,}\"\n)\n\nprint(\n f\"Trainable percentage: \"\n f\"{100.0 * trainable_parameters / total_parameters:.2f}%\"\n)\n\n# ----------------------------------------------------------------\n# MODULE-WISE TRAINABLE COUNTS\n# ----------------------------------------------------------------\n\nmodule_counts = {}\n\nfor name, parameter in model.named_parameters():\n\n if not parameter.requires_grad:\n continue\n\n top_level = name.split(\".\")[0]\n\n module_counts[top_level] = (\n module_counts.get(top_level, 0)\n + parameter.numel()\n )\n\nprint(\"\\nTrainable parameters by top-level module:\")\n\nfor name, count in sorted(\n module_counts.items()\n):\n print(\n f\"{name:<20} {count:,}\"\n )\n\n# ----------------------------------------------------------------\n# LAYER4 COUNT\n# ----------------------------------------------------------------\n\nlayer4_trainable = sum(\n parameter.numel()\n for parameter in model.backbone.layer4.parameters()\n if parameter.requires_grad\n)\n\nconv1_trainable = sum(\n parameter.numel()\n for parameter in model.backbone.conv1.parameters()\n if parameter.requires_grad\n)\n\nclassifier_trainable = sum(\n parameter.numel()\n for parameter in model.classifier.parameters()\n if parameter.requires_grad\n)\n\nprint(\"\\nDetailed trainable counts:\")\n\nprint(\n f\"Conv1: {conv1_trainable:,}\"\n)\n\nprint(\n f\"Layer4: {layer4_trainable:,}\"\n)\n\nprint(\n f\"Classifier: {classifier_trainable:,}\"\n)\n\nexpected_total = (\n conv1_trainable\n + layer4_trainable\n + classifier_trainable\n)\n\nif trainable_parameters != expected_total:\n raise RuntimeError(\n \"Trainable parameter count does not match \"\n \"Conv1 + Layer4 + Classifier.\"\n )\n\nprint(\n \"Trainable parameter arithmetic: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY EARLIER RESNET LAYERS REMAIN FROZEN\n# ----------------------------------------------------------------\n\nearlier_layers = [\n \"bn1\",\n \"layer1\",\n \"layer2\",\n \"layer3\"\n]\n\nunexpected_earlier_training = []\n\nfor layer_name in earlier_layers:\n\n layer = getattr(\n model.backbone,\n layer_name\n )\n\n if any(\n parameter.requires_grad\n for parameter in layer.parameters()\n ):\n unexpected_earlier_training.append(\n layer_name\n )\n\nif unexpected_earlier_training:\n raise RuntimeError(\n \"Unexpectedly trainable earlier layers: \"\n + \", \".join(\n unexpected_earlier_training\n )\n )\n\nprint(\n \"Backbone layers 1-3 frozen: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY CLASSIFIER\n# ----------------------------------------------------------------\n\nif not all(\n parameter.requires_grad\n for parameter in model.classifier.parameters()\n):\n raise RuntimeError(\n \"Classifier is not fully trainable.\"\n )\n\nprint(\n \"Classifier trainability: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY MODEL PARAMETERS ARE FINITE\n# ----------------------------------------------------------------\n\ninvalid_parameters = []\n\nfor name, parameter in model.named_parameters():\n\n if not torch.isfinite(\n parameter.detach()\n ).all():\n\n invalid_parameters.append(\n name\n )\n\nif invalid_parameters:\n raise RuntimeError(\n \"NaN/Inf detected in model parameters: \"\n + \", \".join(invalid_parameters)\n )\n\nprint(\n \"Model parameter validity: PASS\"\n)\n\n# ----------------------------------------------------------------\n# OPTIMIZER PLAN\n#\n# Different learning rates are intentional.\n# Layer4 needs a small update.\n# Classifier and the new 21-channel conv1 adaptation\n# can learn faster.\n# ----------------------------------------------------------------\n\nFINE_TUNE_LR_BACKBONE = 1e-5\nFINE_TUNE_LR_CONV1 = 1e-4\nFINE_TUNE_LR_CLASSIFIER = 1e-3\nFINE_TUNE_WEIGHT_DECAY = 1e-4\nFINE_TUNE_GRAD_CLIP = 1.0\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CONTROLLED FINE-TUNING CONFIGURATION\")\nprint(\"=\" * 70)\n\nprint(\n f\"Conv1 learning rate: {FINE_TUNE_LR_CONV1}\"\n)\n\nprint(\n f\"Layer4 learning rate: {FINE_TUNE_LR_BACKBONE}\"\n)\n\nprint(\n f\"Classifier learning rate: {FINE_TUNE_LR_CLASSIFIER}\"\n)\n\nprint(\n f\"Weight decay: {FINE_TUNE_WEIGHT_DECAY}\"\n)\n\nprint(\n f\"Gradient clipping: {FINE_TUNE_GRAD_CLIP}\"\n)\n\n# ----------------------------------------------------------------\n# CREATE OPTIMIZER\n#\n# Only trainable parameters are included.\n# ----------------------------------------------------------------\n\noptimizer_param_groups = [\n {\n \"params\": model.backbone.conv1.parameters(),\n \"lr\": FINE_TUNE_LR_CONV1\n },\n {\n \"params\": model.backbone.layer4.parameters(),\n \"lr\": FINE_TUNE_LR_BACKBONE\n },\n {\n \"params\": model.classifier.parameters(),\n \"lr\": FINE_TUNE_LR_CLASSIFIER\n }\n]\n\nfine_tune_optimizer = torch.optim.AdamW(\n optimizer_param_groups,\n weight_decay=FINE_TUNE_WEIGHT_DECAY\n)\n\nprint(\n \"\\nFine-tuning optimizer created: PASS\"\n)\n\nprint(\n \"Optimizer: AdamW\"\n)\n\nprint(\n \"Parameter groups: 3\"\n)\n\n# ----------------------------------------------------------------\n# OPTIMIZER PARAMETER COUNT\n# ----------------------------------------------------------------\n\noptimizer_parameter_count = sum(\n parameter.numel()\n for group in fine_tune_optimizer.param_groups\n for parameter in group[\"params\"]\n)\n\nif optimizer_parameter_count != trainable_parameters:\n raise RuntimeError(\n \"Optimizer parameter count mismatch. \"\n f\"Optimizer={optimizer_parameter_count:,}, \"\n f\"Trainable={trainable_parameters:,}\"\n )\n\nprint(\n \"Optimizer/trainable parameter count: PASS\"\n)\n\n# ----------------------------------------------------------------\n# FINAL VERDICT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 50 VERDICT\")\nprint(\"=\" * 70)\n\nprint(\n \"Controlled Layer4 fine-tuning configuration: READY\"\n)\n\nprint(\n \"Conv1 trainable: YES\"\n)\n\nprint(\n \"Layer1 frozen: YES\"\n)\n\nprint(\n \"Layer2 frozen: YES\"\n)\n\nprint(\n \"Layer3 frozen: YES\"\n)\n\nprint(\n \"Layer4 trainable: YES\"\n)\n\nprint(\n \"Classifier trainable: YES\"\n)\n\nprint(\n \"Differential learning rates: READY\"\n)\n\nprint(\n \"No training performed.\"\n)\n\nprint(\n \"No checkpoint overwritten.\"\n)\n\nprint(\n \"No dataset modified.\"\n)\n\nprint(\n \"No validation/test predictions generated.\"\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 50 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"\\nNext step: controlled fine-tuning smoke test.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:35:34.380301Z","iopub.execute_input":"2026-08-11T10:35:34.380648Z","iopub.status.idle":"2026-08-11T10:35:34.458955Z","shell.execute_reply.started":"2026-08-11T10:35:34.380618Z","shell.execute_reply":"2026-08-11T10:35:34.458085Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 51 - CONTROLLED FINE-TUNING SMOKE TEST\n# ================================================================\n\nimport copy\nimport torch\n\nprint(\"=\" * 70)\nprint(\"CELL 51 - CONTROLLED FINE-TUNING SMOKE TEST\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"train_loader\",\n \"TARGETS\",\n \"fine_tune_optimizer\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------\n\ndevice = next(model.parameters()).device\n\nprint(\n f\"Device: {device}\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY TRAINABLE PARAMETER CONFIGURATION\n# ----------------------------------------------------------------\n\nexpected_trainable_prefixes = [\n \"backbone.conv1\",\n \"backbone.layer4\",\n \"classifier\"\n]\n\nunexpected_trainable = []\n\nfor name, parameter in model.named_parameters():\n\n expected = any(\n name.startswith(prefix)\n for prefix in expected_trainable_prefixes\n )\n\n if parameter.requires_grad != expected:\n unexpected_trainable.append(name)\n\nif unexpected_trainable:\n raise RuntimeError(\n \"Unexpected trainability state:\\n\"\n + \"\\n\".join(unexpected_trainable)\n )\n\nprint(\n \"Trainability configuration: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY FROZEN LAYERS\n# ----------------------------------------------------------------\n\nfrozen_prefixes = [\n \"backbone.bn1\",\n \"backbone.layer1\",\n \"backbone.layer2\",\n \"backbone.layer3\"\n]\n\nfor prefix in frozen_prefixes:\n\n matching = [\n parameter\n for name, parameter in model.named_parameters()\n if name.startswith(prefix)\n ]\n\n if not matching:\n raise RuntimeError(\n f\"No parameters found for frozen module: {prefix}\"\n )\n\n if any(\n parameter.requires_grad\n for parameter in matching\n ):\n raise RuntimeError(\n f\"Frozen module became trainable: {prefix}\"\n )\n\nprint(\n \"Frozen backbone layers 1-3: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY OPTIMIZER\n# ----------------------------------------------------------------\n\noptimizer_parameter_ids = {\n id(parameter)\n for group in fine_tune_optimizer.param_groups\n for parameter in group[\"params\"]\n}\n\ntrainable_parameter_ids = {\n id(parameter)\n for parameter in model.parameters()\n if parameter.requires_grad\n}\n\nif optimizer_parameter_ids != trainable_parameter_ids:\n raise RuntimeError(\n \"Optimizer parameters do not exactly match \"\n \"the trainable model parameters.\"\n )\n\nprint(\n \"Optimizer parameter coverage: PASS\"\n)\n\n# ----------------------------------------------------------------\n# SAVE PARAMETER SNAPSHOT\n#\n# Used only to verify that the optimizer step actually changes\n# trainable parameters and does not modify frozen parameters.\n# ----------------------------------------------------------------\n\nbefore_state = {\n name: parameter.detach().clone()\n for name, parameter in model.named_parameters()\n}\n\n# ----------------------------------------------------------------\n# LOAD ONE TRAINING BATCH\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"LOADING ONE TRAINING BATCH\")\nprint(\"=\" * 70)\n\nmodel.train()\n\nbatch = next(iter(train_loader))\n\nif not isinstance(batch, dict):\n raise RuntimeError(\n \"Expected dictionary batch.\"\n )\n\nrequired_batch_keys = [\n \"image\",\n \"label\",\n \"study_id\"\n]\n\nmissing_batch_keys = [\n key\n for key in required_batch_keys\n if key not in batch\n]\n\nif missing_batch_keys:\n raise RuntimeError(\n \"Missing batch keys: \"\n + \", \".join(missing_batch_keys)\n )\n\nimages = batch[\"image\"].to(\n device,\n non_blocking=True\n)\n\nlabels = batch[\"label\"].to(\n device,\n non_blocking=True\n)\n\nprint(\n f\"Images: {tuple(images.shape)}\"\n)\n\nprint(\n f\"Labels: {tuple(labels.shape)}\"\n)\n\n# ----------------------------------------------------------------\n# INPUT VALIDATION\n# ----------------------------------------------------------------\n\nexpected_channels = 21\nexpected_targets = len(TARGETS)\n\nif images.ndim != 4:\n raise RuntimeError(\n f\"Expected 4D image tensor, got {images.ndim}D.\"\n )\n\nif images.shape[1] != expected_channels:\n raise RuntimeError(\n f\"Expected {expected_channels} input channels, \"\n f\"found {images.shape[1]}.\"\n )\n\nif labels.ndim != 2:\n raise RuntimeError(\n f\"Expected 2D label tensor, got {labels.ndim}D.\"\n )\n\nif labels.shape[1] != expected_targets:\n raise RuntimeError(\n f\"Expected {expected_targets} targets, \"\n f\"found {labels.shape[1]}.\"\n )\n\nif not torch.isfinite(images).all():\n raise RuntimeError(\n \"NaN/Inf detected in input images.\"\n )\n\nif not torch.isfinite(labels).all():\n raise RuntimeError(\n \"NaN/Inf detected in labels.\"\n )\n\nprint(\n \"Input validation: PASS\"\n)\n\n# ----------------------------------------------------------------\n# LOSS\n#\n# Recreate the same class-weighted BCE formulation used in the\n# original training configuration.\n# ----------------------------------------------------------------\n\npositive_counts = labels.sum(\n dim=0\n)\n\nnegative_counts = (\n labels.shape[0]\n - positive_counts\n)\n\nbatch_pos_weights = torch.where(\n positive_counts > 0,\n negative_counts / positive_counts,\n torch.ones_like(positive_counts)\n)\n\nbatch_pos_weights = batch_pos_weights.to(\n device=device,\n dtype=torch.float32\n)\n\ncriterion_smoke = torch.nn.BCEWithLogitsLoss(\n pos_weight=batch_pos_weights\n)\n\nprint(\n \"Loss: BCEWithLogitsLoss\"\n)\n\n# ----------------------------------------------------------------\n# FORWARD PASS\n# ----------------------------------------------------------------\n\nfine_tune_optimizer.zero_grad(\n set_to_none=True\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FORWARD PASS\")\nprint(\"=\" * 70)\n\nlogits = model(images)\n\nif logits.shape != labels.shape:\n raise RuntimeError(\n \"Output/label shape mismatch. \"\n f\"Logits={tuple(logits.shape)}, \"\n f\"Labels={tuple(labels.shape)}\"\n )\n\nif not torch.isfinite(logits).all():\n raise RuntimeError(\n \"NaN/Inf detected in logits.\"\n )\n\nloss = criterion_smoke(\n logits,\n labels\n)\n\nif not torch.isfinite(loss):\n raise RuntimeError(\n \"Loss is NaN/Inf.\"\n )\n\nprobabilities = torch.sigmoid(\n logits\n)\n\nif not torch.isfinite(probabilities).all():\n raise RuntimeError(\n \"NaN/Inf detected in probabilities.\"\n )\n\nprint(\n f\"Logit shape: {tuple(logits.shape)}\"\n)\n\nprint(\n f\"Loss before update: {loss.item():.6f}\"\n)\n\nprint(\n f\"Logit min: {logits.min().item():.6f}\"\n)\n\nprint(\n f\"Logit max: {logits.max().item():.6f}\"\n)\n\nprint(\n f\"Probability min: {probabilities.min().item():.6f}\"\n)\n\nprint(\n f\"Probability max: {probabilities.max().item():.6f}\"\n)\n\nprint(\n \"Forward pass: PASS\"\n)\n\n# ----------------------------------------------------------------\n# BACKWARD PASS\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"BACKWARD PASS\")\nprint(\"=\" * 70)\n\nloss.backward()\n\ngradient_values = []\n\ninvalid_gradients = []\n\nfor name, parameter in model.named_parameters():\n\n if not parameter.requires_grad:\n continue\n\n if parameter.grad is None:\n raise RuntimeError(\n f\"Missing gradient for trainable parameter: {name}\"\n )\n\n if not torch.isfinite(\n parameter.grad\n ).all():\n invalid_gradients.append(name)\n\n gradient_values.append(\n parameter.grad.detach().norm().item()\n )\n\nif invalid_gradients:\n raise RuntimeError(\n \"NaN/Inf gradients detected:\\n\"\n + \"\\n\".join(invalid_gradients)\n )\n\ntotal_gradient_norm = torch.nn.utils.clip_grad_norm_(\n model.parameters(),\n max_norm=1.0\n)\n\nif not torch.isfinite(\n torch.as_tensor(total_gradient_norm)\n):\n raise RuntimeError(\n \"Invalid gradient norm.\"\n )\n\nprint(\n f\"Gradient norm before clipping: \"\n f\"{float(total_gradient_norm):.6f}\"\n)\n\nprint(\n \"Gradient clipping: PASS\"\n)\n\nprint(\n \"Backward pass: PASS\"\n)\n\n# ----------------------------------------------------------------\n# OPTIMIZER STEP\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"OPTIMIZER STEP\")\nprint(\"=\" * 70)\n\nfine_tune_optimizer.step()\n\nprint(\n \"Optimizer step: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY PARAMETER CHANGES\n# ----------------------------------------------------------------\n\nchanged_trainable = []\nchanged_frozen = []\n\nfor name, parameter in model.named_parameters():\n\n before = before_state[name]\n after = parameter.detach()\n\n changed = not torch.equal(\n before,\n after\n )\n\n if parameter.requires_grad:\n\n if changed:\n changed_trainable.append(name)\n\n else:\n\n if changed:\n changed_frozen.append(name)\n\nif not changed_trainable:\n raise RuntimeError(\n \"No trainable parameters changed after optimizer step.\"\n )\n\nif changed_frozen:\n raise RuntimeError(\n \"Frozen parameters changed unexpectedly:\\n\"\n + \"\\n\".join(changed_frozen)\n )\n\nprint(\n f\"Trainable tensors changed: \"\n f\"{len(changed_trainable)}\"\n)\n\nprint(\n \"Frozen parameters unchanged: PASS\"\n)\n\nprint(\n \"Optimizer parameter update: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY PARAMETERS REMAIN VALID\n# ----------------------------------------------------------------\n\ninvalid_parameters = []\n\nfor name, parameter in model.named_parameters():\n\n if not torch.isfinite(\n parameter.detach()\n ).all():\n\n invalid_parameters.append(name)\n\nif invalid_parameters:\n raise RuntimeError(\n \"NaN/Inf detected after optimizer step:\\n\"\n + \"\\n\".join(invalid_parameters)\n )\n\nprint(\n \"Post-update parameter validity: PASS\"\n)\n\n# ----------------------------------------------------------------\n# DO NOT SAVE CHECKPOINT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 51 VERDICT\")\nprint(\"=\" * 70)\n\nprint(\n \"21-channel input: PASS\"\n)\n\nprint(\n \"12-target output: PASS\"\n)\n\nprint(\n \"Forward pass: PASS\"\n)\n\nprint(\n \"Backward pass: PASS\"\n)\n\nprint(\n \"Gradient clipping: PASS\"\n)\n\nprint(\n \"Differential-learning-rate optimizer: PASS\"\n)\n\nprint(\n \"Trainable parameter update: PASS\"\n)\n\nprint(\n \"Frozen parameter protection: PASS\"\n)\n\nprint(\n \"NaN/Inf validation: PASS\"\n)\n\nprint(\n \"No checkpoint saved.\"\n)\n\nprint(\n \"No existing fold checkpoint overwritten.\"\n)\n\nprint(\n \"One training batch only.\"\n)\n\nprint(\n \"\\nCELL 51 COMPLETE\"\n)\n\nprint(\n \"Controlled fine-tuning smoke test completed.\"\n)\n\nprint(\n \"Next step requires the complete Cell 51 output.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:36:38.252839Z","iopub.execute_input":"2026-08-11T10:36:38.253742Z","iopub.status.idle":"2026-08-11T10:36:39.767706Z","shell.execute_reply.started":"2026-08-11T10:36:38.253708Z","shell.execute_reply":"2026-08-11T10:36:39.766695Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 52 - CONTROLLED FINE-TUNING FOLD 0 EXPERIMENT\n# ================================================================\n\nimport os\nimport copy\nimport torch\nimport pandas as pd\nimport numpy as np\n\nprint(\"=\" * 70)\nprint(\"CELL 52 - CONTROLLED FINE-TUNING FOLD 0 EXPERIMENT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"TARGETS\",\n \"KneeStudyDataset\",\n \"modeling_table\",\n \"study_folds\"\n]\n\nmissing_objects = [\n name for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------\n\ndevice = next(model.parameters()).device\n\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------\n# CONFIGURATION\n# ----------------------------------------------------------------\n\nFOLD = 0\nMAX_EPOCHS = 8\nEARLY_STOPPING_PATIENCE = 3\n\nCONV1_LR = 1e-4\nLAYER4_LR = 1e-5\nCLASSIFIER_LR = 1e-3\n\nWEIGHT_DECAY = 1e-4\nMAX_GRAD_NORM = 1.0\n\nIMAGE_SIZE = 224\nINPUT_CHANNELS = 21\nNUM_TARGETS = len(TARGETS)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"EXPERIMENT CONFIGURATION\")\nprint(\"=\" * 70)\n\nprint(f\"Fold: {FOLD}\")\nprint(f\"Maximum epochs: {MAX_EPOCHS}\")\nprint(f\"Early stopping patience: {EARLY_STOPPING_PATIENCE}\")\nprint(f\"Conv1 learning rate: {CONV1_LR}\")\nprint(f\"Layer4 learning rate: {LAYER4_LR}\")\nprint(f\"Classifier learning rate: {CLASSIFIER_LR}\")\nprint(f\"Weight decay: {WEIGHT_DECAY}\")\nprint(f\"Gradient clipping: {MAX_GRAD_NORM}\")\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------\n# RESTORE MODEL TO ORIGINAL CELL-50 STRUCTURE\n# ----------------------------------------------------------------\n\nmodel.eval()\n\n# Freeze everything first\nfor parameter in model.parameters():\n parameter.requires_grad = False\n\n# Conv1\nfor parameter in model.backbone.conv1.parameters():\n parameter.requires_grad = True\n\n# Layer4\nfor parameter in model.backbone.layer4.parameters():\n parameter.requires_grad = True\n\n# Classifier\nfor parameter in model.classifier.parameters():\n parameter.requires_grad = True\n\n# ----------------------------------------------------------------\n# VERIFY TRAINABILITY\n# ----------------------------------------------------------------\n\nexpected_trainable_prefixes = [\n \"backbone.conv1\",\n \"backbone.layer4\",\n \"classifier\"\n]\n\nunexpected_trainable = []\n\nfor name, parameter in model.named_parameters():\n\n expected = any(\n name.startswith(prefix)\n for prefix in expected_trainable_prefixes\n )\n\n if parameter.requires_grad != expected:\n unexpected_trainable.append(name)\n\nif unexpected_trainable:\n raise RuntimeError(\n \"Incorrect trainability configuration:\\n\"\n + \"\\n\".join(unexpected_trainable)\n )\n\ntrainable_count = sum(\n parameter.numel()\n for parameter in model.parameters()\n if parameter.requires_grad\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"TRAINABILITY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Trainable parameters: {trainable_count:,}\"\n)\n\nif trainable_count != 8_593_996:\n raise RuntimeError(\n f\"Unexpected trainable parameter count: {trainable_count}\"\n )\n\nprint(\"Trainability configuration: PASS\")\n\n# ----------------------------------------------------------------\n# BUILD FOLD-0 TABLES\n# ----------------------------------------------------------------\n\nif \"fold\" not in study_folds.columns:\n raise RuntimeError(\n \"Fold table does not contain 'fold' column.\"\n )\n\nfold0_ids = set(\n study_folds.loc[\n study_folds[\"fold\"] == FOLD,\n \"StudyInstanceUID\"\n ].astype(str)\n)\n\nif len(fold0_ids) != 20:\n raise RuntimeError(\n f\"Expected 20 Fold-0 validation studies, \"\n f\"found {len(fold0_ids)}.\"\n )\n\nall_ids = set(\n modeling_table[\"StudyInstanceUID\"].astype(str)\n)\n\ntrain_ids = all_ids - fold0_ids\n\nif len(train_ids) != 38:\n raise RuntimeError(\n f\"Expected 38 Fold-0 training studies, \"\n f\"found {len(train_ids)}.\"\n )\n\ntrain_table = modeling_table[\n modeling_table[\"StudyInstanceUID\"].astype(str).isin(train_ids)\n].copy()\n\nval_table = modeling_table[\n modeling_table[\"StudyInstanceUID\"].astype(str).isin(fold0_ids)\n].copy()\n\nif set(train_table[\"StudyInstanceUID\"].astype(str)) & \\\n set(val_table[\"StudyInstanceUID\"].astype(str)):\n raise RuntimeError(\n \"Train/validation overlap detected.\"\n )\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 0 DATA\")\nprint(\"=\" * 70)\n\nprint(f\"Training studies: {len(train_table)}\")\nprint(f\"Validation studies: {len(val_table)}\")\nprint(\"Train/validation overlap: 0\")\n\n# ----------------------------------------------------------------\n# BUILD DATASETS USING THE ESTABLISHED PIPELINE\n# ----------------------------------------------------------------\n\ntrain_dataset_ft = KneeStudyDataset(\n train_table,\n targets=TARGETS\n)\n\nval_dataset_ft = KneeStudyDataset(\n val_table,\n targets=TARGETS\n)\n\n# ----------------------------------------------------------------\n# DATALOADERS\n# ----------------------------------------------------------------\n\ntrain_loader_ft = torch.utils.data.DataLoader(\n train_dataset_ft,\n batch_size=2,\n shuffle=True,\n num_workers=0,\n pin_memory=False\n)\n\nval_loader_ft = torch.utils.data.DataLoader(\n val_dataset_ft,\n batch_size=2,\n shuffle=False,\n num_workers=0,\n pin_memory=False\n)\n\nprint(f\"Training batches: {len(train_loader_ft)}\")\nprint(f\"Validation batches: {len(val_loader_ft)}\")\n\n# ----------------------------------------------------------------\n# COMPUTE TRAINING CLASS WEIGHTS\n# ----------------------------------------------------------------\n\ntrain_labels = train_table[TARGETS].astype(float)\n\npositive_counts = train_labels.sum(axis=0)\nnegative_counts = len(train_labels) - positive_counts\n\npos_weight = np.where(\n positive_counts > 0,\n negative_counts / positive_counts,\n 1.0\n)\n\npos_weight_tensor = torch.tensor(\n pos_weight,\n dtype=torch.float32,\n device=device\n)\n\ncriterion = torch.nn.BCEWithLogitsLoss(\n pos_weight=pos_weight_tensor\n)\n\nprint(\"\\nClass-weighted BCE: PASS\")\n\n# ----------------------------------------------------------------\n# DIFFERENTIAL LEARNING-RATE OPTIMIZER\n# ----------------------------------------------------------------\n\noptimizer = torch.optim.AdamW(\n [\n {\n \"params\": model.backbone.conv1.parameters(),\n \"lr\": CONV1_LR\n },\n {\n \"params\": model.backbone.layer4.parameters(),\n \"lr\": LAYER4_LR\n },\n {\n \"params\": model.classifier.parameters(),\n \"lr\": CLASSIFIER_LR\n }\n ],\n weight_decay=WEIGHT_DECAY\n)\n\nscheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(\n optimizer,\n mode=\"min\",\n factor=0.5,\n patience=2\n)\n\nprint(\"Differential-learning-rate optimizer: PASS\")\n\n# ----------------------------------------------------------------\n# METRIC FUNCTION\n# ----------------------------------------------------------------\n\ndef compute_epoch_metrics(\n probabilities,\n labels,\n threshold=0.50\n):\n\n predictions = (\n probabilities >= threshold\n ).astype(np.int32)\n\n f1_values = []\n\n for target_index in range(\n labels.shape[1]\n ):\n\n y_true = labels[:, target_index]\n y_pred = predictions[:, target_index]\n\n tp = np.sum(\n (y_true == 1) & (y_pred == 1)\n )\n\n fp = np.sum(\n (y_true == 0) & (y_pred == 1)\n )\n\n fn = np.sum(\n (y_true == 1) & (y_pred == 0)\n )\n\n denominator = (\n 2 * tp + fp + fn\n )\n\n if denominator == 0:\n f1 = 0.0\n else:\n f1 = (\n 2 * tp\n / denominator\n )\n\n f1_values.append(float(f1))\n\n return float(np.mean(f1_values))\n\n# ----------------------------------------------------------------\n# VALIDATION FUNCTION\n# ----------------------------------------------------------------\n\ndef evaluate_model():\n\n model.eval()\n\n total_loss = 0.0\n all_probabilities = []\n all_labels = []\n\n with torch.no_grad():\n\n for batch in val_loader_ft:\n\n images = batch[\"image\"].to(\n device,\n non_blocking=True\n )\n\n labels = batch[\"label\"].to(\n device,\n non_blocking=True\n )\n\n logits = model(images)\n\n loss = criterion(\n logits,\n labels\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n total_loss += (\n loss.item()\n * images.size(0)\n )\n\n all_probabilities.append(\n probabilities.cpu().numpy()\n )\n\n all_labels.append(\n labels.cpu().numpy()\n )\n\n probabilities = np.concatenate(\n all_probabilities,\n axis=0\n )\n\n labels = np.concatenate(\n all_labels,\n axis=0\n )\n\n average_loss = (\n total_loss\n / len(val_dataset_ft)\n )\n\n mean_f1 = compute_epoch_metrics(\n probabilities,\n labels\n )\n\n return (\n average_loss,\n mean_f1\n )\n\n# ----------------------------------------------------------------\n# TRAINING\n# ----------------------------------------------------------------\n\nhistory = []\n\nbest_val_loss = float(\"inf\")\nbest_epoch = 0\nepochs_without_improvement = 0\n\nbest_state = None\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"STARTING CONTROLLED FOLD 0 FINE-TUNING\")\nprint(\"=\" * 70)\n\nfor epoch in range(\n 1,\n MAX_EPOCHS + 1\n):\n\n model.train()\n\n running_loss = 0.0\n sample_count = 0\n\n for batch in train_loader_ft:\n\n images = batch[\"image\"].to(\n device,\n non_blocking=True\n )\n\n labels = batch[\"label\"].to(\n device,\n non_blocking=True\n )\n\n optimizer.zero_grad(\n set_to_none=True\n )\n\n logits = model(images)\n\n loss = criterion(\n logits,\n labels\n )\n\n if not torch.isfinite(loss):\n raise RuntimeError(\n f\"Non-finite training loss at epoch {epoch}.\"\n )\n\n loss.backward()\n\n torch.nn.utils.clip_grad_norm_(\n model.parameters(),\n max_norm=MAX_GRAD_NORM\n )\n\n optimizer.step()\n\n batch_size = images.size(0)\n\n running_loss += (\n loss.item()\n * batch_size\n )\n\n sample_count += batch_size\n\n train_loss = (\n running_loss\n / sample_count\n )\n\n val_loss, mean_f1 = evaluate_model()\n\n scheduler.step(val_loss)\n\n current_lr = [\n group[\"lr\"]\n for group in optimizer.param_groups\n ]\n\n history.append(\n {\n \"epoch\": epoch,\n \"train_loss\": train_loss,\n \"val_loss\": val_loss,\n \"mean_f1\": mean_f1,\n \"conv1_lr\": current_lr[0],\n \"layer4_lr\": current_lr[1],\n \"classifier_lr\": current_lr[2]\n }\n )\n\n if val_loss < best_val_loss:\n\n best_val_loss = val_loss\n best_epoch = epoch\n epochs_without_improvement = 0\n\n best_state = copy.deepcopy(\n model.state_dict()\n )\n\n else:\n\n epochs_without_improvement += 1\n\n print(\n f\"\\nEpoch {epoch:02d}/{MAX_EPOCHS}\"\n )\n\n print(\n f\"Train Loss: {train_loss:.5f}\"\n )\n\n print(\n f\"Val Loss: {val_loss:.5f}\"\n )\n\n print(\n f\"Mean F1: {mean_f1:.4f}\"\n )\n\n print(\n \"Learning Rates: \"\n f\"Conv1={current_lr[0]:.6f} | \"\n f\"Layer4={current_lr[1]:.6f} | \"\n f\"Classifier={current_lr[2]:.6f}\"\n )\n\n print(\n f\"Best Epoch: {best_epoch}\"\n )\n\n print(\n f\"Best Val Loss: {best_val_loss:.5f}\"\n )\n\n if epochs_without_improvement >= \\\n EARLY_STOPPING_PATIENCE:\n\n print(\n \"\\nEarly stopping triggered.\"\n )\n\n break\n\n# ----------------------------------------------------------------\n# RESTORE BEST EXPERIMENT STATE\n# ----------------------------------------------------------------\n\nif best_state is None:\n raise RuntimeError(\n \"No valid fine-tuning checkpoint state was produced.\"\n )\n\nmodel.load_state_dict(\n best_state\n)\n\nmodel.eval()\n\n# ----------------------------------------------------------------\n# SAVE ONLY EXPERIMENTAL OUTPUT\n# ----------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nos.makedirs(\n audit_dir,\n exist_ok=True\n)\n\ncheckpoint_path = os.path.join(\n audit_dir,\n \"cell52_fold0_finetuned_experiment.pt\"\n)\n\nhistory_path = os.path.join(\n audit_dir,\n \"cell52_fold0_finetuned_history.csv\"\n)\n\ntorch.save(\n model.state_dict(),\n checkpoint_path\n)\n\npd.DataFrame(history).to_csv(\n history_path,\n index=False\n)\n\n# ----------------------------------------------------------------\n# FINAL OUTPUT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 52 - EXPERIMENT COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n f\"Best epoch: {best_epoch}\"\n)\n\nprint(\n f\"Best validation loss: \"\n f\"{best_val_loss:.6f}\"\n)\n\nprint(\n f\"Final validation loss: \"\n f\"{best_val_loss:.6f}\"\n)\n\nprint(\n f\"Experimental checkpoint saved: \"\n f\"{checkpoint_path}\"\n)\n\nprint(\n f\"History saved: \"\n f\"{history_path}\"\n)\n\nprint(\n \"\\nIMPORTANT:\"\n)\n\nprint(\n \"Original Cell-37 Fold-0 checkpoint was NOT overwritten.\"\n)\n\nprint(\n \"This is a controlled fine-tuning experiment only.\"\n)\n\nprint(\n \"\\nCELL 52 COMPLETE\"\n)\n\nprint(\n \"Send the complete Cell 52 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:39:15.588406Z","iopub.execute_input":"2026-08-11T10:39:15.58873Z","iopub.status.idle":"2026-08-11T10:41:50.887839Z","shell.execute_reply.started":"2026-08-11T10:39:15.588705Z","shell.execute_reply":"2026-08-11T10:41:50.886722Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 53 - FINE-TUNED FOLD 0 DIAGNOSTIC AUDIT\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom sklearn.metrics import roc_auc_score\n\nprint(\"=\" * 70)\nprint(\"CELL 53 - FINE-TUNED FOLD 0 DIAGNOSTIC AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"val_loader_ft\",\n \"TARGETS\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n )\n\nprint(\"\\nRequired notebook objects: PASS\")\n\n# ----------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------\n\ndevice = next(model.parameters()).device\n\nprint(\n f\"Device: {device}\"\n)\n\n# ----------------------------------------------------------------\n# CHECK EXPERIMENTAL CHECKPOINT\n# ----------------------------------------------------------------\n\ncheckpoint_path = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell52_fold0_finetuned_experiment.pt\"\n)\n\nif not os.path.exists(checkpoint_path):\n raise RuntimeError(\n \"Fine-tuned Fold-0 checkpoint not found:\\n\"\n + checkpoint_path\n )\n\nprint(\n \"Fine-tuned Fold-0 checkpoint: EXISTS\"\n)\n\n# ----------------------------------------------------------------\n# LOAD EXPERIMENTAL CHECKPOINT\n# ----------------------------------------------------------------\n\ncheckpoint_state = torch.load(\n checkpoint_path,\n map_location=device\n)\n\nmodel.load_state_dict(\n checkpoint_state\n)\n\nmodel.eval()\n\nprint(\n \"Fine-tuned Fold-0 checkpoint: LOADED\"\n)\n\nprint(\n \"Model mode: evaluation\"\n)\n\n# ----------------------------------------------------------------\n# COLLECT VALIDATION PREDICTIONS\n# ----------------------------------------------------------------\n\nall_probabilities = []\nall_labels = []\nall_study_ids = []\n\nwith torch.no_grad():\n\n for batch in val_loader_ft:\n\n images = batch[\"image\"].to(\n device,\n non_blocking=True\n )\n\n labels = batch[\"label\"].to(\n device,\n non_blocking=True\n )\n\n logits = model(images)\n\n if logits.shape != labels.shape:\n raise RuntimeError(\n \"Logit/label shape mismatch: \"\n f\"{tuple(logits.shape)} vs \"\n f\"{tuple(labels.shape)}\"\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n if not torch.isfinite(\n probabilities\n ).all():\n raise RuntimeError(\n \"NaN/Inf detected in probabilities.\"\n )\n\n all_probabilities.append(\n probabilities.cpu().numpy()\n )\n\n all_labels.append(\n labels.cpu().numpy()\n )\n\n all_study_ids.extend(\n batch[\"study_id\"]\n )\n\nprobabilities = np.concatenate(\n all_probabilities,\n axis=0\n)\n\nlabels = np.concatenate(\n all_labels,\n axis=0\n)\n\n# ----------------------------------------------------------------\n# BASIC VALIDATION\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"PREDICTION VALIDATION\")\nprint(\"=\" * 70)\n\nprint(\n f\"Probability shape: {probabilities.shape}\"\n)\n\nprint(\n f\"Label shape: {labels.shape}\"\n)\n\nif probabilities.shape != labels.shape:\n raise RuntimeError(\n \"Prediction/label shape mismatch.\"\n )\n\nif probabilities.shape[1] != len(TARGETS):\n raise RuntimeError(\n \"Unexpected target dimension.\"\n )\n\nif len(all_study_ids) != probabilities.shape[0]:\n raise RuntimeError(\n \"Study ID count does not match predictions.\"\n )\n\nif len(set(all_study_ids)) != len(all_study_ids):\n raise RuntimeError(\n \"Duplicate validation study predictions detected.\"\n )\n\nprint(\n \"Probability validity: PASS\"\n)\n\nprint(\n \"Shape consistency: PASS\"\n)\n\nprint(\n \"One prediction per validation study: PASS\"\n)\n\n# ----------------------------------------------------------------\n# ROC-AUC + PROBABILITY SEPARATION\n# ----------------------------------------------------------------\n\nrows = []\n\nfor index, target in enumerate(TARGETS):\n\n y_true = labels[:, index]\n y_prob = probabilities[:, index]\n\n positive_mask = (\n y_true == 1\n )\n\n negative_mask = (\n y_true == 0\n )\n\n positive_count = int(\n positive_mask.sum()\n )\n\n negative_count = int(\n negative_mask.sum()\n )\n\n positive_mean = (\n float(y_prob[positive_mask].mean())\n if positive_count > 0\n else np.nan\n )\n\n negative_mean = (\n float(y_prob[negative_mask].mean())\n if negative_count > 0\n else np.nan\n )\n\n positive_median = (\n float(np.median(y_prob[positive_mask]))\n if positive_count > 0\n else np.nan\n )\n\n negative_median = (\n float(np.median(y_prob[negative_mask]))\n if negative_count > 0\n else np.nan\n )\n\n separation = (\n positive_mean\n - negative_mean\n )\n\n if (\n positive_count > 0\n and negative_count > 0\n ):\n roc_auc = roc_auc_score(\n y_true,\n y_prob\n )\n else:\n roc_auc = np.nan\n\n rows.append(\n {\n \"target\": target,\n \"positive_count\": positive_count,\n \"negative_count\": negative_count,\n \"positive_mean\": positive_mean,\n \"negative_mean\": negative_mean,\n \"positive_median\": positive_median,\n \"negative_median\": negative_median,\n \"probability_separation\": separation,\n \"roc_auc\": roc_auc\n }\n )\n\nseparation_df = pd.DataFrame(rows)\n\n# ----------------------------------------------------------------\n# THRESHOLD AUDIT\n# ----------------------------------------------------------------\n\nthresholds = np.arange(\n 0.10,\n 0.901,\n 0.025\n)\n\nthreshold_rows = []\n\nfor index, target in enumerate(TARGETS):\n\n y_true = labels[:, index]\n y_prob = probabilities[:, index]\n\n best_threshold = 0.50\n best_f1 = 0.0\n f1_at_050 = 0.0\n\n for threshold in thresholds:\n\n y_pred = (\n y_prob >= threshold\n ).astype(np.int32)\n\n tp = np.sum(\n (y_true == 1)\n & (y_pred == 1)\n )\n\n fp = np.sum(\n (y_true == 0)\n & (y_pred == 1)\n )\n\n fn = np.sum(\n (y_true == 1)\n & (y_pred == 0)\n )\n\n denominator = (\n 2 * tp\n + fp\n + fn\n )\n\n if denominator == 0:\n f1 = 0.0\n else:\n f1 = (\n 2 * tp\n / denominator\n )\n\n if abs(threshold - 0.50) < 1e-8:\n f1_at_050 = float(f1)\n\n if f1 > best_f1:\n best_f1 = float(f1)\n best_threshold = float(threshold)\n\n threshold_rows.append(\n {\n \"target\": target,\n \"f1_at_0.50\": f1_at_050,\n \"best_threshold\": best_threshold,\n \"best_f1\": best_f1,\n \"f1_improvement\": (\n best_f1 - f1_at_050\n )\n }\n )\n\nthreshold_df = pd.DataFrame(\n threshold_rows\n)\n\n# ----------------------------------------------------------------\n# MERGE DIAGNOSTICS\n# ----------------------------------------------------------------\n\ndiagnostics_df = separation_df.merge(\n threshold_df,\n on=\"target\",\n how=\"inner\"\n)\n\n# ----------------------------------------------------------------\n# SUMMARY\n# ----------------------------------------------------------------\n\nmean_roc_auc = float(\n diagnostics_df[\"roc_auc\"].mean()\n)\n\nmean_separation = float(\n diagnostics_df[\n \"probability_separation\"\n ].mean()\n)\n\nmean_f1_050 = float(\n diagnostics_df[\n \"f1_at_0.50\"\n ].mean()\n)\n\nmean_best_f1 = float(\n diagnostics_df[\n \"best_f1\"\n ].mean()\n)\n\nf1_improvement = (\n mean_best_f1\n - mean_f1_050\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FINE-TUNED FOLD 0 DIAGNOSTICS\")\nprint(\"=\" * 70)\n\nprint(\n diagnostics_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Mean ROC-AUC: {mean_roc_auc:.4f}\"\n)\n\nprint(\n f\"Mean probability separation: \"\n f\"{mean_separation:.4f}\"\n)\n\nprint(\n f\"Mean F1 @ 0.50: {mean_f1_050:.4f}\"\n)\n\nprint(\n f\"Mean best-threshold F1: \"\n f\"{mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Potential F1 improvement: \"\n f\"{f1_improvement:.4f}\"\n)\n\nhigh_auc_targets = diagnostics_df.loc[\n diagnostics_df[\"roc_auc\"] >= 0.65,\n \"target\"\n].tolist()\n\nnegative_separation_targets = diagnostics_df.loc[\n diagnostics_df[\n \"probability_separation\"\n ] < 0,\n \"target\"\n].tolist()\n\nlarge_f1_gain_targets = diagnostics_df.loc[\n diagnostics_df[\n \"f1_improvement\"\n ] > 0.05,\n \"target\"\n].tolist()\n\nprint(\n f\"\\nTargets with ROC-AUC >= 0.65: \"\n f\"{len(high_auc_targets)}\"\n)\n\nprint(\n high_auc_targets\n)\n\nprint(\n f\"\\nTargets with negative probability separation: \"\n f\"{len(negative_separation_targets)}\"\n)\n\nprint(\n negative_separation_targets\n)\n\nprint(\n f\"\\nTargets with >0.05 F1 improvement: \"\n f\"{len(large_f1_gain_targets)}\"\n)\n\nprint(\n large_f1_gain_targets\n)\n\n# ----------------------------------------------------------------\n# SAVE\n# ----------------------------------------------------------------\n\naudit_dir = (\n \"/kaggle/working/rsna_knee_audit\"\n)\n\ndiagnostics_path = os.path.join(\n audit_dir,\n \"cell53_fold0_finetuned_diagnostics.csv\"\n)\n\nthreshold_path = os.path.join(\n audit_dir,\n \"cell53_fold0_finetuned_threshold_audit.csv\"\n)\n\ndiagnostics_df.to_csv(\n diagnostics_path,\n index=False\n)\n\nthreshold_df.to_csv(\n threshold_path,\n index=False\n)\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 53 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n f\"Saved: {diagnostics_path}\"\n)\n\nprint(\n f\"Saved: {threshold_path}\"\n)\n\nprint(\n \"No training performed.\"\n)\n\nprint(\n \"No original checkpoint modified.\"\n)\n\nprint(\n \"No test data used.\"\n)\n\nprint(\n \"Send the complete Cell 53 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:44:35.222842Z","iopub.execute_input":"2026-08-11T10:44:35.223327Z","iopub.status.idle":"2026-08-11T10:44:44.741429Z","shell.execute_reply.started":"2026-08-11T10:44:35.223296Z","shell.execute_reply":"2026-08-11T10:44:44.740432Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 54R - CONTROLLED FINE-TUNING FOLD 1\n# CORRECTED CONTINUATION AFTER MODEL-RESET FAILURE\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nimport torch.nn as nn\n\nprint(\"=\" * 70)\nprint(\"CELL 54R - CONTROLLED FINE-TUNING FOLD 1\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# VERIFY OBJECTS CREATED SUCCESSFULLY BY CELL 54\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"train_loader_ft\",\n \"val_loader_ft\",\n \"train_table\",\n \"val_table\",\n \"TARGETS\",\n \"device\"\n]\n\nmissing_objects = [\n name for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------\n# PATHS\n# ----------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nbaseline_checkpoint = os.path.join(\n audit_dir,\n \"cell40_fold1_best_model.pt\"\n)\n\nexperimental_checkpoint = os.path.join(\n audit_dir,\n \"cell54_fold1_finetuned_experiment.pt\"\n)\n\nhistory_path = os.path.join(\n audit_dir,\n \"cell54_fold1_finetuned_history.csv\"\n)\n\nmetrics_path = os.path.join(\n audit_dir,\n \"cell54_fold1_finetuned_metrics.csv\"\n)\n\nif not os.path.exists(baseline_checkpoint):\n raise RuntimeError(\n \"Original Cell-40 Fold-1 checkpoint not found: \"\n + baseline_checkpoint\n )\n\nprint(\n \"Original Fold-1 baseline checkpoint: EXISTS\"\n)\n\n# ----------------------------------------------------------------\n# LOAD ORIGINAL FOLD-1 BASELINE MODEL\n# ----------------------------------------------------------------\n# IMPORTANT:\n# Cell 40 is the baseline Fold-1 model.\n# Cell 54 fine-tuning must start from that checkpoint.\n# Cell 40 checkpoint itself is never modified.\n\nbaseline_state = torch.load(\n baseline_checkpoint,\n map_location=device\n)\n\n# Support both a plain state_dict and a checkpoint dictionary,\n# without assuming a checkpoint structure that was not established.\n\nif isinstance(baseline_state, dict):\n\n if \"state_dict\" in baseline_state:\n baseline_state_dict = baseline_state[\"state_dict\"]\n\n elif \"model_state_dict\" in baseline_state:\n baseline_state_dict = baseline_state[\"model_state_dict\"]\n\n else:\n baseline_state_dict = baseline_state\n\nelse:\n raise RuntimeError(\n \"Unsupported Fold-1 checkpoint format.\"\n )\n\nmodel.load_state_dict(\n baseline_state_dict,\n strict=True\n)\n\nmodel = model.to(device)\n\nprint(\n \"Original Cell-40 Fold-1 checkpoint loaded into existing model: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY MODEL ARCHITECTURE\n# ----------------------------------------------------------------\n\nif not hasattr(model, \"backbone\"):\n raise RuntimeError(\n \"Existing model does not contain 'backbone'.\"\n )\n\nif not hasattr(model.backbone, \"conv1\"):\n raise RuntimeError(\n \"Existing model backbone does not contain 'conv1'.\"\n )\n\nif not hasattr(model.backbone, \"layer1\"):\n raise RuntimeError(\n \"Existing model backbone does not contain 'layer1'.\"\n )\n\nif not hasattr(model.backbone, \"layer2\"):\n raise RuntimeError(\n \"Existing model backbone does not contain 'layer2'.\"\n )\n\nif not hasattr(model.backbone, \"layer3\"):\n raise RuntimeError(\n \"Existing model backbone does not contain 'layer3'.\"\n )\n\nif not hasattr(model.backbone, \"layer4\"):\n raise RuntimeError(\n \"Existing model backbone does not contain 'layer4'.\"\n )\n\nif not hasattr(model, \"classifier\"):\n raise RuntimeError(\n \"Existing model does not contain 'classifier'.\"\n )\n\nprint(\"Existing model architecture: PASS\")\n\n# ----------------------------------------------------------------\n# VERIFY 21-CHANNEL INPUT AND 12-CLASS OUTPUT\n# ----------------------------------------------------------------\n\nsmoke_batch = next(iter(train_loader_ft))\n\nimages = smoke_batch[\"image\"].to(device)\nlabels = smoke_batch[\"label\"].to(device)\n\nif tuple(images.shape) != (2, 21, 224, 224):\n raise RuntimeError(\n \"Unexpected input shape: \"\n + str(tuple(images.shape))\n )\n\nif tuple(labels.shape) != (2, 12):\n raise RuntimeError(\n \"Unexpected label shape: \"\n + str(tuple(labels.shape))\n )\n\nmodel.eval()\n\nwith torch.no_grad():\n smoke_logits = model(images)\n\nif tuple(smoke_logits.shape) != (2, 12):\n raise RuntimeError(\n \"Unexpected model output shape: \"\n + str(tuple(smoke_logits.shape))\n )\n\nif not torch.isfinite(smoke_logits).all():\n raise RuntimeError(\n \"NaN/Inf detected in model output.\"\n )\n\nprint(\"21-channel input: PASS\")\nprint(\"12-target output: PASS\")\nprint(\"Forward-pass validation: PASS\")\n\n# ----------------------------------------------------------------\n# CONTROLLED TRAINABILITY\n# ----------------------------------------------------------------\n\nfor parameter in model.parameters():\n parameter.requires_grad = False\n\n# Conv1 trainable\nfor parameter in model.backbone.conv1.parameters():\n parameter.requires_grad = True\n\n# Layer4 trainable\nfor parameter in model.backbone.layer4.parameters():\n parameter.requires_grad = True\n\n# Classifier trainable\nfor parameter in model.classifier.parameters():\n parameter.requires_grad = True\n\n# ----------------------------------------------------------------\n# TRAINABILITY VERIFICATION\n# ----------------------------------------------------------------\n\nlayer1_trainable = any(\n parameter.requires_grad\n for parameter in model.backbone.layer1.parameters()\n)\n\nlayer2_trainable = any(\n parameter.requires_grad\n for parameter in model.backbone.layer2.parameters()\n)\n\nlayer3_trainable = any(\n parameter.requires_grad\n for parameter in model.backbone.layer3.parameters()\n)\n\nconv1_trainable = any(\n parameter.requires_grad\n for parameter in model.backbone.conv1.parameters()\n)\n\nlayer4_trainable = any(\n parameter.requires_grad\n for parameter in model.backbone.layer4.parameters()\n)\n\nclassifier_trainable = any(\n parameter.requires_grad\n for parameter in model.classifier.parameters()\n)\n\nif layer1_trainable:\n raise RuntimeError(\n \"Layer1 is unexpectedly trainable.\"\n )\n\nif layer2_trainable:\n raise RuntimeError(\n \"Layer2 is unexpectedly trainable.\"\n )\n\nif layer3_trainable:\n raise RuntimeError(\n \"Layer3 is unexpectedly trainable.\"\n )\n\nif not conv1_trainable:\n raise RuntimeError(\n \"Conv1 is not trainable.\"\n )\n\nif not layer4_trainable:\n raise RuntimeError(\n \"Layer4 is not trainable.\"\n )\n\nif not classifier_trainable:\n raise RuntimeError(\n \"Classifier is not trainable.\"\n )\n\nprint(\"Controlled trainability: PASS\")\nprint(\"Conv1 trainable: YES\")\nprint(\"Layer1 frozen: YES\")\nprint(\"Layer2 frozen: YES\")\nprint(\"Layer3 frozen: YES\")\nprint(\"Layer4 trainable: YES\")\nprint(\"Classifier trainable: YES\")\n\n# ----------------------------------------------------------------\n# TRAINABLE PARAMETER COUNT\n# ----------------------------------------------------------------\n\ntrainable_parameter_count = sum(\n parameter.numel()\n for parameter in model.parameters()\n if parameter.requires_grad\n)\n\nif trainable_parameter_count != 8_593_996:\n raise RuntimeError(\n \"Unexpected trainable parameter count: \"\n + str(trainable_parameter_count)\n )\n\nprint(\n f\"Trainable parameters: {trainable_parameter_count:,}\"\n)\nprint(\"Trainable parameter count: PASS\")\n\n# ----------------------------------------------------------------\n# FOLD 1 CLASS WEIGHTS\n# ----------------------------------------------------------------\n\npos_weights = []\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"FOLD 1 CLASS WEIGHTS\")\nprint(\"=\" * 70)\n\nfor target in TARGETS:\n\n positive = int(\n train_table[target].sum()\n )\n\n negative = (\n len(train_table) - positive\n )\n\n if positive <= 0:\n raise RuntimeError(\n f\"No positive training samples for {target}.\"\n )\n\n pos_weight = (\n negative / positive\n )\n\n pos_weights.append(\n pos_weight\n )\n\n print(\n f\"{target:<20} \"\n f\"positive={positive:2d} \"\n f\"negative={negative:2d} \"\n f\"pos_weight={pos_weight:.4f}\"\n )\n\npos_weight_tensor = torch.tensor(\n pos_weights,\n dtype=torch.float32,\n device=device\n)\n\ncriterion = nn.BCEWithLogitsLoss(\n pos_weight=pos_weight_tensor\n)\n\nprint(\"Class-weighted BCE: PASS\")\n\n# ----------------------------------------------------------------\n# DIFFERENTIAL LEARNING-RATE OPTIMIZER\n# ----------------------------------------------------------------\n\noptimizer = torch.optim.AdamW(\n [\n {\n \"params\": model.backbone.conv1.parameters(),\n \"lr\": 1e-4\n },\n {\n \"params\": model.backbone.layer4.parameters(),\n \"lr\": 1e-5\n },\n {\n \"params\": model.classifier.parameters(),\n \"lr\": 1e-3\n }\n ],\n weight_decay=1e-4\n)\n\nprint(\"Differential-learning-rate optimizer: PASS\")\n\n# ----------------------------------------------------------------\n# EXPERIMENT CONFIGURATION\n# ----------------------------------------------------------------\n\nmax_epochs = 8\npatience = 3\ngradient_clip_norm = 1.0\n\nbest_val_loss = float(\"inf\")\nbest_epoch = 0\nepochs_without_improvement = 0\n\nhistory = []\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"EXPERIMENT CONFIGURATION\")\nprint(\"=\" * 70)\n\nprint(\"Fold: 1\")\nprint(f\"Maximum epochs: {max_epochs}\")\nprint(f\"Early stopping patience: {patience}\")\nprint(\"Conv1 learning rate: 0.0001\")\nprint(\"Layer4 learning rate: 0.00001\")\nprint(\"Classifier learning rate: 0.001\")\nprint(\"Weight decay: 0.0001\")\nprint(\"Gradient clipping: 1.0\")\nprint(f\"Device: {device}\")\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"STARTING CONTROLLED FOLD 1 FINE-TUNING\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# TRAINING LOOP\n# ----------------------------------------------------------------\n\nfor epoch in range(1, max_epochs + 1):\n\n model.train()\n\n train_loss_sum = 0.0\n train_batch_count = 0\n\n for batch in train_loader_ft:\n\n batch_images = batch[\"image\"].to(\n device\n )\n\n batch_labels = batch[\"label\"].to(\n device\n )\n\n optimizer.zero_grad(\n set_to_none=True\n )\n\n logits = model(\n batch_images\n )\n\n loss = criterion(\n logits,\n batch_labels\n )\n\n if not torch.isfinite(loss):\n raise RuntimeError(\n \"NaN/Inf training loss detected.\"\n )\n\n loss.backward()\n\n torch.nn.utils.clip_grad_norm_(\n model.parameters(),\n max_norm=gradient_clip_norm\n )\n\n optimizer.step()\n\n train_loss_sum += loss.item()\n train_batch_count += 1\n\n train_loss = (\n train_loss_sum\n / train_batch_count\n )\n\n # ------------------------------------------------------------\n # VALIDATION\n # ------------------------------------------------------------\n\n model.eval()\n\n val_loss_sum = 0.0\n val_batch_count = 0\n\n all_val_probabilities = []\n all_val_labels = []\n\n with torch.no_grad():\n\n for batch in val_loader_ft:\n\n batch_images = batch[\"image\"].to(\n device\n )\n\n batch_labels = batch[\"label\"].to(\n device\n )\n\n logits = model(\n batch_images\n )\n\n loss = criterion(\n logits,\n batch_labels\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n if not torch.isfinite(loss):\n raise RuntimeError(\n \"NaN/Inf validation loss detected.\"\n )\n\n if not torch.isfinite(\n probabilities\n ).all():\n raise RuntimeError(\n \"NaN/Inf validation probabilities detected.\"\n )\n\n val_loss_sum += loss.item()\n val_batch_count += 1\n\n all_val_probabilities.append(\n probabilities.cpu()\n )\n\n all_val_labels.append(\n batch_labels.cpu()\n )\n\n val_loss = (\n val_loss_sum\n / val_batch_count\n )\n\n val_probabilities = torch.cat(\n all_val_probabilities,\n dim=0\n ).numpy()\n\n val_labels_np = torch.cat(\n all_val_labels,\n dim=0\n ).numpy()\n\n if val_probabilities.shape != (\n len(val_table),\n len(TARGETS)\n ):\n raise RuntimeError(\n \"Unexpected validation probability shape: \"\n + str(val_probabilities.shape)\n )\n\n # ------------------------------------------------------------\n # F1 @ 0.50\n # ------------------------------------------------------------\n\n f1_values = []\n\n for target_index in range(\n len(TARGETS)\n ):\n\n y_true = val_labels_np[\n :, target_index\n ]\n\n y_pred = (\n val_probabilities[\n :, target_index\n ] >= 0.50\n ).astype(np.int32)\n\n tp = np.sum(\n (y_true == 1)\n & (y_pred == 1)\n )\n\n fp = np.sum(\n (y_true == 0)\n & (y_pred == 1)\n )\n\n fn = np.sum(\n (y_true == 1)\n & (y_pred == 0)\n )\n\n denominator = (\n 2 * tp + fp + fn\n )\n\n if denominator == 0:\n f1 = 0.0\n else:\n f1 = (\n 2 * tp\n / denominator\n )\n\n f1_values.append(\n f1\n )\n\n mean_f1 = float(\n np.mean(f1_values)\n )\n\n # ------------------------------------------------------------\n # REDUCE LR ON PLATEAU\n # ------------------------------------------------------------\n\n if val_loss < best_val_loss - 1e-6:\n\n best_val_loss = val_loss\n best_epoch = epoch\n epochs_without_improvement = 0\n\n torch.save(\n model.state_dict(),\n experimental_checkpoint\n )\n\n else:\n\n epochs_without_improvement += 1\n\n if epochs_without_improvement >= 2:\n\n for group in optimizer.param_groups:\n group[\"lr\"] *= 0.5\n\n epochs_without_improvement = 0\n\n current_lrs = [\n group[\"lr\"]\n for group in optimizer.param_groups\n ]\n\n history.append(\n {\n \"epoch\": epoch,\n \"train_loss\": train_loss,\n \"val_loss\": val_loss,\n \"mean_f1\": mean_f1,\n \"conv1_lr\": current_lrs[0],\n \"layer4_lr\": current_lrs[1],\n \"classifier_lr\": current_lrs[2]\n }\n )\n\n print(\n f\"\\nEpoch {epoch:02d}/{max_epochs}\"\n )\n\n print(\n f\"Train Loss: {train_loss:.5f}\"\n )\n\n print(\n f\"Val Loss: {val_loss:.5f}\"\n )\n\n print(\n f\"Mean F1: {mean_f1:.4f}\"\n )\n\n print(\n \"Learning Rates: \"\n f\"Conv1={current_lrs[0]:.6f} | \"\n f\"Layer4={current_lrs[1]:.6f} | \"\n f\"Classifier={current_lrs[2]:.6f}\"\n )\n\n print(\n f\"Best Epoch: {best_epoch}\"\n )\n\n print(\n f\"Best Val Loss: {best_val_loss:.5f}\"\n )\n\n # ------------------------------------------------------------\n # EARLY STOPPING\n # ------------------------------------------------------------\n\n if (\n epoch - best_epoch\n >= patience\n ):\n print(\n \"\\nEarly stopping triggered.\"\n )\n break\n\n# ----------------------------------------------------------------\n# RESTORE BEST EXPERIMENTAL CHECKPOINT\n# ----------------------------------------------------------------\n\nif not os.path.exists(\n experimental_checkpoint\n):\n raise RuntimeError(\n \"Experimental Fold-1 checkpoint was not saved.\"\n )\n\nbest_state = torch.load(\n experimental_checkpoint,\n map_location=device\n)\n\nmodel.load_state_dict(\n best_state,\n strict=True\n)\n\nmodel.eval()\n\n# ----------------------------------------------------------------\n# SAVE HISTORY\n# ----------------------------------------------------------------\n\nhistory_df = pd.DataFrame(\n history\n)\n\nhistory_df.to_csv(\n history_path,\n index=False\n)\n\nmetrics_df = pd.DataFrame(\n {\n \"fold\": [1],\n \"best_epoch\": [best_epoch],\n \"best_validation_loss\": [\n best_val_loss\n ]\n }\n)\n\nmetrics_df.to_csv(\n metrics_path,\n index=False\n)\n\n# ----------------------------------------------------------------\n# FINAL VERIFICATION\n# ----------------------------------------------------------------\n\nif not torch.isfinite(\n torch.cat(\n [\n parameter.detach().flatten()\n for parameter in model.parameters()\n ]\n )\n).all():\n raise RuntimeError(\n \"NaN/Inf detected in final model parameters.\"\n )\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 54R - EXPERIMENT COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n f\"Best epoch: {best_epoch}\"\n)\n\nprint(\n f\"Best validation loss: \"\n f\"{best_val_loss:.6f}\"\n)\n\nprint(\n f\"Experimental checkpoint saved: \"\n f\"{experimental_checkpoint}\"\n)\n\nprint(\n f\"History saved: {history_path}\"\n)\n\nprint(\n f\"Metrics saved: {metrics_path}\"\n)\n\nprint(\n \"\\nIMPORTANT:\"\n)\n\nprint(\n \"Original Cell-40 Fold-1 checkpoint was NOT overwritten.\"\n)\n\nprint(\n \"This is a controlled fine-tuning experiment only.\"\n)\n\nprint(\n \"\\nCELL 54R COMPLETE\"\n)\n\nprint(\n \"Send the complete Cell 54R output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:48:53.779405Z","iopub.execute_input":"2026-08-11T10:48:53.77977Z","iopub.status.idle":"2026-08-11T10:52:25.939632Z","shell.execute_reply.started":"2026-08-11T10:48:53.779741Z","shell.execute_reply":"2026-08-11T10:52:25.938458Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 55 - FOLD 1 FINE-TUNED DIAGNOSTIC\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom sklearn.metrics import roc_auc_score\n\nprint(\"=\" * 70)\nprint(\"CELL 55 - FOLD 1 FINE-TUNED DIAGNOSTIC\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"val_loader_ft\",\n \"TARGETS\",\n \"device\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------\n# CHECKPOINT\n# ----------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nfinetuned_checkpoint = os.path.join(\n audit_dir,\n \"cell54_fold1_finetuned_experiment.pt\"\n)\n\noutput_separation = os.path.join(\n audit_dir,\n \"cell55_fold1_finetuned_probability_separation.csv\"\n)\n\noutput_threshold = os.path.join(\n audit_dir,\n \"cell55_fold1_finetuned_threshold_audit.csv\"\n)\n\nif not os.path.exists(\n finetuned_checkpoint\n):\n raise RuntimeError(\n \"Fine-tuned Fold-1 checkpoint not found: \"\n + finetuned_checkpoint\n )\n\nprint(\n \"Fine-tuned Fold-1 checkpoint: EXISTS\"\n)\n\n# ----------------------------------------------------------------\n# LOAD BEST FINE-TUNED CHECKPOINT\n# ----------------------------------------------------------------\n\ncheckpoint = torch.load(\n finetuned_checkpoint,\n map_location=device\n)\n\nif isinstance(checkpoint, dict):\n\n if \"state_dict\" in checkpoint:\n checkpoint_state = checkpoint[\"state_dict\"]\n\n elif \"model_state_dict\" in checkpoint:\n checkpoint_state = checkpoint[\"model_state_dict\"]\n\n else:\n checkpoint_state = checkpoint\n\nelse:\n raise RuntimeError(\n \"Unsupported fine-tuned checkpoint format.\"\n )\n\nmodel.load_state_dict(\n checkpoint_state,\n strict=True\n)\n\nmodel = model.to(device)\nmodel.eval()\n\nprint(\n \"Fine-tuned Fold-1 checkpoint: LOADED\"\n)\n\nprint(\n \"Model mode: evaluation\"\n)\n\n# ----------------------------------------------------------------\n# COLLECT VALIDATION PREDICTIONS\n# ----------------------------------------------------------------\n\nall_probabilities = []\nall_labels = []\nall_study_ids = []\n\nwith torch.no_grad():\n\n for batch in val_loader_ft:\n\n if not isinstance(batch, dict):\n raise RuntimeError(\n \"Unexpected validation batch type.\"\n )\n\n required_keys = [\n \"image\",\n \"label\",\n \"study_id\"\n ]\n\n missing_keys = [\n key\n for key in required_keys\n if key not in batch\n ]\n\n if missing_keys:\n raise RuntimeError(\n \"Missing validation batch keys: \"\n + \", \".join(missing_keys)\n )\n\n images = batch[\"image\"].to(\n device\n )\n\n labels = batch[\"label\"].to(\n device\n )\n\n logits = model(\n images\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n if tuple(probabilities.shape) != (\n images.shape[0],\n len(TARGETS)\n ):\n raise RuntimeError(\n \"Probability shape mismatch: \"\n + str(tuple(probabilities.shape))\n )\n\n if not torch.isfinite(\n probabilities\n ).all():\n raise RuntimeError(\n \"NaN/Inf detected in probabilities.\"\n )\n\n all_probabilities.append(\n probabilities.cpu()\n )\n\n all_labels.append(\n labels.cpu()\n )\n\n all_study_ids.extend(\n batch[\"study_id\"]\n )\n\nprobabilities = torch.cat(\n all_probabilities,\n dim=0\n).numpy()\n\nlabels = torch.cat(\n all_labels,\n dim=0\n).numpy()\n\nprint(\n f\"\\nProbability shape: {probabilities.shape}\"\n)\n\nprint(\n f\"Label shape: {labels.shape}\"\n)\n\nif probabilities.shape != (\n len(all_study_ids),\n len(TARGETS)\n):\n raise RuntimeError(\n \"Final prediction shape is inconsistent.\"\n )\n\nif labels.shape != probabilities.shape:\n raise RuntimeError(\n \"Probability/label shape mismatch.\"\n )\n\nif len(set(all_study_ids)) != len(\n all_study_ids\n):\n raise RuntimeError(\n \"Duplicate validation study predictions detected.\"\n )\n\nprint(\n \"Probability validity: PASS\"\n)\n\nprint(\n \"Shape consistency: PASS\"\n)\n\nprint(\n \"One prediction per validation study: PASS\"\n)\n\n# ----------------------------------------------------------------\n# TARGET-LEVEL DIAGNOSTICS\n# ----------------------------------------------------------------\n\nrows = []\n\nfor index, target in enumerate(TARGETS):\n\n y_true = labels[:, index]\n y_prob = probabilities[:, index]\n\n positive_mask = (\n y_true == 1\n )\n\n negative_mask = (\n y_true == 0\n )\n\n positive_count = int(\n positive_mask.sum()\n )\n\n negative_count = int(\n negative_mask.sum()\n )\n\n if (\n positive_count == 0\n or negative_count == 0\n ):\n raise RuntimeError(\n f\"Both classes are required for {target}.\"\n )\n\n positive_values = (\n y_prob[positive_mask]\n )\n\n negative_values = (\n y_prob[negative_mask]\n )\n\n positive_mean = float(\n np.mean(positive_values)\n )\n\n negative_mean = float(\n np.mean(negative_values)\n )\n\n positive_median = float(\n np.median(positive_values)\n )\n\n negative_median = float(\n np.median(negative_values)\n )\n\n probability_separation = (\n positive_mean\n - negative_mean\n )\n\n roc_auc = float(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n\n # ------------------------------------------------------------\n # F1 @ 0.50\n # ------------------------------------------------------------\n\n predictions_050 = (\n y_prob >= 0.50\n ).astype(np.int32)\n\n tp = np.sum(\n (y_true == 1)\n & (predictions_050 == 1)\n )\n\n fp = np.sum(\n (y_true == 0)\n & (predictions_050 == 1)\n )\n\n fn = np.sum(\n (y_true == 1)\n & (predictions_050 == 0)\n )\n\n denominator = (\n 2 * tp + fp + fn\n )\n\n if denominator == 0:\n f1_at_050 = 0.0\n else:\n f1_at_050 = (\n 2 * tp\n / denominator\n )\n\n # ------------------------------------------------------------\n # BEST THRESHOLD\n # ------------------------------------------------------------\n\n thresholds = np.arange(\n 0.10,\n 0.9001,\n 0.025\n )\n\n best_threshold = 0.50\n best_f1 = 0.0\n\n for threshold in thresholds:\n\n predictions = (\n y_prob >= threshold\n ).astype(np.int32)\n\n tp_t = np.sum(\n (y_true == 1)\n & (predictions == 1)\n )\n\n fp_t = np.sum(\n (y_true == 0)\n & (predictions == 1)\n )\n\n fn_t = np.sum(\n (y_true == 1)\n & (predictions == 0)\n )\n\n denominator_t = (\n 2 * tp_t\n + fp_t\n + fn_t\n )\n\n if denominator_t == 0:\n f1_t = 0.0\n else:\n f1_t = (\n 2 * tp_t\n / denominator_t\n )\n\n if f1_t > best_f1:\n best_f1 = float(f1_t)\n best_threshold = float(\n threshold\n )\n\n f1_improvement = (\n best_f1\n - f1_at_050\n )\n\n rows.append(\n {\n \"target\": target,\n \"positive_count\": positive_count,\n \"negative_count\": negative_count,\n \"positive_mean\": positive_mean,\n \"negative_mean\": negative_mean,\n \"positive_median\": positive_median,\n \"negative_median\": negative_median,\n \"probability_separation\": probability_separation,\n \"roc_auc\": roc_auc,\n \"f1_at_0.50\": f1_at_050,\n \"best_threshold\": best_threshold,\n \"best_f1\": best_f1,\n \"f1_improvement\": f1_improvement\n }\n )\n\ndiagnostics_df = pd.DataFrame(\n rows\n)\n\n# ----------------------------------------------------------------\n# DISPLAY DIAGNOSTICS\n# ----------------------------------------------------------------\n\nprint(\"\\n\")\nprint(\n diagnostics_df.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# SUMMARY\n# ----------------------------------------------------------------\n\nmean_roc_auc = float(\n diagnostics_df[\"roc_auc\"].mean()\n)\n\nmean_separation = float(\n diagnostics_df[\n \"probability_separation\"\n ].mean()\n)\n\nmean_f1_050 = float(\n diagnostics_df[\n \"f1_at_0.50\"\n ].mean()\n)\n\nmean_best_f1 = float(\n diagnostics_df[\n \"best_f1\"\n ].mean()\n)\n\npotential_f1_improvement = (\n mean_best_f1\n - mean_f1_050\n)\n\nroc_auc_good = diagnostics_df[\n diagnostics_df[\"roc_auc\"] >= 0.65\n][\"target\"].tolist()\n\nnegative_separation = diagnostics_df[\n diagnostics_df[\n \"probability_separation\"\n ] < 0\n][\"target\"].tolist()\n\nlarge_f1_improvement = diagnostics_df[\n diagnostics_df[\n \"f1_improvement\"\n ] > 0.05\n][\"target\"].tolist()\n\nprint(\n f\"\\nMean ROC-AUC: {mean_roc_auc:.4f}\"\n)\n\nprint(\n f\"Mean probability separation: \"\n f\"{mean_separation:.4f}\"\n)\n\nprint(\n f\"Mean F1 @ 0.50: \"\n f\"{mean_f1_050:.4f}\"\n)\n\nprint(\n f\"Mean best-threshold F1: \"\n f\"{mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Potential F1 improvement: \"\n f\"{potential_f1_improvement:.4f}\"\n)\n\nprint(\n f\"\\nTargets with ROC-AUC >= 0.65: \"\n f\"{len(roc_auc_good)}\"\n)\n\nprint(\n roc_auc_good\n)\n\nprint(\n f\"\\nTargets with negative probability separation: \"\n f\"{len(negative_separation)}\"\n)\n\nprint(\n negative_separation\n)\n\nprint(\n f\"\\nTargets with >0.05 F1 improvement: \"\n f\"{len(large_f1_improvement)}\"\n)\n\nprint(\n large_f1_improvement\n)\n\n# ----------------------------------------------------------------\n# SAVE\n# ----------------------------------------------------------------\n\ndiagnostics_df.to_csv(\n output_separation,\n index=False\n)\n\nthreshold_df = diagnostics_df[\n [\n \"target\",\n \"f1_at_0.50\",\n \"best_threshold\",\n \"best_f1\",\n \"f1_improvement\"\n ]\n].copy()\n\nthreshold_df.to_csv(\n output_threshold,\n index=False\n)\n\nprint(\n f\"\\nSaved: {output_separation}\"\n)\n\nprint(\n f\"Saved: {output_threshold}\"\n)\n\nprint(\n \"No training performed.\"\n)\n\nprint(\n \"No original checkpoint modified.\"\n)\n\nprint(\n \"No test data used.\"\n)\n\nprint(\n \"\\nCELL 55 COMPLETE\"\n)\n\nprint(\n \"Send the complete Cell 55 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T10:54:23.707323Z","iopub.execute_input":"2026-08-11T10:54:23.707739Z","iopub.status.idle":"2026-08-11T10:54:35.833341Z","shell.execute_reply.started":"2026-08-11T10:54:23.707706Z","shell.execute_reply":"2026-08-11T10:54:35.832518Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 56R - FOLD 2 FINE-TUNING FINALIZATION\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\n\nprint(\"=\" * 70)\nprint(\"CELL 56R - FOLD 2 FINE-TUNING FINALIZATION\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"TARGETS\",\n \"device\",\n \"best_state_dict\",\n \"best_epoch\",\n \"best_val_loss\",\n \"history\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------\n# VERIFY BEST TRAINING RESULT\n# ----------------------------------------------------------------\n\nif best_epoch != 2:\n raise RuntimeError(\n f\"Unexpected best epoch: {best_epoch}\"\n )\n\nif not np.isfinite(\n float(best_val_loss)\n):\n raise RuntimeError(\n f\"Invalid best validation loss: {best_val_loss}\"\n )\n\nprint(\n f\"Best epoch from completed training: {best_epoch}\"\n)\n\nprint(\n f\"Best validation loss from completed training: \"\n f\"{float(best_val_loss):.6f}\"\n)\n\nprint(\n \"Best training result: PASS\"\n)\n\n# ----------------------------------------------------------------\n# RESTORE BEST MODEL STATE\n# ----------------------------------------------------------------\n\nmodel.load_state_dict(\n best_state_dict,\n strict=True\n)\n\nmodel = model.to(device)\nmodel.eval()\n\nprint(\n \"Best Fold-2 fine-tuned model state restored: PASS\"\n)\n\n# ----------------------------------------------------------------\n# MODEL VALIDITY CHECK\n# ----------------------------------------------------------------\n\nfor parameter in model.parameters():\n\n if not torch.isfinite(\n parameter\n ).all():\n\n raise RuntimeError(\n \"NaN/Inf detected in restored model parameters.\"\n )\n\nprint(\n \"Restored model parameter validity: PASS\"\n)\n\n# ----------------------------------------------------------------\n# FORWARD-PASS CHECK\n# ----------------------------------------------------------------\n\nif \"val_loader\" in globals():\n\n validation_batch = next(\n iter(val_loader)\n )\n\n if not isinstance(\n validation_batch,\n dict\n ):\n raise RuntimeError(\n \"Unexpected validation batch type.\"\n )\n\n required_keys = [\n \"image\",\n \"label\",\n \"study_id\"\n ]\n\n missing_keys = [\n key\n for key in required_keys\n if key not in validation_batch\n ]\n\n if missing_keys:\n raise RuntimeError(\n \"Missing validation batch keys: \"\n + \", \".join(missing_keys)\n )\n\n images = validation_batch[\n \"image\"\n ].to(device)\n\n labels = validation_batch[\n \"label\"\n ].to(device)\n\n with torch.no_grad():\n\n logits = model(\n images\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n if tuple(images.shape[1:]) != (\n 21,\n 224,\n 224\n ):\n raise RuntimeError(\n \"Unexpected input shape: \"\n + str(tuple(images.shape))\n )\n\n if tuple(logits.shape[1:]) != (\n len(TARGETS),\n ):\n raise RuntimeError(\n \"Unexpected output shape: \"\n + str(tuple(logits.shape))\n )\n\n if not torch.isfinite(\n logits\n ).all():\n raise RuntimeError(\n \"NaN/Inf detected in logits.\"\n )\n\n if not torch.isfinite(\n probabilities\n ).all():\n raise RuntimeError(\n \"NaN/Inf detected in probabilities.\"\n )\n\n print(\n \"Forward-pass validation: PASS\"\n )\n\n# ----------------------------------------------------------------\n# SAVE EXPERIMENTAL CHECKPOINT\n# ----------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nfinetuned_checkpoint = os.path.join(\n audit_dir,\n \"cell56_fold2_finetuned_experiment.pt\"\n)\n\nhistory_path = os.path.join(\n audit_dir,\n \"cell56_fold2_finetuned_history.csv\"\n)\n\nmetrics_path = os.path.join(\n audit_dir,\n \"cell56_fold2_finetuned_metrics.csv\"\n)\n\ntorch.save(\n {\n \"state_dict\": model.state_dict(),\n \"fold\": 2,\n \"best_epoch\": int(best_epoch),\n \"best_val_loss\": float(best_val_loss),\n \"trainable_parameters\": 8_593_996,\n \"config\": {\n \"conv1_lr\": 1e-4,\n \"layer4_lr\": 1e-5,\n \"classifier_lr\": 1e-3,\n \"weight_decay\": 1e-4,\n \"gradient_clip\": 1.0,\n \"max_epochs\": 8,\n \"early_stopping_patience\": 3\n }\n },\n finetuned_checkpoint\n)\n\nprint(\n \"Experimental checkpoint saved: PASS\"\n)\n\n# ----------------------------------------------------------------\n# SAVE HISTORY\n# ----------------------------------------------------------------\n\nhistory_df = pd.DataFrame(\n history\n)\n\nhistory_df.to_csv(\n history_path,\n index=False\n)\n\nprint(\n \"History saved: PASS\"\n)\n\n# ----------------------------------------------------------------\n# SAVE METADATA\n# ----------------------------------------------------------------\n\nmetrics_df = pd.DataFrame(\n [\n {\n \"fold\": 2,\n \"best_epoch\": int(best_epoch),\n \"best_val_loss\":\n float(best_val_loss),\n \"trainable_parameters\":\n 8_593_996,\n \"conv1_lr\": 1e-4,\n \"layer4_lr\": 1e-5,\n \"classifier_lr\": 1e-3,\n \"weight_decay\": 1e-4,\n \"gradient_clip\": 1.0,\n \"max_epochs\": 8,\n \"early_stopping_patience\": 3\n }\n ]\n)\n\nmetrics_df.to_csv(\n metrics_path,\n index=False\n)\n\nprint(\n \"Metrics saved: PASS\"\n)\n\n# ----------------------------------------------------------------\n# VERIFY FILES\n# ----------------------------------------------------------------\n\nfor path in [\n finetuned_checkpoint,\n history_path,\n metrics_path\n]:\n\n if not os.path.exists(path):\n raise RuntimeError(\n \"Expected output file was not created: \"\n + path\n )\n\nprint(\n \"Output-file verification: PASS\"\n)\n\n# ----------------------------------------------------------------\n# FINAL VERDICT\n# ----------------------------------------------------------------\n\nprint(\"\\n\" + \"=\" * 70)\nprint(\"CELL 56R COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n f\"Fold: 2\"\n)\n\nprint(\n f\"Best epoch: {best_epoch}\"\n)\n\nprint(\n f\"Best validation loss: \"\n f\"{float(best_val_loss):.6f}\"\n)\n\nprint(\n f\"Checkpoint: {finetuned_checkpoint}\"\n)\n\nprint(\n \"Original Cell-44 Fold-2 checkpoint was NOT overwritten.\"\n)\n\nprint(\n \"No additional training performed.\"\n)\n\nprint(\n \"No test data used.\"\n)\n\nprint(\n \"Fold-2 controlled fine-tuning experiment finalized successfully.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:01:17.217455Z","iopub.execute_input":"2026-08-11T11:01:17.217799Z","iopub.status.idle":"2026-08-11T11:01:18.356082Z","shell.execute_reply.started":"2026-08-11T11:01:17.217771Z","shell.execute_reply":"2026-08-11T11:01:18.355062Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 57 - FOLD 2 FINE-TUNED DIAGNOSTIC\n# ================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom sklearn.metrics import roc_auc_score\n\nprint(\"=\" * 70)\nprint(\"CELL 57 - FOLD 2 FINE-TUNED DIAGNOSTIC\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"val_loader\",\n \"TARGETS\",\n \"device\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------\n# CHECKPOINT\n# ----------------------------------------------------------------\n\ncheckpoint_path = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell56_fold2_finetuned_experiment.pt\"\n)\n\nif not os.path.exists(\n checkpoint_path\n):\n raise RuntimeError(\n \"Fine-tuned Fold-2 checkpoint not found: \"\n + checkpoint_path\n )\n\nprint(\n \"Fine-tuned Fold-2 checkpoint: EXISTS\"\n)\n\n# ----------------------------------------------------------------\n# LOAD CHECKPOINT\n# ----------------------------------------------------------------\n\ncheckpoint = torch.load(\n checkpoint_path,\n map_location=device\n)\n\nif not isinstance(\n checkpoint,\n dict\n):\n raise RuntimeError(\n \"Unexpected checkpoint format.\"\n )\n\nif \"state_dict\" not in checkpoint:\n raise RuntimeError(\n \"Checkpoint does not contain state_dict.\"\n )\n\nmodel.load_state_dict(\n checkpoint[\"state_dict\"],\n strict=True\n)\n\nmodel = model.to(device)\nmodel.eval()\n\nprint(\n \"Fine-tuned Fold-2 checkpoint: LOADED\"\n)\n\nprint(\n \"Model mode: evaluation\"\n)\n\n# ----------------------------------------------------------------\n# COLLECT VALIDATION PREDICTIONS\n# ----------------------------------------------------------------\n\nprobability_batches = []\nlabel_batches = []\nstudy_ids = []\n\nwith torch.no_grad():\n\n for batch in val_loader:\n\n if not isinstance(\n batch,\n dict\n ):\n raise RuntimeError(\n \"Unexpected validation batch type.\"\n )\n\n required_keys = [\n \"image\",\n \"label\",\n \"study_id\"\n ]\n\n missing_keys = [\n key\n for key in required_keys\n if key not in batch\n ]\n\n if missing_keys:\n raise RuntimeError(\n \"Missing validation batch keys: \"\n + \", \".join(missing_keys)\n )\n\n images = batch[\n \"image\"\n ].to(device)\n\n labels = batch[\n \"label\"\n ]\n\n logits = model(\n images\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n if not torch.isfinite(\n probabilities\n ).all():\n raise RuntimeError(\n \"NaN/Inf detected in probabilities.\"\n )\n\n probability_batches.append(\n probabilities.cpu().numpy()\n )\n\n label_batches.append(\n labels.cpu().numpy()\n )\n\n study_ids.extend(\n list(batch[\"study_id\"])\n )\n\nprobabilities = np.concatenate(\n probability_batches,\n axis=0\n)\n\nlabels = np.concatenate(\n label_batches,\n axis=0\n)\n\n# ----------------------------------------------------------------\n# SHAPE VALIDATION\n# ----------------------------------------------------------------\n\nif probabilities.shape != (\n len(val_loader.dataset),\n len(TARGETS)\n):\n raise RuntimeError(\n \"Probability shape mismatch: \"\n + str(probabilities.shape)\n )\n\nif labels.shape != (\n len(val_loader.dataset),\n len(TARGETS)\n):\n raise RuntimeError(\n \"Label shape mismatch: \"\n + str(labels.shape)\n )\n\nif len(study_ids) != len(\n val_loader.dataset\n):\n raise RuntimeError(\n \"Study ID count does not match validation dataset.\"\n )\n\nif len(set(study_ids)) != len(\n study_ids\n):\n raise RuntimeError(\n \"Duplicate study IDs detected.\"\n )\n\nprint(\n f\"Probability shape: {probabilities.shape}\"\n)\n\nprint(\n f\"Label shape: {labels.shape}\"\n)\n\nprint(\n \"Probability validity: PASS\"\n)\n\nprint(\n \"Shape consistency: PASS\"\n)\n\nprint(\n \"One prediction per validation study: PASS\"\n)\n\n# ----------------------------------------------------------------\n# TARGET DIAGNOSTICS\n# ----------------------------------------------------------------\n\ndiagnostics = []\nthreshold_diagnostics = []\n\nthreshold_grid = np.arange(\n 0.10,\n 0.9001,\n 0.025\n)\n\nfor index, target in enumerate(\n TARGETS\n):\n\n y_true = labels[\n :,\n index\n ].astype(int)\n\n y_prob = probabilities[\n :,\n index\n ]\n\n positive_mask = (\n y_true == 1\n )\n\n negative_mask = (\n y_true == 0\n )\n\n positive_count = int(\n positive_mask.sum()\n )\n\n negative_count = int(\n negative_mask.sum()\n )\n\n positive_mean = float(\n y_prob[\n positive_mask\n ].mean()\n )\n\n negative_mean = float(\n y_prob[\n negative_mask\n ].mean()\n )\n\n positive_median = float(\n np.median(\n y_prob[\n positive_mask\n ]\n )\n )\n\n negative_median = float(\n np.median(\n y_prob[\n negative_mask\n ]\n )\n )\n\n probability_separation = (\n positive_mean\n - negative_mean\n )\n\n if (\n positive_count > 0\n and negative_count > 0\n ):\n\n roc_auc = float(\n roc_auc_score(\n y_true,\n y_prob\n )\n )\n\n else:\n\n roc_auc = np.nan\n\n # ------------------------------------------------------------\n # F1 AT 0.50\n # ------------------------------------------------------------\n\n y_pred_050 = (\n y_prob >= 0.50\n ).astype(int)\n\n tp = int(\n np.sum(\n (y_true == 1)\n & (y_pred_050 == 1)\n )\n )\n\n fp = int(\n np.sum(\n (y_true == 0)\n & (y_pred_050 == 1)\n )\n )\n\n fn = int(\n np.sum(\n (y_true == 1)\n & (y_pred_050 == 0)\n )\n )\n\n f1_denominator = (\n 2 * tp\n + fp\n + fn\n )\n\n if f1_denominator == 0:\n f1_at_050 = 0.0\n else:\n f1_at_050 = (\n 2 * tp\n / f1_denominator\n )\n\n # ------------------------------------------------------------\n # BEST THRESHOLD\n # ------------------------------------------------------------\n\n best_threshold = 0.50\n best_f1 = f1_at_050\n\n for threshold in threshold_grid:\n\n y_pred = (\n y_prob >= threshold\n ).astype(int)\n\n tp_t = int(\n np.sum(\n (y_true == 1)\n & (y_pred == 1)\n )\n )\n\n fp_t = int(\n np.sum(\n (y_true == 0)\n & (y_pred == 1)\n )\n )\n\n fn_t = int(\n np.sum(\n (y_true == 1)\n & (y_pred == 0)\n )\n )\n\n denominator = (\n 2 * tp_t\n + fp_t\n + fn_t\n )\n\n if denominator == 0:\n f1_t = 0.0\n else:\n f1_t = (\n 2 * tp_t\n / denominator\n )\n\n if f1_t > best_f1:\n best_f1 = f1_t\n best_threshold = float(\n threshold\n )\n\n f1_improvement = (\n best_f1\n - f1_at_050\n )\n\n diagnostics.append(\n {\n \"target\": target,\n \"positive_count\":\n positive_count,\n \"negative_count\":\n negative_count,\n \"positive_mean\":\n positive_mean,\n \"negative_mean\":\n negative_mean,\n \"positive_median\":\n positive_median,\n \"negative_median\":\n negative_median,\n \"probability_separation\":\n probability_separation,\n \"roc_auc\":\n roc_auc,\n \"f1_at_0.50\":\n f1_at_050,\n \"best_threshold\":\n best_threshold,\n \"best_f1\":\n best_f1,\n \"f1_improvement\":\n f1_improvement\n }\n )\n\n threshold_diagnostics.append(\n {\n \"target\": target,\n \"f1_at_0.50\":\n f1_at_050,\n \"best_threshold\":\n best_threshold,\n \"best_f1\":\n best_f1,\n \"f1_improvement\":\n f1_improvement\n }\n )\n\ndiagnostics_df = pd.DataFrame(\n diagnostics\n)\n\nthreshold_df = pd.DataFrame(\n threshold_diagnostics\n)\n\n# ----------------------------------------------------------------\n# DISPLAY\n# ----------------------------------------------------------------\n\nprint()\n\nprint(\n diagnostics_df.to_string(\n index=False,\n float_format=lambda x:\n f\"{x:.4f}\"\n )\n)\n\nprint()\n\nprint(\n threshold_df.to_string(\n index=False,\n float_format=lambda x:\n f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------\n# SUMMARY\n# ----------------------------------------------------------------\n\nmean_roc_auc = float(\n diagnostics_df[\n \"roc_auc\"\n ].mean()\n)\n\nmean_separation = float(\n diagnostics_df[\n \"probability_separation\"\n ].mean()\n)\n\nmean_f1_050 = float(\n diagnostics_df[\n \"f1_at_0.50\"\n ].mean()\n)\n\nmean_best_f1 = float(\n diagnostics_df[\n \"best_f1\"\n ].mean()\n)\n\npotential_improvement = (\n mean_best_f1\n - mean_f1_050\n)\n\nhigh_auc_targets = diagnostics_df.loc[\n diagnostics_df[\"roc_auc\"] >= 0.65,\n \"target\"\n].tolist()\n\nnegative_separation_targets = diagnostics_df.loc[\n diagnostics_df[\n \"probability_separation\"\n ] < 0,\n \"target\"\n].tolist()\n\nimproved_targets = diagnostics_df.loc[\n diagnostics_df[\n \"f1_improvement\"\n ] > 0.05,\n \"target\"\n].tolist()\n\nprint()\n\nprint(\n f\"Mean ROC-AUC: {mean_roc_auc:.4f}\"\n)\n\nprint(\n f\"Mean probability separation: \"\n f\"{mean_separation:.4f}\"\n)\n\nprint(\n f\"Mean F1 @ 0.50: {mean_f1_050:.4f}\"\n)\n\nprint(\n f\"Mean best-threshold F1: \"\n f\"{mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Potential F1 improvement: \"\n f\"{potential_improvement:.4f}\"\n)\n\nprint()\n\nprint(\n f\"Targets with ROC-AUC >= 0.65: \"\n f\"{len(high_auc_targets)}\"\n)\n\nprint(\n high_auc_targets\n)\n\nprint()\n\nprint(\n f\"Targets with negative probability separation: \"\n f\"{len(negative_separation_targets)}\"\n)\n\nprint(\n negative_separation_targets\n)\n\nprint()\n\nprint(\n f\"Targets with >0.05 F1 improvement: \"\n f\"{len(improved_targets)}\"\n)\n\nprint(\n improved_targets\n)\n\n# ----------------------------------------------------------------\n# SAVE\n# ----------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\ndiagnostics_path = os.path.join(\n audit_dir,\n \"cell57_fold2_finetuned_probability_separation.csv\"\n)\n\nthreshold_path = os.path.join(\n audit_dir,\n \"cell57_fold2_finetuned_threshold_audit.csv\"\n)\n\ndiagnostics_df.to_csv(\n diagnostics_path,\n index=False\n)\n\nthreshold_df.to_csv(\n threshold_path,\n index=False\n)\n\nprint()\n\nprint(\n f\"Saved: {diagnostics_path}\"\n)\n\nprint(\n f\"Saved: {threshold_path}\"\n)\n\nprint(\n \"No training performed.\"\n)\n\nprint(\n \"No original checkpoint modified.\"\n)\n\nprint(\n \"No test data used.\"\n)\n\nprint()\n\nprint(\"=\" * 70)\nprint(\"CELL 57 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Send the complete Cell 57 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:02:14.728607Z","iopub.execute_input":"2026-08-11T11:02:14.729Z","iopub.status.idle":"2026-08-11T11:02:23.934557Z","shell.execute_reply.started":"2026-08-11T11:02:14.728968Z","shell.execute_reply":"2026-08-11T11:02:23.933615Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ================================================================\n# CELL 58A - FINE-TUNED CHECKPOINT FORMAT AUDIT\n# ================================================================\n\nimport os\nimport torch\n\nprint(\"=\" * 70)\nprint(\"CELL 58A - FINE-TUNED CHECKPOINT FORMAT AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"TARGETS\",\n \"device\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------\n# CHECKPOINT PATHS\n# ----------------------------------------------------------------\n\ncheckpoint_paths = {\n 0: \"/kaggle/working/rsna_knee_audit/\"\n \"cell52_fold0_finetuned_experiment.pt\",\n\n 1: \"/kaggle/working/rsna_knee_audit/\"\n \"cell54_fold1_finetuned_experiment.pt\",\n\n 2: \"/kaggle/working/rsna_knee_audit/\"\n \"cell56_fold2_finetuned_experiment.pt\"\n}\n\nprint()\nprint(\"=\" * 70)\nprint(\"CHECKPOINT FILE VALIDATION\")\nprint(\"=\" * 70)\n\nfor fold, path in checkpoint_paths.items():\n\n exists = os.path.exists(path)\n\n print(\n f\"Fold-{fold} checkpoint: \"\n f\"{'EXISTS' if exists else 'MISSING'}\"\n )\n\n if not exists:\n raise RuntimeError(\n f\"Missing Fold-{fold} fine-tuned checkpoint: \"\n + path\n )\n\nprint(\"All three fine-tuned checkpoints: PASS\")\n\n# ----------------------------------------------------------------\n# MODEL ARCHITECTURE VALIDATION\n# ----------------------------------------------------------------\n\nif not isinstance(\n model,\n torch.nn.Module\n):\n raise RuntimeError(\n \"Existing model is not a torch.nn.Module.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"EXISTING MODEL VALIDATION\")\nprint(\"=\" * 70)\n\nprint(\n f\"Model type: {type(model)}\"\n)\n\nprint(\n f\"Model device: \"\n f\"{next(model.parameters()).device}\"\n)\n\n# ----------------------------------------------------------------\n# CHECKPOINT FORMAT INSPECTION\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CHECKPOINT FORMAT INSPECTION\")\nprint(\"=\" * 70)\n\ncheckpoint_audit = {}\n\nfor fold, path in checkpoint_paths.items():\n\n print()\n print(\n f\"--- Fold {fold} ---\"\n )\n\n checkpoint = torch.load(\n path,\n map_location=\"cpu\",\n weights_only=False\n )\n\n checkpoint_type = type(\n checkpoint\n )\n\n print(\n f\"Object type: {checkpoint_type}\"\n )\n\n if isinstance(\n checkpoint,\n dict\n ):\n\n print(\n f\"Dictionary keys: \"\n f\"{list(checkpoint.keys())}\"\n )\n\n for key, value in checkpoint.items():\n\n print(\n f\" {key}: \"\n f\"type={type(value)}\"\n )\n\n if isinstance(\n value,\n dict\n ):\n print(\n f\" nested keys=\"\n f\"{list(value.keys())[:20]}\"\n )\n\n elif isinstance(\n checkpoint,\n torch.nn.Module\n ):\n\n print(\n \"Checkpoint contains a complete \"\n \"PyTorch model object.\"\n )\n\n else:\n\n print(\n \"Checkpoint is neither a standard \"\n \"dictionary nor a torch.nn.Module.\"\n )\n\n checkpoint_audit[fold] = checkpoint\n\n# ----------------------------------------------------------------\n# DETERMINE LOADABLE STATE FORMAT WITHOUT ASSUMPTIONS\n# ----------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CHECKPOINT LOAD FORMAT DETECTION\")\nprint(\"=\" * 70)\n\ndef detect_checkpoint_format(\n checkpoint\n):\n\n # Direct state_dict:\n # dictionary whose values are tensors/parameters.\n if isinstance(\n checkpoint,\n dict\n ):\n\n if len(checkpoint) > 0:\n\n tensor_values = all(\n isinstance(\n value,\n (\n torch.Tensor,\n torch.nn.Parameter\n )\n )\n for value in checkpoint.values()\n )\n\n if tensor_values:\n return \"direct_state_dict\"\n\n # Common wrapped formats.\n for key in [\n \"state_dict\",\n \"model_state_dict\",\n \"model\"\n ]:\n\n if key in checkpoint:\n\n value = checkpoint[key]\n\n if isinstance(\n value,\n dict\n ):\n\n if len(value) == 0:\n continue\n\n tensor_values = all(\n isinstance(\n item,\n (\n torch.Tensor,\n torch.nn.Parameter\n )\n )\n for item in value.values()\n )\n\n if tensor_values:\n return (\n f\"wrapped_state_dict:{key}\"\n )\n\n if isinstance(\n value,\n torch.nn.Module\n ):\n return (\n f\"wrapped_model:{key}\"\n )\n\n if isinstance(\n checkpoint,\n torch.nn.Module\n ):\n return \"complete_model\"\n\n return \"unknown\"\n\ndetected_formats = {}\n\nfor fold, checkpoint in checkpoint_audit.items():\n\n fmt = detect_checkpoint_format(\n checkpoint\n )\n\n detected_formats[fold] = fmt\n\n print(\n f\"Fold-{fold}: {fmt}\"\n )\n\n if fmt == \"unknown\":\n\n raise RuntimeError(\n f\"Unable to safely determine \"\n f\"Fold-{fold} checkpoint format. \"\n f\"Do not continue.\"\n )\n\n# ----------------------------------------------------------------\n# SAVE AUDIT IN NOTEBOOK STATE\n# ----------------------------------------------------------------\n\nfine_tuned_checkpoint_audit = {\n \"paths\": checkpoint_paths,\n \"formats\": detected_formats\n}\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 58A VERDICT\")\nprint(\"=\" * 70)\n\nprint(\n \"All three checkpoint files: PASS\"\n)\n\nprint(\n \"Checkpoint formats identified: PASS\"\n)\n\nfor fold in [0, 1, 2]:\n\n print(\n f\"Fold-{fold} format: \"\n f\"{detected_formats[fold]}\"\n )\n\nprint()\nprint(\n \"No training performed.\"\n)\n\nprint(\n \"No checkpoint modified.\"\n)\n\nprint(\n \"No dataset modified.\"\n)\n\nprint(\n \"No validation predictions generated.\"\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 58A COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Send the complete Cell 58A output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:08:54.287484Z","iopub.execute_input":"2026-08-11T11:08:54.287833Z","iopub.status.idle":"2026-08-11T11:08:54.392823Z","shell.execute_reply.started":"2026-08-11T11:08:54.2878Z","shell.execute_reply":"2026-08-11T11:08:54.391881Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 58R - FINE-TUNED 3-FOLD OOF EVALUATION\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom torch.utils.data import DataLoader\n\nprint(\"=\" * 70)\nprint(\"CELL 58R - FINE-TUNED 3-FOLD OOF EVALUATION\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"TARGETS\",\n \"KneeStudyDataset\",\n \"modeling_table\",\n \"study_folds\",\n \"device\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(f\"Device: {device}\")\n\n# ----------------------------------------------------------------------\n# BASIC VALIDATION\n# ----------------------------------------------------------------------\n\nif not isinstance(model, torch.nn.Module):\n raise RuntimeError(\n \"Existing model is not a torch.nn.Module.\"\n )\n\nif len(TARGETS) != 12:\n raise RuntimeError(\n f\"Expected 12 targets, found {len(TARGETS)}.\"\n )\n\nrequired_primary_columns = [\n \"Sagittal_SeriesInstanceUID\",\n \"Coronal_SeriesInstanceUID\",\n \"Axial_SeriesInstanceUID\"\n]\n\nmissing_primary_columns = [\n col\n for col in required_primary_columns\n if col not in modeling_table.columns\n]\n\nif missing_primary_columns:\n raise RuntimeError(\n \"Missing primary-series columns: \"\n + \", \".join(missing_primary_columns)\n )\n\nif \"StudyInstanceUID\" not in modeling_table.columns:\n raise RuntimeError(\n \"Missing StudyInstanceUID in modeling_table.\"\n )\n\nif \"fold\" not in study_folds.columns:\n raise RuntimeError(\n \"Missing fold column in study_folds.\"\n )\n\nprint(f\"Modeling table shape: {modeling_table.shape}\")\nprint(f\"Fold table shape: {study_folds.shape}\")\nprint(\"Target columns: PASS\")\nprint(\"Primary-series columns: PASS\")\n\n# ----------------------------------------------------------------------\n# BUILD / VERIFY FOLD ASSIGNMENT\n# ----------------------------------------------------------------------\n\nmodeling_ids = set(\n modeling_table[\"StudyInstanceUID\"].astype(str)\n)\n\nfold_ids = set(\n study_folds[\"StudyInstanceUID\"].astype(str)\n)\n\nif modeling_ids != fold_ids:\n missing_fold_ids = modeling_ids - fold_ids\n extra_fold_ids = fold_ids - modeling_ids\n\n raise RuntimeError(\n \"Modeling/fold study IDs do not match. \"\n f\"Missing in fold table={len(missing_fold_ids)}, \"\n f\"extra in fold table={len(extra_fold_ids)}.\"\n )\n\nfold_assignment = study_folds[\n [\"StudyInstanceUID\", \"fold\"]\n].copy()\n\nfold_assignment[\"StudyInstanceUID\"] = (\n fold_assignment[\"StudyInstanceUID\"].astype(str)\n)\n\nfold_assignment[\"fold\"] = (\n fold_assignment[\"fold\"].astype(int)\n)\n\nif fold_assignment[\"StudyInstanceUID\"].duplicated().any():\n raise RuntimeError(\n \"Duplicate StudyInstanceUID values in fold table.\"\n )\n\noof_table_base = modeling_table.copy()\n\noof_table_base[\"StudyInstanceUID\"] = (\n oof_table_base[\"StudyInstanceUID\"].astype(str)\n)\n\nif \"fold\" in oof_table_base.columns:\n oof_table_base = oof_table_base.drop(\n columns=[\"fold\"]\n )\n\noof_table_base = oof_table_base.merge(\n fold_assignment,\n on=\"StudyInstanceUID\",\n how=\"left\",\n validate=\"one_to_one\"\n)\n\nif oof_table_base[\"fold\"].isna().any():\n raise RuntimeError(\n \"Some modeling studies have no fold assignment.\"\n )\n\noof_table_base[\"fold\"] = (\n oof_table_base[\"fold\"].astype(int)\n)\n\nprint(\"Fold assignment merge: PASS\")\nprint(\n f\"Total modeling studies: \"\n f\"{len(oof_table_base)}\"\n)\n\nprint(\n f\"Unique studies: \"\n f\"{oof_table_base['StudyInstanceUID'].nunique()}\"\n)\n\navailable_folds = sorted(\n oof_table_base[\"fold\"].unique().tolist()\n)\n\nprint(\n f\"Available folds: {available_folds}\"\n)\n\nif available_folds != [0, 1, 2]:\n raise RuntimeError(\n \"Expected exactly folds [0, 1, 2]. \"\n f\"Found {available_folds}.\"\n )\n\nprint(\"Three-fold configuration: PASS\")\n\n# ----------------------------------------------------------------------\n# FINE-TUNED CHECKPOINTS\n# ----------------------------------------------------------------------\n\ncheckpoint_paths = {\n 0: \"/kaggle/working/rsna_knee_audit/\"\n \"cell52_fold0_finetuned_experiment.pt\",\n\n 1: \"/kaggle/working/rsna_knee_audit/\"\n \"cell54_fold1_finetuned_experiment.pt\",\n\n 2: \"/kaggle/working/rsna_knee_audit/\"\n \"cell56_fold2_finetuned_experiment.pt\"\n}\n\nfor fold, path in checkpoint_paths.items():\n\n if not os.path.exists(path):\n raise RuntimeError(\n f\"Missing fine-tuned Fold-{fold} checkpoint: \"\n + path\n )\n\n print(\n f\"Fine-tuned Fold-{fold} checkpoint: EXISTS\"\n )\n\nprint(\"All three fine-tuned checkpoints: PASS\")\n\n# ----------------------------------------------------------------------\n# CHECKPOINT STATE-DICT EXTRACTION\n# ----------------------------------------------------------------------\n\ndef extract_state_dict(checkpoint, fold):\n\n # Fold 0 and Fold 1 were verified as direct state_dict objects.\n if isinstance(\n checkpoint,\n dict\n ):\n\n # Direct state_dict:\n # every value is a tensor/parameter.\n if len(checkpoint) > 0:\n\n direct_state_dict = all(\n isinstance(\n value,\n (\n torch.Tensor,\n torch.nn.Parameter\n )\n )\n for value in checkpoint.values()\n )\n\n if direct_state_dict:\n return checkpoint\n\n # Fold 2 was verified as a wrapped state_dict.\n if \"state_dict\" in checkpoint:\n\n state_dict = checkpoint[\"state_dict\"]\n\n if not isinstance(\n state_dict,\n dict\n ):\n raise RuntimeError(\n f\"Fold-{fold} state_dict is not a dictionary.\"\n )\n\n return state_dict\n\n raise RuntimeError(\n f\"Unable to extract state_dict from \"\n f\"Fold-{fold} checkpoint.\"\n )\n\n# ----------------------------------------------------------------------\n# MODEL STATE VALIDATION\n# ----------------------------------------------------------------------\n\nreference_model_state = model.state_dict()\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINE-TUNED CHECKPOINT COMPATIBILITY\")\nprint(\"=\" * 70)\n\nfor fold, path in checkpoint_paths.items():\n\n checkpoint = torch.load(\n path,\n map_location=\"cpu\",\n weights_only=False\n )\n\n state_dict = extract_state_dict(\n checkpoint,\n fold\n )\n\n checkpoint_keys = set(\n state_dict.keys()\n )\n\n model_keys = set(\n reference_model_state.keys()\n )\n\n missing_keys = model_keys - checkpoint_keys\n unexpected_keys = checkpoint_keys - model_keys\n\n if missing_keys:\n raise RuntimeError(\n f\"Fold-{fold} checkpoint missing model keys: \"\n + \", \".join(\n sorted(missing_keys)[:10]\n )\n )\n\n if unexpected_keys:\n raise RuntimeError(\n f\"Fold-{fold} checkpoint has unexpected \"\n f\"model keys: \"\n + \", \".join(\n sorted(unexpected_keys)[:10]\n )\n )\n\n shape_mismatches = []\n\n for key in model_keys:\n\n if (\n tuple(\n state_dict[key].shape\n )\n !=\n tuple(\n reference_model_state[key].shape\n )\n ):\n shape_mismatches.append(key)\n\n if shape_mismatches:\n raise RuntimeError(\n f\"Fold-{fold} checkpoint has parameter \"\n f\"shape mismatches: \"\n + \", \".join(\n shape_mismatches[:10]\n )\n )\n\n print(\n f\"Fold-{fold} state_dict compatibility: PASS\"\n )\n\n# ----------------------------------------------------------------------\n# OOF INFERENCE\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINE-TUNED OOF INFERENCE\")\nprint(\"=\" * 70)\n\noof_prediction_rows = []\n\nfor fold in [0, 1, 2]:\n\n print()\n print(\n f\"FINE-TUNED FOLD {fold} OOF INFERENCE\"\n )\n print(\"-\" * 70)\n\n fold_validation_table = oof_table_base[\n oof_table_base[\"fold\"] == fold\n ].copy()\n\n fold_training_table = oof_table_base[\n oof_table_base[\"fold\"] != fold\n ].copy()\n\n training_ids = set(\n fold_training_table[\n \"StudyInstanceUID\"\n ].astype(str)\n )\n\n validation_ids = set(\n fold_validation_table[\n \"StudyInstanceUID\"\n ].astype(str)\n )\n\n overlap = training_ids & validation_ids\n\n print(\n f\"Training studies: \"\n f\"{len(training_ids)}\"\n )\n\n print(\n f\"Validation studies: \"\n f\"{len(validation_ids)}\"\n )\n\n print(\n f\"Train/validation overlap: \"\n f\"{len(overlap)}\"\n )\n\n if overlap:\n raise RuntimeError(\n f\"Fold-{fold} has train/validation \"\n f\"study overlap.\"\n )\n\n # --------------------------------------------------------------\n # EXACT EXISTING DATASET PIPELINE\n # --------------------------------------------------------------\n\n validation_dataset = KneeStudyDataset(\n fold_validation_table,\n targets=TARGETS\n )\n\n validation_loader = DataLoader(\n validation_dataset,\n batch_size=2,\n shuffle=False,\n num_workers=0,\n pin_memory=False\n )\n\n print(\n f\"Validation dataset: \"\n f\"{len(validation_dataset)}\"\n )\n\n print(\n f\"Validation batches: \"\n f\"{len(validation_loader)}\"\n )\n\n if len(validation_dataset) != len(\n fold_validation_table\n ):\n raise RuntimeError(\n f\"Fold-{fold} validation dataset size \"\n f\"does not match validation table.\"\n )\n\n # --------------------------------------------------------------\n # LOAD CORRECT FINE-TUNED CHECKPOINT\n # --------------------------------------------------------------\n\n checkpoint = torch.load(\n checkpoint_paths[fold],\n map_location=\"cpu\",\n weights_only=False\n )\n\n state_dict = extract_state_dict(\n checkpoint,\n fold\n )\n\n model.load_state_dict(\n state_dict,\n strict=True\n )\n\n model = model.to(device)\n model.eval()\n\n print(\n f\"Fine-tuned Fold-{fold} checkpoint: LOADED\"\n )\n\n # --------------------------------------------------------------\n # INFERENCE\n # --------------------------------------------------------------\n\n fold_prediction_count = 0\n fold_study_ids = []\n\n with torch.no_grad():\n\n for batch in validation_loader:\n\n if not isinstance(\n batch,\n dict\n ):\n raise RuntimeError(\n f\"Fold-{fold} batch is not a dictionary.\"\n )\n\n required_batch_keys = [\n \"image\",\n \"label\",\n \"study_id\"\n ]\n\n missing_batch_keys = [\n key\n for key in required_batch_keys\n if key not in batch\n ]\n\n if missing_batch_keys:\n raise RuntimeError(\n f\"Fold-{fold} batch missing keys: \"\n + \", \".join(\n missing_batch_keys\n )\n )\n\n images = batch[\"image\"].to(\n device,\n non_blocking=True\n )\n\n labels = batch[\"label\"]\n\n study_ids = batch[\"study_id\"]\n\n if images.ndim != 4:\n raise RuntimeError(\n f\"Unexpected image shape in \"\n f\"Fold-{fold}: \"\n f\"{tuple(images.shape)}\"\n )\n\n if images.shape[1:] != (\n 21,\n 224,\n 224\n ):\n raise RuntimeError(\n f\"Unexpected image shape in \"\n f\"Fold-{fold}: \"\n f\"{tuple(images.shape)}\"\n )\n\n if labels.ndim != 2:\n raise RuntimeError(\n f\"Unexpected label shape in \"\n f\"Fold-{fold}: \"\n f\"{tuple(labels.shape)}\"\n )\n\n if labels.shape[1] != len(TARGETS):\n raise RuntimeError(\n f\"Unexpected label target count \"\n f\"in Fold-{fold}: \"\n f\"{labels.shape[1]}\"\n )\n\n logits = model(images)\n\n if logits.ndim != 2:\n raise RuntimeError(\n f\"Unexpected logit shape in \"\n f\"Fold-{fold}: \"\n f\"{tuple(logits.shape)}\"\n )\n\n if logits.shape[1] != len(TARGETS):\n raise RuntimeError(\n f\"Expected {len(TARGETS)} outputs, \"\n f\"got {logits.shape[1]}.\"\n )\n\n probabilities = torch.sigmoid(\n logits\n ).detach().cpu().numpy()\n\n labels_np = (\n labels.detach()\n .cpu()\n .numpy()\n )\n\n probabilities = np.asarray(\n probabilities,\n dtype=np.float64\n )\n\n labels_np = np.asarray(\n labels_np,\n dtype=np.float64\n )\n\n if not np.isfinite(\n probabilities\n ).all():\n raise RuntimeError(\n f\"NaN/Inf probabilities in \"\n f\"Fold-{fold}.\"\n )\n\n if not np.isfinite(\n labels_np\n ).all():\n raise RuntimeError(\n f\"NaN/Inf labels in \"\n f\"Fold-{fold}.\"\n )\n\n if (\n probabilities.shape\n != labels_np.shape\n ):\n raise RuntimeError(\n f\"Probability/label shape mismatch \"\n f\"in Fold-{fold}: \"\n f\"{probabilities.shape} vs \"\n f\"{labels_np.shape}\"\n )\n\n if len(study_ids) != len(\n probabilities\n ):\n raise RuntimeError(\n f\"Study ID count does not match \"\n f\"prediction count in Fold-{fold}.\"\n )\n\n for row_idx, study_id in enumerate(\n study_ids\n ):\n\n study_id = str(\n study_id\n )\n\n row = {\n \"StudyInstanceUID\": study_id,\n \"fold\": fold\n }\n\n for target_idx, target in enumerate(\n TARGETS\n ):\n\n row[\n f\"{target}_prob\"\n ] = float(\n probabilities[\n row_idx,\n target_idx\n ]\n )\n\n row[\n target\n ] = float(\n labels_np[\n row_idx,\n target_idx\n ]\n )\n\n oof_prediction_rows.append(\n row\n )\n\n fold_study_ids.append(\n study_id\n )\n\n fold_prediction_count += len(\n probabilities\n )\n\n # --------------------------------------------------------------\n # FOLD COVERAGE VALIDATION\n # --------------------------------------------------------------\n\n expected_ids = set(\n fold_validation_table[\n \"StudyInstanceUID\"\n ].astype(str)\n )\n\n predicted_ids = set(\n fold_study_ids\n )\n\n if expected_ids != predicted_ids:\n\n missing_predictions = (\n expected_ids - predicted_ids\n )\n\n unexpected_predictions = (\n predicted_ids - expected_ids\n )\n\n raise RuntimeError(\n f\"Fold-{fold} OOF coverage mismatch. \"\n f\"Missing={len(missing_predictions)}, \"\n f\"Unexpected={len(unexpected_predictions)}.\"\n )\n\n if len(fold_study_ids) != len(\n expected_ids\n ):\n raise RuntimeError(\n f\"Fold-{fold} contains duplicate \"\n f\"validation predictions.\"\n )\n\n print(\n f\"Fold-{fold} predictions: \"\n f\"{fold_prediction_count}\"\n )\n\n print(\n f\"Fold-{fold} coverage: PASS\"\n )\n\n# ----------------------------------------------------------------------\n# BUILD FINAL OOF TABLE\n# ----------------------------------------------------------------------\n\nfine_tuned_oof = pd.DataFrame(\n oof_prediction_rows\n)\n\nif fine_tuned_oof.empty:\n raise RuntimeError(\n \"Fine-tuned OOF prediction table is empty.\"\n )\n\nexpected_oof_rows = len(\n oof_table_base\n)\n\nif len(fine_tuned_oof) != expected_oof_rows:\n raise RuntimeError(\n f\"Expected {expected_oof_rows} OOF rows, \"\n f\"got {len(fine_tuned_oof)}.\"\n )\n\nif (\n fine_tuned_oof[\"StudyInstanceUID\"]\n .nunique()\n != expected_oof_rows\n):\n raise RuntimeError(\n \"Fine-tuned OOF contains duplicate \"\n \"StudyInstanceUID values.\"\n )\n\nif (\n fine_tuned_oof[\"fold\"]\n .value_counts()\n .sort_index()\n .to_dict()\n != {0: 20, 1: 19, 2: 19}\n):\n raise RuntimeError(\n \"Unexpected fine-tuned OOF fold distribution.\"\n )\n\n# ----------------------------------------------------------------------\n# VALIDATE PROBABILITIES\n# ----------------------------------------------------------------------\n\nprobability_columns = [\n f\"{target}_prob\"\n for target in TARGETS\n]\n\nprobability_matrix = (\n fine_tuned_oof[\n probability_columns\n ]\n .to_numpy(\n dtype=np.float64\n )\n)\n\nlabel_matrix = (\n fine_tuned_oof[\n TARGETS\n ]\n .to_numpy(\n dtype=np.float64\n )\n)\n\nif not np.isfinite(\n probability_matrix\n).all():\n raise RuntimeError(\n \"Fine-tuned OOF contains NaN/Inf probabilities.\"\n )\n\nif not np.isfinite(\n label_matrix\n).all():\n raise RuntimeError(\n \"Fine-tuned OOF contains NaN/Inf labels.\"\n )\n\nif (\n probability_matrix.min() < 0.0\n or probability_matrix.max() > 1.0\n):\n raise RuntimeError(\n \"Fine-tuned OOF probabilities are outside [0, 1].\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINE-TUNED OOF VALIDATION\")\nprint(\"=\" * 70)\n\nprint(\n f\"OOF shape: {fine_tuned_oof.shape}\"\n)\n\nprint(\n f\"OOF rows: {len(fine_tuned_oof)}\"\n)\n\nprint(\n f\"Unique OOF studies: \"\n f\"{fine_tuned_oof['StudyInstanceUID'].nunique()}\"\n)\n\nprint(\n \"One prediction per study: PASS\"\n)\n\nprint()\nprint(\"OOF fold distribution:\")\n\nprint(\n fine_tuned_oof[\n \"fold\"\n ].value_counts()\n .sort_index()\n)\n\nprint(\n \"Fold coverage: PASS\"\n)\n\nprint(\n \"Probability validity: PASS\"\n)\n\nprint(\n \"Label validity: PASS\"\n)\n\n# ----------------------------------------------------------------------\n# SAVE OOF PREDICTIONS\n# ----------------------------------------------------------------------\n\noutput_path = (\n \"/kaggle/working/rsna_knee_audit/\"\n \"cell58_finetuned_oof_predictions.csv\"\n)\n\nfine_tuned_oof.to_csv(\n output_path,\n index=False\n)\n\nif not os.path.exists(\n output_path\n):\n raise RuntimeError(\n \"Fine-tuned OOF output file was not created.\"\n )\n\nprint()\nprint(\n f\"Saved: {output_path}\"\n)\n\n# ----------------------------------------------------------------------\n# FINAL VERDICT\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 58R VERDICT\")\nprint(\"=\" * 70)\n\nprint(\n \"Fine-tuned 3-fold OOF inference: PASS\"\n)\n\nprint(\n \"58/58 studies covered exactly once: PASS\"\n)\n\nprint(\n \"Fold distribution 20/19/19: PASS\"\n)\n\nprint(\n \"21-channel input: PASS\"\n)\n\nprint(\n \"12-target output: PASS\"\n)\n\nprint(\n \"Checkpoint compatibility: PASS\"\n)\n\nprint(\n \"Probability validity: PASS\"\n)\n\nprint(\n \"No training performed.\"\n)\n\nprint(\n \"No original checkpoint modified.\"\n)\n\nprint(\n \"No test data used.\"\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 58R COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Fine-tuned OOF predictions are ready \"\n \"for comparison against baseline Cell-46 OOF.\"\n)\n\nprint(\n \"Send the complete Cell 58R output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:14:12.918128Z","iopub.execute_input":"2026-08-11T11:14:12.918493Z","iopub.status.idle":"2026-08-11T11:14:45.324307Z","shell.execute_reply.started":"2026-08-11T11:14:12.918465Z","shell.execute_reply":"2026-08-11T11:14:45.323075Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 59R - BASELINE vs FINE-TUNED OOF COMPARISON\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nfrom sklearn.metrics import (\n roc_auc_score,\n average_precision_score,\n f1_score,\n brier_score_loss\n)\n\nprint(\"=\" * 70)\nprint(\"CELL 59R - BASELINE vs FINE-TUNED OOF COMPARISON\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECT\n# ----------------------------------------------------------------------\n\nif \"TARGETS\" not in globals():\n raise RuntimeError(\n \"Missing notebook object: TARGETS. Do not continue.\"\n )\n\nif len(TARGETS) != 12:\n raise RuntimeError(\n f\"Expected 12 targets, found {len(TARGETS)}.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(f\"Targets: {len(TARGETS)}\")\n\n# ----------------------------------------------------------------------\n# FILE PATHS\n# ----------------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nbaseline_path = os.path.join(\n audit_dir,\n \"cell46_oof_predictions.csv\"\n)\n\nfinetuned_path = os.path.join(\n audit_dir,\n \"cell58_finetuned_oof_predictions.csv\"\n)\n\nif not os.path.exists(baseline_path):\n raise RuntimeError(\n \"Baseline OOF file missing: \"\n + baseline_path\n )\n\nif not os.path.exists(finetuned_path):\n raise RuntimeError(\n \"Fine-tuned OOF file missing: \"\n + finetuned_path\n )\n\nprint(\"Baseline OOF file: EXISTS\")\nprint(\"Fine-tuned OOF file: EXISTS\")\n\n# ----------------------------------------------------------------------\n# LOAD FILES\n# ----------------------------------------------------------------------\n\nbaseline_oof = pd.read_csv(\n baseline_path\n)\n\nfinetuned_oof = pd.read_csv(\n finetuned_path\n)\n\nprint()\nprint(\n f\"Baseline OOF shape: {baseline_oof.shape}\"\n)\n\nprint(\n f\"Fine-tuned OOF shape: {finetuned_oof.shape}\"\n)\n\n# ----------------------------------------------------------------------\n# PRINT ACTUAL SCHEMA FIRST\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"ACTUAL BASELINE OOF SCHEMA\")\nprint(\"=\" * 70)\n\nfor i, column in enumerate(\n baseline_oof.columns\n):\n print(\n f\"{i:02d}: {column}\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"ACTUAL FINE-TUNED OOF SCHEMA\")\nprint(\"=\" * 70)\n\nfor i, column in enumerate(\n finetuned_oof.columns\n):\n print(\n f\"{i:02d}: {column}\"\n )\n\n# ----------------------------------------------------------------------\n# IDENTIFY ID / FOLD COLUMNS\n# ----------------------------------------------------------------------\n\ndef find_column(\n columns,\n candidates\n):\n\n for candidate in candidates:\n if candidate in columns:\n return candidate\n\n return None\n\n\nbaseline_id_col = find_column(\n baseline_oof.columns,\n [\n \"StudyInstanceUID\",\n \"study_id\",\n \"StudyID\"\n ]\n)\n\nfinetuned_id_col = find_column(\n finetuned_oof.columns,\n [\n \"StudyInstanceUID\",\n \"study_id\",\n \"StudyID\"\n ]\n)\n\nif baseline_id_col is None:\n raise RuntimeError(\n \"Could not identify baseline study-ID column.\"\n )\n\nif finetuned_id_col is None:\n raise RuntimeError(\n \"Could not identify fine-tuned study-ID column.\"\n )\n\nif \"fold\" not in baseline_oof.columns:\n raise RuntimeError(\n \"Baseline OOF is missing the fold column.\"\n )\n\nif \"fold\" not in finetuned_oof.columns:\n raise RuntimeError(\n \"Fine-tuned OOF is missing the fold column.\"\n )\n\nprint()\nprint(\n f\"Baseline study-ID column: {baseline_id_col}\"\n)\n\nprint(\n f\"Fine-tuned study-ID column: {finetuned_id_col}\"\n)\n\nprint(\"Fold column: PASS\")\n\n# ----------------------------------------------------------------------\n# IDENTIFY PROBABILITY COLUMNS\n# ----------------------------------------------------------------------\n\ndef identify_probability_columns(\n table\n):\n\n columns = list(table.columns)\n\n candidates = [\n column\n for column in columns\n if column.endswith(\"_prob\")\n ]\n\n if len(candidates) != len(TARGETS):\n raise RuntimeError(\n \"Expected exactly \"\n f\"{len(TARGETS)} probability columns, \"\n f\"found {len(candidates)}: \"\n + \", \".join(candidates)\n )\n\n return candidates\n\n\nbaseline_prob_columns = (\n identify_probability_columns(\n baseline_oof\n )\n)\n\nfinetuned_prob_columns = (\n identify_probability_columns(\n finetuned_oof\n )\n)\n\nprint()\nprint(\n \"Baseline probability columns: \"\n f\"{len(baseline_prob_columns)}\"\n)\n\nprint(\n \"Fine-tuned probability columns: \"\n f\"{len(finetuned_prob_columns)}\"\n)\n\nprint(\"Probability-column count: PASS\")\n\n# ----------------------------------------------------------------------\n# IDENTIFY LABEL COLUMNS FROM ACTUAL FILE STRUCTURE\n# ----------------------------------------------------------------------\n#\n# We do NOT assume that the label columns are named exactly\n# \"ACL\", \"MCL\", etc.\n#\n# The known structure is:\n#\n# study ID\n# fold\n# 12 label columns\n# 12 probability columns\n#\n# Therefore the 12 columns which are neither ID/fold/probability\n# columns are the label columns.\n# ----------------------------------------------------------------------\n\ndef identify_label_columns(\n table,\n id_col,\n probability_columns\n):\n\n excluded = {\n id_col,\n \"fold\"\n }\n\n excluded.update(\n probability_columns\n )\n\n label_columns = [\n column\n for column in table.columns\n if column not in excluded\n ]\n\n if len(label_columns) != len(TARGETS):\n raise RuntimeError(\n \"Could not identify exactly 12 label columns. \"\n f\"Found {len(label_columns)}: \"\n + \", \".join(label_columns)\n )\n\n return label_columns\n\n\nbaseline_label_columns = (\n identify_label_columns(\n baseline_oof,\n baseline_id_col,\n baseline_prob_columns\n )\n)\n\nfinetuned_label_columns = (\n identify_label_columns(\n finetuned_oof,\n finetuned_id_col,\n finetuned_prob_columns\n )\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"IDENTIFIED LABEL COLUMNS\")\nprint(\"=\" * 70)\n\nprint(\"Baseline labels:\")\n\nfor target, column in zip(\n TARGETS,\n baseline_label_columns\n):\n print(\n f\"{target:<22} -> {column}\"\n )\n\nprint()\nprint(\"Fine-tuned labels:\")\n\nfor target, column in zip(\n TARGETS,\n finetuned_label_columns\n):\n print(\n f\"{target:<22} -> {column}\"\n )\n\nprint()\nprint(\"Label-column count: PASS\")\n\n# ----------------------------------------------------------------------\n# BUILD CANONICAL TARGET MAPPINGS\n# ----------------------------------------------------------------------\n\nbaseline_target_map = {\n target: column\n for target, column in zip(\n TARGETS,\n baseline_label_columns\n )\n}\n\nfinetuned_target_map = {\n target: column\n for target, column in zip(\n TARGETS,\n finetuned_label_columns\n )\n}\n\nbaseline_probability_map = {\n target: column\n for target, column in zip(\n TARGETS,\n baseline_prob_columns\n )\n}\n\nfinetuned_probability_map = {\n target: column\n for target, column in zip(\n TARGETS,\n finetuned_prob_columns\n )\n}\n\nprint(\"Target-column mapping: PASS\")\n\n# ----------------------------------------------------------------------\n# STUDY ALIGNMENT\n# ----------------------------------------------------------------------\n\nbaseline_ids = (\n baseline_oof[\n baseline_id_col\n ]\n .astype(str)\n)\n\nfinetuned_ids = (\n finetuned_oof[\n finetuned_id_col\n ]\n .astype(str)\n)\n\nif len(baseline_oof) != 58:\n raise RuntimeError(\n \"Baseline OOF does not contain 58 rows.\"\n )\n\nif len(finetuned_oof) != 58:\n raise RuntimeError(\n \"Fine-tuned OOF does not contain 58 rows.\"\n )\n\nif baseline_ids.nunique() != 58:\n raise RuntimeError(\n \"Baseline OOF does not contain 58 unique studies.\"\n )\n\nif finetuned_ids.nunique() != 58:\n raise RuntimeError(\n \"Fine-tuned OOF does not contain 58 unique studies.\"\n )\n\nif set(baseline_ids) != set(finetuned_ids):\n raise RuntimeError(\n \"Baseline and fine-tuned OOF study sets differ.\"\n )\n\n# ----------------------------------------------------------------------\n# CREATE SORTED COPIES FOR EXACT STUDY ALIGNMENT\n# ----------------------------------------------------------------------\n\nbaseline_eval = baseline_oof.copy()\n\nfinetuned_eval = finetuned_oof.copy()\n\nbaseline_eval[\"_study_key\"] = (\n baseline_eval[\n baseline_id_col\n ].astype(str)\n)\n\nfinetuned_eval[\"_study_key\"] = (\n finetuned_eval[\n finetuned_id_col\n ].astype(str)\n)\n\nbaseline_eval = (\n baseline_eval\n .sort_values(\"_study_key\")\n .reset_index(drop=True)\n)\n\nfinetuned_eval = (\n finetuned_eval\n .sort_values(\"_study_key\")\n .reset_index(drop=True)\n)\n\nif not (\n baseline_eval[\"_study_key\"]\n ==\n finetuned_eval[\"_study_key\"]\n).all():\n raise RuntimeError(\n \"Study ordering/alignment failed.\"\n )\n\n# ----------------------------------------------------------------------\n# FOLD ALIGNMENT\n# ----------------------------------------------------------------------\n\nif not (\n baseline_eval[\"fold\"].astype(int)\n ==\n finetuned_eval[\"fold\"].astype(int)\n).all():\n raise RuntimeError(\n \"Baseline and fine-tuned fold assignments differ.\"\n )\n\nfold_counts = (\n baseline_eval[\n \"fold\"\n ]\n .astype(int)\n .value_counts()\n .sort_index()\n .to_dict()\n)\n\nexpected_fold_counts = {\n 0: 20,\n 1: 19,\n 2: 19\n}\n\nif fold_counts != expected_fold_counts:\n raise RuntimeError(\n \"Unexpected baseline fold distribution: \"\n + str(fold_counts)\n )\n\nfinetuned_fold_counts = (\n finetuned_eval[\n \"fold\"\n ]\n .astype(int)\n .value_counts()\n .sort_index()\n .to_dict()\n)\n\nif finetuned_fold_counts != expected_fold_counts:\n raise RuntimeError(\n \"Unexpected fine-tuned fold distribution: \"\n + str(finetuned_fold_counts)\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"OOF ALIGNMENT\")\nprint(\"=\" * 70)\n\nprint(\"Baseline studies: 58\")\nprint(\"Fine-tuned studies: 58\")\nprint(\"Same study set: PASS\")\nprint(\"Same fold assignments: PASS\")\nprint(\"Fold distribution: 20 / 19 / 19\")\nprint(\"OOF alignment: PASS\")\n\n# ----------------------------------------------------------------------\n# LABEL CONSISTENCY\n# ----------------------------------------------------------------------\n\nfor target in TARGETS:\n\n baseline_labels = (\n baseline_eval[\n baseline_target_map[target]\n ]\n .to_numpy(\n dtype=np.float64\n )\n )\n\n finetuned_labels = (\n finetuned_eval[\n finetuned_target_map[target]\n ]\n .to_numpy(\n dtype=np.float64\n )\n )\n\n if not np.array_equal(\n baseline_labels,\n finetuned_labels\n ):\n raise RuntimeError(\n f\"Label mismatch for target: {target}\"\n )\n\nprint(\"Label consistency: PASS\")\n\n# ----------------------------------------------------------------------\n# PROBABILITY VALIDITY\n# ----------------------------------------------------------------------\n\nfor name, table, probability_map in [\n (\n \"Baseline\",\n baseline_eval,\n baseline_probability_map\n ),\n (\n \"Fine-tuned\",\n finetuned_eval,\n finetuned_probability_map\n )\n]:\n\n probability_matrix = np.column_stack(\n [\n table[\n probability_map[target]\n ].to_numpy(\n dtype=np.float64\n )\n for target in TARGETS\n ]\n )\n\n if not np.isfinite(\n probability_matrix\n ).all():\n raise RuntimeError(\n f\"{name} probabilities contain NaN/Inf.\"\n )\n\n if (\n probability_matrix.min() < 0.0\n or probability_matrix.max() > 1.0\n ):\n raise RuntimeError(\n f\"{name} probabilities outside [0,1].\"\n )\n\nprint(\"Baseline probability validity: PASS\")\nprint(\"Fine-tuned probability validity: PASS\")\n\n# ----------------------------------------------------------------------\n# THRESHOLD GRID\n# ----------------------------------------------------------------------\n\nthresholds = np.linspace(\n 0.10,\n 0.90,\n 33\n)\n\n# ----------------------------------------------------------------------\n# METRIC FUNCTION\n# ----------------------------------------------------------------------\n\ndef calculate_metrics(\n table,\n target,\n label_map,\n probability_map\n):\n\n y_true = (\n table[\n label_map[target]\n ]\n .to_numpy(\n dtype=np.float64\n )\n )\n\n y_prob = (\n table[\n probability_map[target]\n ]\n .to_numpy(\n dtype=np.float64\n )\n )\n\n positive_count = int(\n y_true.sum()\n )\n\n negative_count = int(\n len(y_true) - positive_count\n )\n\n if positive_count == 0:\n raise RuntimeError(\n f\"{target}: no positive OOF samples.\"\n )\n\n if negative_count == 0:\n raise RuntimeError(\n f\"{target}: no negative OOF samples.\"\n )\n\n roc_auc = roc_auc_score(\n y_true,\n y_prob\n )\n\n pr_auc = average_precision_score(\n y_true,\n y_prob\n )\n\n brier = brier_score_loss(\n y_true,\n y_prob\n )\n\n pred_050 = (\n y_prob >= 0.50\n ).astype(int)\n\n f1_050 = f1_score(\n y_true,\n pred_050,\n zero_division=0\n )\n\n best_threshold = None\n best_f1 = -1.0\n\n for threshold in thresholds:\n\n predictions = (\n y_prob >= threshold\n ).astype(int)\n\n current_f1 = f1_score(\n y_true,\n predictions,\n zero_division=0\n )\n\n if (\n current_f1 > best_f1\n or (\n current_f1 == best_f1\n and (\n best_threshold is None\n or threshold > best_threshold\n )\n )\n ):\n best_f1 = current_f1\n best_threshold = float(\n threshold\n )\n\n positive_mean = float(\n y_prob[y_true == 1].mean()\n )\n\n negative_mean = float(\n y_prob[y_true == 0].mean()\n )\n\n separation = (\n positive_mean\n -\n negative_mean\n )\n\n return {\n \"target\": target,\n \"positive_count\": positive_count,\n \"negative_count\": negative_count,\n \"roc_auc\": float(roc_auc),\n \"pr_auc\": float(pr_auc),\n \"brier_score\": float(brier),\n \"f1_at_0.50\": float(f1_050),\n \"best_threshold\": float(\n best_threshold\n ),\n \"best_f1\": float(best_f1),\n \"f1_improvement\": float(\n best_f1 - f1_050\n ),\n \"probability_separation\": float(\n separation\n )\n }\n\n# ----------------------------------------------------------------------\n# CALCULATE METRICS\n# ----------------------------------------------------------------------\n\nbaseline_metrics = pd.DataFrame(\n [\n calculate_metrics(\n baseline_eval,\n target,\n baseline_target_map,\n baseline_probability_map\n )\n for target in TARGETS\n ]\n)\n\nfinetuned_metrics = pd.DataFrame(\n [\n calculate_metrics(\n finetuned_eval,\n target,\n finetuned_target_map,\n finetuned_probability_map\n )\n for target in TARGETS\n ]\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"BASELINE OOF METRICS\")\nprint(\"=\" * 70)\n\nprint(\n baseline_metrics.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINE-TUNED OOF METRICS\")\nprint(\"=\" * 70)\n\nprint(\n finetuned_metrics.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------------\n# COMPARISON TABLE\n# ----------------------------------------------------------------------\n\ncomparison = pd.DataFrame(\n {\n \"target\": TARGETS,\n\n \"baseline_roc_auc\":\n baseline_metrics[\n \"roc_auc\"\n ],\n\n \"finetuned_roc_auc\":\n finetuned_metrics[\n \"roc_auc\"\n ],\n\n \"baseline_pr_auc\":\n baseline_metrics[\n \"pr_auc\"\n ],\n\n \"finetuned_pr_auc\":\n finetuned_metrics[\n \"pr_auc\"\n ],\n\n \"baseline_brier\":\n baseline_metrics[\n \"brier_score\"\n ],\n\n \"finetuned_brier\":\n finetuned_metrics[\n \"brier_score\"\n ],\n\n \"baseline_f1_at_0.50\":\n baseline_metrics[\n \"f1_at_0.50\"\n ],\n\n \"finetuned_f1_at_0.50\":\n finetuned_metrics[\n \"f1_at_0.50\"\n ],\n\n \"baseline_best_f1\":\n baseline_metrics[\n \"best_f1\"\n ],\n\n \"finetuned_best_f1\":\n finetuned_metrics[\n \"best_f1\"\n ],\n\n \"baseline_best_threshold\":\n baseline_metrics[\n \"best_threshold\"\n ],\n\n \"finetuned_best_threshold\":\n finetuned_metrics[\n \"best_threshold\"\n ],\n\n \"baseline_probability_separation\":\n baseline_metrics[\n \"probability_separation\"\n ],\n\n \"finetuned_probability_separation\":\n finetuned_metrics[\n \"probability_separation\"\n ]\n }\n)\n\ncomparison[\n \"delta_roc_auc\"\n] = (\n comparison[\n \"finetuned_roc_auc\"\n ]\n -\n comparison[\n \"baseline_roc_auc\"\n ]\n)\n\ncomparison[\n \"delta_pr_auc\"\n] = (\n comparison[\n \"finetuned_pr_auc\"\n ]\n -\n comparison[\n \"baseline_pr_auc\"\n ]\n)\n\ncomparison[\n \"delta_brier\"\n] = (\n comparison[\n \"finetuned_brier\"\n ]\n -\n comparison[\n \"baseline_brier\"\n ]\n)\n\ncomparison[\n \"delta_f1_at_0.50\"\n] = (\n comparison[\n \"finetuned_f1_at_0.50\"\n ]\n -\n comparison[\n \"baseline_f1_at_0.50\"\n ]\n)\n\ncomparison[\n \"delta_best_f1\"\n] = (\n comparison[\n \"finetuned_best_f1\"\n ]\n -\n comparison[\n \"baseline_best_f1\"\n ]\n)\n\ncomparison[\n \"delta_probability_separation\"\n] = (\n comparison[\n \"finetuned_probability_separation\"\n ]\n -\n comparison[\n \"baseline_probability_separation\"\n ]\n)\n\n# ----------------------------------------------------------------------\n# SUMMARY STATISTICS\n# ----------------------------------------------------------------------\n\nbaseline_mean_roc_auc = float(\n baseline_metrics[\n \"roc_auc\"\n ].mean()\n)\n\nfinetuned_mean_roc_auc = float(\n finetuned_metrics[\n \"roc_auc\"\n ].mean()\n)\n\nbaseline_mean_pr_auc = float(\n baseline_metrics[\n \"pr_auc\"\n ].mean()\n)\n\nfinetuned_mean_pr_auc = float(\n finetuned_metrics[\n \"pr_auc\"\n ].mean()\n)\n\nbaseline_mean_brier = float(\n baseline_metrics[\n \"brier_score\"\n ].mean()\n)\n\nfinetuned_mean_brier = float(\n finetuned_metrics[\n \"brier_score\"\n ].mean()\n)\n\nbaseline_mean_f1 = float(\n baseline_metrics[\n \"f1_at_0.50\"\n ].mean()\n)\n\nfinetuned_mean_f1 = float(\n finetuned_metrics[\n \"f1_at_0.50\"\n ].mean()\n)\n\nbaseline_mean_best_f1 = float(\n baseline_metrics[\n \"best_f1\"\n ].mean()\n)\n\nfinetuned_mean_best_f1 = float(\n finetuned_metrics[\n \"best_f1\"\n ].mean()\n)\n\nbaseline_mean_separation = float(\n baseline_metrics[\n \"probability_separation\"\n ].mean()\n)\n\nfinetuned_mean_separation = float(\n finetuned_metrics[\n \"probability_separation\"\n ].mean()\n)\n\n# ----------------------------------------------------------------------\n# PRINT COMPARISON\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"BASELINE vs FINE-TUNED OOF COMPARISON\")\nprint(\"=\" * 70)\n\nprint(\n comparison[\n [\n \"target\",\n \"baseline_roc_auc\",\n \"finetuned_roc_auc\",\n \"delta_roc_auc\",\n \"baseline_f1_at_0.50\",\n \"finetuned_f1_at_0.50\",\n \"delta_f1_at_0.50\",\n \"baseline_best_f1\",\n \"finetuned_best_f1\",\n \"delta_best_f1\"\n ]\n ].to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------------\n# MEAN SUMMARY\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"MEAN OOF SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Baseline mean ROC-AUC: \"\n f\"{baseline_mean_roc_auc:.4f}\"\n)\n\nprint(\n f\"Fine-tuned mean ROC-AUC: \"\n f\"{finetuned_mean_roc_auc:.4f}\"\n)\n\nprint(\n f\"ROC-AUC delta: \"\n f\"{finetuned_mean_roc_auc - baseline_mean_roc_auc:+.4f}\"\n)\n\nprint()\n\nprint(\n f\"Baseline mean PR-AUC: \"\n f\"{baseline_mean_pr_auc:.4f}\"\n)\n\nprint(\n f\"Fine-tuned mean PR-AUC: \"\n f\"{finetuned_mean_pr_auc:.4f}\"\n)\n\nprint(\n f\"PR-AUC delta: \"\n f\"{finetuned_mean_pr_auc - baseline_mean_pr_auc:+.4f}\"\n)\n\nprint()\n\nprint(\n f\"Baseline mean Brier: \"\n f\"{baseline_mean_brier:.4f}\"\n)\n\nprint(\n f\"Fine-tuned mean Brier: \"\n f\"{finetuned_mean_brier:.4f}\"\n)\n\nprint(\n f\"Brier delta: \"\n f\"{finetuned_mean_brier - baseline_mean_brier:+.4f}\"\n)\n\nprint()\n\nprint(\n f\"Baseline mean F1 @ 0.50: \"\n f\"{baseline_mean_f1:.4f}\"\n)\n\nprint(\n f\"Fine-tuned mean F1 @ 0.50: \"\n f\"{finetuned_mean_f1:.4f}\"\n)\n\nprint(\n f\"F1 @ 0.50 delta: \"\n f\"{finetuned_mean_f1 - baseline_mean_f1:+.4f}\"\n)\n\nprint()\n\nprint(\n f\"Baseline mean best F1: \"\n f\"{baseline_mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Fine-tuned mean best F1: \"\n f\"{finetuned_mean_best_f1:.4f}\"\n)\n\nprint(\n f\"Best F1 delta: \"\n f\"{finetuned_mean_best_f1 - baseline_mean_best_f1:+.4f}\"\n)\n\nprint()\n\nprint(\n f\"Baseline mean probability separation: \"\n f\"{baseline_mean_separation:.4f}\"\n)\n\nprint(\n f\"Fine-tuned mean probability separation: \"\n f\"{finetuned_mean_separation:.4f}\"\n)\n\nprint(\n f\"Probability separation delta: \"\n f\"{finetuned_mean_separation - baseline_mean_separation:+.4f}\"\n)\n\n# ----------------------------------------------------------------------\n# WIN COUNTS\n# ----------------------------------------------------------------------\n\nroc_auc_wins = int(\n (\n comparison[\n \"delta_roc_auc\"\n ] > 0\n ).sum()\n)\n\npr_auc_wins = int(\n (\n comparison[\n \"delta_pr_auc\"\n ] > 0\n ).sum()\n)\n\nf1_wins = int(\n (\n comparison[\n \"delta_f1_at_0.50\"\n ] > 0\n ).sum()\n)\n\nbest_f1_wins = int(\n (\n comparison[\n \"delta_best_f1\"\n ] > 0\n ).sum()\n)\n\nbrier_wins = int(\n (\n comparison[\n \"delta_brier\"\n ] < 0\n ).sum()\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"TARGET-LEVEL IMPROVEMENT COUNTS\")\nprint(\"=\" * 70)\n\nprint(\n f\"ROC-AUC improvement: \"\n f\"{roc_auc_wins}/12\"\n)\n\nprint(\n f\"PR-AUC improvement: \"\n f\"{pr_auc_wins}/12\"\n)\n\nprint(\n f\"F1 @ 0.50 improvement: \"\n f\"{f1_wins}/12\"\n)\n\nprint(\n f\"Best-F1 improvement: \"\n f\"{best_f1_wins}/12\"\n)\n\nprint(\n f\"Brier improvement: \"\n f\"{brier_wins}/12\"\n)\n\n# ----------------------------------------------------------------------\n# SAVE RESULTS\n# ----------------------------------------------------------------------\n\ncomparison_path = os.path.join(\n audit_dir,\n \"cell59_baseline_vs_finetuned_oof.csv\"\n)\n\nbaseline_summary_path = os.path.join(\n audit_dir,\n \"cell59_baseline_oof_summary.csv\"\n)\n\nfinetuned_summary_path = os.path.join(\n audit_dir,\n \"cell59_finetuned_oof_summary.csv\"\n)\n\ncomparison.to_csv(\n comparison_path,\n index=False\n)\n\nbaseline_metrics.to_csv(\n baseline_summary_path,\n index=False\n)\n\nfinetuned_metrics.to_csv(\n finetuned_summary_path,\n index=False\n)\n\nif not all(\n os.path.exists(path)\n for path in [\n comparison_path,\n baseline_summary_path,\n finetuned_summary_path\n ]\n):\n raise RuntimeError(\n \"One or more Cell-59 output files were not created.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"OUTPUT FILES\")\nprint(\"=\" * 70)\n\nprint(\n f\"Saved: {comparison_path}\"\n)\n\nprint(\n f\"Saved: {baseline_summary_path}\"\n)\n\nprint(\n f\"Saved: {finetuned_summary_path}\"\n)\n\n# ----------------------------------------------------------------------\n# FINAL VERIFICATION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 59R VERIFICATION\")\nprint(\"=\" * 70)\n\nprint(\"58-study baseline OOF: PASS\")\nprint(\"58-study fine-tuned OOF: PASS\")\nprint(\"Same study set: PASS\")\nprint(\"Same fold assignments: PASS\")\nprint(\"12 target labels aligned: PASS\")\nprint(\"12 probability outputs aligned: PASS\")\nprint(\"Probability validity: PASS\")\nprint(\"Baseline metrics calculated: PASS\")\nprint(\"Fine-tuned metrics calculated: PASS\")\nprint(\"Comparison table: PASS\")\n\nprint()\nprint(\"No training performed.\")\nprint(\"No checkpoint modified.\")\nprint(\"No test data used.\")\nprint(\"No final model selected automatically.\")\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 59R COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Send the complete Cell 59R output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:20:07.284167Z","iopub.execute_input":"2026-08-11T11:20:07.284512Z","iopub.status.idle":"2026-08-11T11:20:09.182046Z","shell.execute_reply.started":"2026-08-11T11:20:07.284485Z","shell.execute_reply":"2026-08-11T11:20:09.180833Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 60 - OOF MODEL SELECTION + THRESHOLD DECISION AUDIT\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"CELL 60 - OOF MODEL SELECTION + THRESHOLD DECISION AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED OBJECTS\n# ----------------------------------------------------------------------\n\nif \"TARGETS\" not in globals():\n raise RuntimeError(\n \"Missing notebook object: TARGETS. Do not continue.\"\n )\n\nif len(TARGETS) != 12:\n raise RuntimeError(\n f\"Expected 12 targets, found {len(TARGETS)}.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(f\"Targets: {len(TARGETS)}\")\n\n# ----------------------------------------------------------------------\n# PATHS\n# ----------------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\ncomparison_path = os.path.join(\n audit_dir,\n \"cell59_baseline_vs_finetuned_oof.csv\"\n)\n\nbaseline_summary_path = os.path.join(\n audit_dir,\n \"cell59_baseline_oof_summary.csv\"\n)\n\nfinetuned_summary_path = os.path.join(\n audit_dir,\n \"cell59_finetuned_oof_summary.csv\"\n)\n\nrequired_paths = [\n comparison_path,\n baseline_summary_path,\n finetuned_summary_path\n]\n\nfor path in required_paths:\n if not os.path.exists(path):\n raise RuntimeError(\n f\"Required Cell-59 output missing: {path}\"\n )\n\nprint(\"Cell-59 comparison file: EXISTS\")\nprint(\"Cell-59 baseline summary: EXISTS\")\nprint(\"Cell-59 fine-tuned summary: EXISTS\")\n\n# ----------------------------------------------------------------------\n# LOAD RESULTS\n# ----------------------------------------------------------------------\n\ncomparison = pd.read_csv(\n comparison_path\n)\n\nbaseline_metrics = pd.read_csv(\n baseline_summary_path\n)\n\nfinetuned_metrics = pd.read_csv(\n finetuned_summary_path\n)\n\n# ----------------------------------------------------------------------\n# SHAPE VALIDATION\n# ----------------------------------------------------------------------\n\nif comparison.shape[0] != 12:\n raise RuntimeError(\n f\"Expected 12 comparison rows, found {comparison.shape[0]}.\"\n )\n\nif baseline_metrics.shape[0] != 12:\n raise RuntimeError(\n f\"Expected 12 baseline rows, found {baseline_metrics.shape[0]}.\"\n )\n\nif finetuned_metrics.shape[0] != 12:\n raise RuntimeError(\n f\"Expected 12 fine-tuned rows, found {finetuned_metrics.shape[0]}.\"\n )\n\nprint(\"Comparison rows: 12\")\nprint(\"Baseline metric rows: 12\")\nprint(\"Fine-tuned metric rows: 12\")\nprint(\"Target coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# TARGET ORDER VALIDATION\n# ----------------------------------------------------------------------\n\nif list(comparison[\"target\"]) != list(TARGETS):\n raise RuntimeError(\n \"Comparison target order does not match TARGETS.\"\n )\n\nif list(baseline_metrics[\"target\"]) != list(TARGETS):\n raise RuntimeError(\n \"Baseline target order does not match TARGETS.\"\n )\n\nif list(finetuned_metrics[\"target\"]) != list(TARGETS):\n raise RuntimeError(\n \"Fine-tuned target order does not match TARGETS.\"\n )\n\nprint(\"Target ordering: PASS\")\n\n# ----------------------------------------------------------------------\n# REQUIRED COLUMNS\n# ----------------------------------------------------------------------\n\nrequired_comparison_columns = [\n \"target\",\n \"baseline_roc_auc\",\n \"finetuned_roc_auc\",\n \"delta_roc_auc\",\n \"baseline_pr_auc\",\n \"finetuned_pr_auc\",\n \"delta_pr_auc\",\n \"baseline_brier\",\n \"finetuned_brier\",\n \"delta_brier\",\n \"baseline_f1_at_0.50\",\n \"finetuned_f1_at_0.50\",\n \"delta_f1_at_0.50\",\n \"baseline_best_f1\",\n \"finetuned_best_f1\",\n \"delta_best_f1\",\n \"baseline_probability_separation\",\n \"finetuned_probability_separation\",\n \"delta_probability_separation\"\n]\n\nmissing_columns = [\n column\n for column in required_comparison_columns\n if column not in comparison.columns\n]\n\nif missing_columns:\n raise RuntimeError(\n \"Cell-59 comparison missing columns: \"\n + \", \".join(missing_columns)\n )\n\nprint(\"Comparison schema: PASS\")\n\n# ----------------------------------------------------------------------\n# CREATE MODEL DECISION\n# ----------------------------------------------------------------------\n#\n# We do not select a model from a single metric.\n#\n# Priority:\n# 1. ROC-AUC\n# 2. PR-AUC\n# 3. Best-F1\n#\n# Brier is treated separately because lower is better.\n#\n# A model is considered clearly better only when it improves at\n# least two of the three primary discrimination metrics.\n# ----------------------------------------------------------------------\n\ndecision_rows = []\n\nfor target in TARGETS:\n\n row = comparison[\n comparison[\"target\"] == target\n ].iloc[0]\n\n baseline_roc = float(\n row[\"baseline_roc_auc\"]\n )\n\n finetuned_roc = float(\n row[\"finetuned_roc_auc\"]\n )\n\n baseline_pr = float(\n row[\"baseline_pr_auc\"]\n )\n\n finetuned_pr = float(\n row[\"finetuned_pr_auc\"]\n )\n\n baseline_best_f1 = float(\n row[\"baseline_best_f1\"]\n )\n\n finetuned_best_f1 = float(\n row[\"finetuned_best_f1\"]\n )\n\n baseline_brier = float(\n row[\"baseline_brier\"]\n )\n\n finetuned_brier = float(\n row[\"finetuned_brier\"]\n )\n\n roc_improved = (\n finetuned_roc > baseline_roc\n )\n\n pr_improved = (\n finetuned_pr > baseline_pr\n )\n\n best_f1_improved = (\n finetuned_best_f1 > baseline_best_f1\n )\n\n brier_improved = (\n finetuned_brier < baseline_brier\n )\n\n primary_wins = int(roc_improved) + int(\n pr_improved\n ) + int(\n best_f1_improved\n )\n\n primary_losses = int(\n finetuned_roc < baseline_roc\n ) + int(\n finetuned_pr < baseline_pr\n ) + int(\n finetuned_best_f1 < baseline_best_f1\n )\n\n if primary_wins >= 2 and primary_losses == 0:\n decision = \"fine_tuned\"\n\n elif primary_losses >= 2 and primary_wins == 0:\n decision = \"baseline\"\n\n else:\n decision = \"uncertain\"\n\n decision_rows.append(\n {\n \"target\": target,\n \"baseline_roc_auc\": baseline_roc,\n \"finetuned_roc_auc\": finetuned_roc,\n \"delta_roc_auc\": finetuned_roc - baseline_roc,\n \"baseline_pr_auc\": baseline_pr,\n \"finetuned_pr_auc\": finetuned_pr,\n \"delta_pr_auc\": finetuned_pr - baseline_pr,\n \"baseline_best_f1\": baseline_best_f1,\n \"finetuned_best_f1\": finetuned_best_f1,\n \"delta_best_f1\": (\n finetuned_best_f1\n -\n baseline_best_f1\n ),\n \"baseline_brier\": baseline_brier,\n \"finetuned_brier\": finetuned_brier,\n \"delta_brier\": (\n finetuned_brier\n -\n baseline_brier\n ),\n \"roc_auc_improved\": roc_improved,\n \"pr_auc_improved\": pr_improved,\n \"best_f1_improved\": best_f1_improved,\n \"brier_improved\": brier_improved,\n \"primary_metric_wins\": primary_wins,\n \"primary_metric_losses\": primary_losses,\n \"decision\": decision\n }\n )\n\ndecision_table = pd.DataFrame(\n decision_rows\n)\n\n# ----------------------------------------------------------------------\n# THRESHOLD SELECTION\n# ----------------------------------------------------------------------\n\nthreshold_rows = []\n\nfor target in TARGETS:\n\n baseline_row = baseline_metrics[\n baseline_metrics[\"target\"] == target\n ].iloc[0]\n\n finetuned_row = finetuned_metrics[\n finetuned_metrics[\"target\"] == target\n ].iloc[0]\n\n decision_row = decision_table[\n decision_table[\"target\"] == target\n ].iloc[0]\n\n if decision_row[\"decision\"] == \"baseline\":\n\n selected_model = \"baseline\"\n\n selected_threshold = float(\n baseline_row[\"best_threshold\"]\n )\n\n selected_best_f1 = float(\n baseline_row[\"best_f1\"]\n )\n\n elif decision_row[\"decision\"] == \"fine_tuned\":\n\n selected_model = \"fine_tuned\"\n\n selected_threshold = float(\n finetuned_row[\"best_threshold\"]\n )\n\n selected_best_f1 = float(\n finetuned_row[\"best_f1\"]\n )\n\n else:\n\n # For uncertain targets, choose the model with the\n # higher best-F1 while keeping the decision explicitly\n # marked as evidence-based rather than automatically\n # declaring the architecture globally superior.\n\n if (\n float(finetuned_row[\"best_f1\"])\n >\n float(baseline_row[\"best_f1\"])\n ):\n\n selected_model = \"fine_tuned\"\n\n selected_threshold = float(\n finetuned_row[\"best_threshold\"]\n )\n\n selected_best_f1 = float(\n finetuned_row[\"best_f1\"]\n )\n\n else:\n\n selected_model = \"baseline\"\n\n selected_threshold = float(\n baseline_row[\"best_threshold\"]\n )\n\n selected_best_f1 = float(\n baseline_row[\"best_f1\"]\n )\n\n threshold_rows.append(\n {\n \"target\": target,\n \"decision\": decision_row[\"decision\"],\n \"selected_model\": selected_model,\n \"selected_threshold\": selected_threshold,\n \"selected_best_f1\": selected_best_f1,\n \"baseline_threshold\": float(\n baseline_row[\"best_threshold\"]\n ),\n \"finetuned_threshold\": float(\n finetuned_row[\"best_threshold\"]\n )\n }\n )\n\nselection_table = pd.DataFrame(\n threshold_rows\n)\n\n# ----------------------------------------------------------------------\n# PRINT MODEL DECISION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TARGET-LEVEL MODEL DECISION\")\nprint(\"=\" * 70)\n\nprint(\n decision_table[\n [\n \"target\",\n \"delta_roc_auc\",\n \"delta_pr_auc\",\n \"delta_best_f1\",\n \"delta_brier\",\n \"primary_metric_wins\",\n \"primary_metric_losses\",\n \"decision\"\n ]\n ].to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------------\n# PRINT THRESHOLD SELECTION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"OOF THRESHOLD SELECTION\")\nprint(\"=\" * 70)\n\nprint(\n selection_table.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------------\n# SUMMARY\n# ----------------------------------------------------------------------\n\nbaseline_count = int(\n (\n decision_table[\"decision\"]\n == \"baseline\"\n ).sum()\n)\n\nfinetuned_count = int(\n (\n decision_table[\"decision\"]\n == \"fine_tuned\"\n ).sum()\n)\n\nuncertain_count = int(\n (\n decision_table[\"decision\"]\n == \"uncertain\"\n ).sum()\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"MODEL DECISION SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Baseline clearly preferred: \"\n f\"{baseline_count}/12\"\n)\n\nprint(\n f\"Fine-tuned clearly preferred: \"\n f\"{finetuned_count}/12\"\n)\n\nprint(\n f\"Uncertain targets: \"\n f\"{uncertain_count}/12\"\n)\n\n# ----------------------------------------------------------------------\n# SANITY CHECKS\n# ----------------------------------------------------------------------\n\nif not selection_table[\n \"selected_threshold\"\n].between(\n 0.10,\n 0.90\n).all():\n\n raise RuntimeError(\n \"Selected threshold outside audited range.\"\n )\n\nif not selection_table[\n \"selected_model\"\n].isin(\n [\n \"baseline\",\n \"fine_tuned\"\n ]\n).all():\n\n raise RuntimeError(\n \"Invalid selected model.\"\n )\n\nprint()\nprint(\"Threshold range validation: PASS\")\nprint(\"Model-selection validation: PASS\")\n\n# ----------------------------------------------------------------------\n# SAVE\n# ----------------------------------------------------------------------\n\ndecision_path = os.path.join(\n audit_dir,\n \"cell60_model_selection_audit.csv\"\n)\n\nthreshold_path = os.path.join(\n audit_dir,\n \"cell60_selected_thresholds.csv\"\n)\n\ndecision_table.to_csv(\n decision_path,\n index=False\n)\n\nselection_table.to_csv(\n threshold_path,\n index=False\n)\n\nif not os.path.exists(\n decision_path\n):\n raise RuntimeError(\n \"Model-selection audit was not saved.\"\n )\n\nif not os.path.exists(\n threshold_path\n):\n raise RuntimeError(\n \"Threshold selection was not saved.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"OUTPUT FILES\")\nprint(\"=\" * 70)\n\nprint(\n f\"Saved: {decision_path}\"\n)\n\nprint(\n f\"Saved: {threshold_path}\"\n)\n\n# ----------------------------------------------------------------------\n# FINAL VERIFICATION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 60 VERIFICATION\")\nprint(\"=\" * 70)\n\nprint(\"12 targets evaluated: PASS\")\nprint(\"Baseline OOF evidence used: PASS\")\nprint(\"Fine-tuned OOF evidence used: PASS\")\nprint(\"Thresholds derived from OOF: PASS\")\nprint(\"No training performed: PASS\")\nprint(\"No checkpoint modified: PASS\")\nprint(\"No test data used: PASS\")\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 60 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Send the complete Cell 60 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:31:57.24894Z","iopub.execute_input":"2026-08-11T11:31:57.249316Z","iopub.status.idle":"2026-08-11T11:31:57.322457Z","shell.execute_reply.started":"2026-08-11T11:31:57.249284Z","shell.execute_reply":"2026-08-11T11:31:57.321612Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 61 - FINAL OOF-BACKED MODEL + THRESHOLD PLAN AUDIT\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"CELL 61 - FINAL OOF-BACKED MODEL + THRESHOLD PLAN AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED OBJECTS\n# ----------------------------------------------------------------------\n\nif \"TARGETS\" not in globals():\n raise RuntimeError(\n \"Missing notebook object: TARGETS. Do not continue.\"\n )\n\nif len(TARGETS) != 12:\n raise RuntimeError(\n f\"Expected 12 targets, found {len(TARGETS)}.\"\n )\n\nprint(\"TARGETS: PASS\")\nprint(f\"Target count: {len(TARGETS)}\")\n\n# ----------------------------------------------------------------------\n# AUDIT DIRECTORY\n# ----------------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nif not os.path.isdir(audit_dir):\n raise RuntimeError(\n f\"Audit directory missing: {audit_dir}\"\n )\n\nprint(\"Audit directory: PASS\")\n\n# ----------------------------------------------------------------------\n# REQUIRED CELL-60 OUTPUTS\n# ----------------------------------------------------------------------\n\nmodel_selection_path = os.path.join(\n audit_dir,\n \"cell60_model_selection_audit.csv\"\n)\n\nthreshold_selection_path = os.path.join(\n audit_dir,\n \"cell60_selected_thresholds.csv\"\n)\n\nrequired_files = [\n model_selection_path,\n threshold_selection_path\n]\n\nfor path in required_files:\n if not os.path.exists(path):\n raise RuntimeError(\n f\"Required Cell-60 output missing: {path}\"\n )\n\nprint(\"Cell-60 model-selection audit: EXISTS\")\nprint(\"Cell-60 threshold-selection audit: EXISTS\")\n\n# ----------------------------------------------------------------------\n# LOAD CELL-60 RESULTS\n# ----------------------------------------------------------------------\n\nmodel_selection = pd.read_csv(\n model_selection_path\n)\n\nthreshold_selection = pd.read_csv(\n threshold_selection_path\n)\n\n# ----------------------------------------------------------------------\n# BASIC VALIDATION\n# ----------------------------------------------------------------------\n\nif len(model_selection) != 12:\n raise RuntimeError(\n \"Cell-60 model-selection table must contain exactly 12 targets.\"\n )\n\nif len(threshold_selection) != 12:\n raise RuntimeError(\n \"Cell-60 threshold table must contain exactly 12 targets.\"\n )\n\nprint(\"Model-selection rows: 12\")\nprint(\"Threshold-selection rows: 12\")\n\n# ----------------------------------------------------------------------\n# TARGET ORDER\n# ----------------------------------------------------------------------\n\nif list(model_selection[\"target\"]) != list(TARGETS):\n raise RuntimeError(\n \"Model-selection target order does not match TARGETS.\"\n )\n\nif list(threshold_selection[\"target\"]) != list(TARGETS):\n raise RuntimeError(\n \"Threshold-selection target order does not match TARGETS.\"\n )\n\nprint(\"Target ordering: PASS\")\n\n# ----------------------------------------------------------------------\n# REQUIRED COLUMNS\n# ----------------------------------------------------------------------\n\nrequired_model_columns = [\n \"target\",\n \"decision\",\n \"primary_metric_wins\",\n \"primary_metric_losses\"\n]\n\nrequired_threshold_columns = [\n \"target\",\n \"selected_model\",\n \"selected_threshold\",\n \"selected_best_f1\",\n \"baseline_threshold\",\n \"finetuned_threshold\"\n]\n\nmissing_model_columns = [\n c for c in required_model_columns\n if c not in model_selection.columns\n]\n\nmissing_threshold_columns = [\n c for c in required_threshold_columns\n if c not in threshold_selection.columns\n]\n\nif missing_model_columns:\n raise RuntimeError(\n \"Cell-60 model-selection table missing columns: \"\n + \", \".join(missing_model_columns)\n )\n\nif missing_threshold_columns:\n raise RuntimeError(\n \"Cell-60 threshold table missing columns: \"\n + \", \".join(missing_threshold_columns)\n )\n\nprint(\"Model-selection schema: PASS\")\nprint(\"Threshold-selection schema: PASS\")\n\n# ----------------------------------------------------------------------\n# VALID MODEL NAMES\n# ----------------------------------------------------------------------\n\nvalid_models = {\n \"baseline\",\n \"fine_tuned\"\n}\n\ninvalid_models = [\n value\n for value in threshold_selection[\"selected_model\"]\n if value not in valid_models\n]\n\nif invalid_models:\n raise RuntimeError(\n \"Invalid selected model values: \"\n + \", \".join(sorted(set(invalid_models)))\n )\n\nprint(\"Selected model names: PASS\")\n\n# ----------------------------------------------------------------------\n# VALID THRESHOLDS\n# ----------------------------------------------------------------------\n\nthreshold_values = pd.to_numeric(\n threshold_selection[\"selected_threshold\"],\n errors=\"coerce\"\n)\n\nif threshold_values.isna().any():\n raise RuntimeError(\n \"Selected threshold contains NaN or non-numeric values.\"\n )\n\nif not threshold_values.between(\n 0.10,\n 0.90\n).all():\n raise RuntimeError(\n \"Selected threshold outside audited range 0.10-0.90.\"\n )\n\nprint(\"Selected threshold range: PASS\")\n\n# ----------------------------------------------------------------------\n# MODEL DECISION DISTRIBUTION\n# ----------------------------------------------------------------------\n\nbaseline_targets = threshold_selection.loc[\n threshold_selection[\"selected_model\"] == \"baseline\",\n \"target\"\n].tolist()\n\nfinetuned_targets = threshold_selection.loc[\n threshold_selection[\"selected_model\"] == \"fine_tuned\",\n \"target\"\n].tolist()\n\nif len(baseline_targets) + len(finetuned_targets) != 12:\n raise RuntimeError(\n \"Model-selection coverage does not equal 12 targets.\"\n )\n\nif set(baseline_targets).intersection(\n set(finetuned_targets)\n):\n raise RuntimeError(\n \"A target was assigned to both baseline and fine-tuned models.\"\n )\n\nprint(\"Target model coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# IMPORTANT OOF DECISION CHECK\n# ----------------------------------------------------------------------\n#\n# We preserve the Cell-60 decision exactly.\n#\n# We do NOT overwrite uncertain decisions.\n# We do NOT invent new thresholds.\n# We do NOT retrain here.\n# ----------------------------------------------------------------------\n\nmerged_plan = threshold_selection[\n [\n \"target\",\n \"selected_model\",\n \"selected_threshold\",\n \"selected_best_f1\"\n ]\n].copy()\n\nmerged_plan[\"selected_threshold\"] = pd.to_numeric(\n merged_plan[\"selected_threshold\"]\n)\n\nmerged_plan[\"selected_best_f1\"] = pd.to_numeric(\n merged_plan[\"selected_best_f1\"]\n)\n\n# ----------------------------------------------------------------------\n# FINAL PLAN TABLE\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL OOF-BACKED TARGET PLAN\")\nprint(\"=\" * 70)\n\nprint(\n merged_plan.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------------\n# MODEL COUNTS\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL MODEL ASSIGNMENT SUMMARY\")\nprint(\"=\" * 70)\n\nprint(\n f\"Baseline-selected targets: \"\n f\"{len(baseline_targets)}/12\"\n)\n\nprint(\n f\"Fine-tuned-selected targets: \"\n f\"{len(finetuned_targets)}/12\"\n)\n\nprint()\nprint(\"Baseline targets:\")\nprint(baseline_targets)\n\nprint()\nprint(\"Fine-tuned targets:\")\nprint(finetuned_targets)\n\n# ----------------------------------------------------------------------\n# CHECKPOINT AVAILABILITY\n# ----------------------------------------------------------------------\n\nbaseline_checkpoint_paths = {\n 0: os.path.join(\n audit_dir,\n \"cell37_fold0_best_model.pt\"\n ),\n 1: os.path.join(\n audit_dir,\n \"cell40_fold1_best_model.pt\"\n ),\n 2: os.path.join(\n audit_dir,\n \"cell44_fold2_best_model.pt\"\n )\n}\n\nfinetuned_checkpoint_paths = {\n 0: os.path.join(\n audit_dir,\n \"cell52_fold0_finetuned_experiment.pt\"\n ),\n 1: os.path.join(\n audit_dir,\n \"cell54_fold1_finetuned_experiment.pt\"\n ),\n 2: os.path.join(\n audit_dir,\n \"cell56_fold2_finetuned_experiment.pt\"\n )\n}\n\nprint()\nprint(\"=\" * 70)\nprint(\"CHECKPOINT AVAILABILITY\")\nprint(\"=\" * 70)\n\nfor fold, path in baseline_checkpoint_paths.items():\n\n if not os.path.exists(path):\n raise RuntimeError(\n f\"Missing baseline Fold-{fold} checkpoint: {path}\"\n )\n\n print(\n f\"Baseline Fold-{fold} checkpoint: EXISTS\"\n )\n\nfor fold, path in finetuned_checkpoint_paths.items():\n\n if not os.path.exists(path):\n raise RuntimeError(\n f\"Missing fine-tuned Fold-{fold} checkpoint: {path}\"\n )\n\n print(\n f\"Fine-tuned Fold-{fold} checkpoint: EXISTS\"\n )\n\nprint(\"All six validated checkpoints: PASS\")\n\n# ----------------------------------------------------------------------\n# SAVE FINAL PLAN\n# ----------------------------------------------------------------------\n\nfinal_plan_path = os.path.join(\n audit_dir,\n \"cell61_final_oof_backed_model_plan.csv\"\n)\n\nmerged_plan.to_csv(\n final_plan_path,\n index=False\n)\n\nif not os.path.exists(final_plan_path):\n raise RuntimeError(\n \"Final OOF-backed model plan was not saved.\"\n )\n\n# ----------------------------------------------------------------------\n# FINAL VERIFICATION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 61 VERIFICATION\")\nprint(\"=\" * 70)\n\nprint(\"12 targets covered: PASS\")\nprint(\"OOF-derived model decisions preserved: PASS\")\nprint(\"OOF-derived thresholds preserved: PASS\")\nprint(\"Baseline checkpoints verified: PASS\")\nprint(\"Fine-tuned checkpoints verified: PASS\")\nprint(\"No new training performed: PASS\")\nprint(\"No checkpoint modified: PASS\")\nprint(\"No test data used: PASS\")\n\nprint()\nprint(\"=\" * 70)\nprint(\"OUTPUT\")\nprint(\"=\" * 70)\n\nprint(\n f\"Saved: {final_plan_path}\"\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 61 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Final OOF-backed model plan is ready.\"\n)\n\nprint(\n \"Send the complete Cell 61 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:33:06.217536Z","iopub.execute_input":"2026-08-11T11:33:06.217914Z","iopub.status.idle":"2026-08-11T11:33:06.263371Z","shell.execute_reply.started":"2026-08-11T11:33:06.217863Z","shell.execute_reply":"2026-08-11T11:33:06.262098Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 62 - FINAL CHECKPOINT LOADING + INFERENCE ARCHITECTURE AUDIT\n# ======================================================================\n\nimport os\nimport copy\nimport torch\nimport pandas as pd\nimport numpy as np\n\nprint(\"=\" * 70)\nprint(\"CELL 62 - FINAL CHECKPOINT LOADING + INFERENCE ARCHITECTURE AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"model\",\n \"TARGETS\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\n\n# ----------------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------------\n\nif \"device\" in globals():\n inference_device = device\nelse:\n inference_device = torch.device(\n \"cuda\" if torch.cuda.is_available() else \"cpu\"\n )\n\nprint(f\"Device: {inference_device}\")\n\n# ----------------------------------------------------------------------\n# BASIC MODEL VALIDATION\n# ----------------------------------------------------------------------\n\nif not isinstance(model, torch.nn.Module):\n raise RuntimeError(\n \"Existing model object is not a torch.nn.Module.\"\n )\n\nif len(TARGETS) != 12:\n raise RuntimeError(\n f\"Expected 12 targets, found {len(TARGETS)}.\"\n )\n\nmodel = model.to(inference_device)\n\nprint(\"Existing model: PASS\")\nprint(\"12-target configuration: PASS\")\n\n# ----------------------------------------------------------------------\n# AUDIT DIRECTORY\n# ----------------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nif not os.path.isdir(audit_dir):\n raise RuntimeError(\n f\"Audit directory missing: {audit_dir}\"\n )\n\nprint(\"Audit directory: PASS\")\n\n# ----------------------------------------------------------------------\n# FINAL PLAN\n# ----------------------------------------------------------------------\n\nplan_path = os.path.join(\n audit_dir,\n \"cell61_final_oof_backed_model_plan.csv\"\n)\n\nif not os.path.exists(plan_path):\n raise RuntimeError(\n f\"Final model plan missing: {plan_path}\"\n )\n\nfinal_plan = pd.read_csv(plan_path)\n\nif len(final_plan) != 12:\n raise RuntimeError(\n \"Final model plan must contain exactly 12 targets.\"\n )\n\nrequired_plan_columns = [\n \"target\",\n \"selected_model\",\n \"selected_threshold\",\n \"selected_best_f1\"\n]\n\nmissing_plan_columns = [\n c\n for c in required_plan_columns\n if c not in final_plan.columns\n]\n\nif missing_plan_columns:\n raise RuntimeError(\n \"Final model plan missing columns: \"\n + \", \".join(missing_plan_columns)\n )\n\nif list(final_plan[\"target\"]) != list(TARGETS):\n raise RuntimeError(\n \"Final model plan target order does not match TARGETS.\"\n )\n\nprint(\"Final OOF-backed model plan: PASS\")\nprint(\"Final target ordering: PASS\")\n\n# ----------------------------------------------------------------------\n# CHECKPOINT PATHS\n# ----------------------------------------------------------------------\n\ncheckpoint_paths = {\n \"baseline_fold0\": os.path.join(\n audit_dir,\n \"cell37_fold0_best_model.pt\"\n ),\n \"baseline_fold1\": os.path.join(\n audit_dir,\n \"cell40_fold1_best_model.pt\"\n ),\n \"baseline_fold2\": os.path.join(\n audit_dir,\n \"cell44_fold2_best_model.pt\"\n ),\n \"finetuned_fold0\": os.path.join(\n audit_dir,\n \"cell52_fold0_finetuned_experiment.pt\"\n ),\n \"finetuned_fold1\": os.path.join(\n audit_dir,\n \"cell54_fold1_finetuned_experiment.pt\"\n ),\n \"finetuned_fold2\": os.path.join(\n audit_dir,\n \"cell56_fold2_finetuned_experiment.pt\"\n )\n}\n\nprint()\nprint(\"=\" * 70)\nprint(\"CHECKPOINT FILE VALIDATION\")\nprint(\"=\" * 70)\n\nfor name, path in checkpoint_paths.items():\n\n if not os.path.exists(path):\n raise RuntimeError(\n f\"Missing checkpoint: {path}\"\n )\n\n if os.path.getsize(path) <= 0:\n raise RuntimeError(\n f\"Checkpoint is empty: {path}\"\n )\n\n print(f\"{name}: EXISTS\")\n\nprint(\"Six checkpoint files: PASS\")\n\n# ----------------------------------------------------------------------\n# CHECKPOINT STATE-DICT EXTRACTION\n# ----------------------------------------------------------------------\n\ndef extract_state_dict(checkpoint, checkpoint_name):\n \"\"\"\n Extract a PyTorch state_dict without assuming a single\n checkpoint wrapper format.\n \"\"\"\n\n if isinstance(checkpoint, dict):\n\n if \"state_dict\" in checkpoint:\n state_dict = checkpoint[\"state_dict\"]\n\n elif \"model_state_dict\" in checkpoint:\n state_dict = checkpoint[\"model_state_dict\"]\n\n elif \"model\" in checkpoint and isinstance(\n checkpoint[\"model\"],\n dict\n ):\n state_dict = checkpoint[\"model\"]\n\n else:\n # Some checkpoints are themselves state_dict dictionaries.\n if all(\n isinstance(k, str)\n and isinstance(v, torch.Tensor)\n for k, v in checkpoint.items()\n ):\n state_dict = checkpoint\n else:\n raise RuntimeError(\n f\"{checkpoint_name}: unable to identify state_dict.\"\n )\n\n else:\n raise RuntimeError(\n f\"{checkpoint_name}: unsupported checkpoint type \"\n f\"{type(checkpoint)}.\"\n )\n\n if not isinstance(state_dict, dict):\n raise RuntimeError(\n f\"{checkpoint_name}: extracted state_dict is invalid.\"\n )\n\n if len(state_dict) == 0:\n raise RuntimeError(\n f\"{checkpoint_name}: state_dict is empty.\"\n )\n\n return state_dict\n\n# ----------------------------------------------------------------------\n# STATE-DICT COMPATIBILITY CHECK\n# ----------------------------------------------------------------------\n\ntemplate_state = model.state_dict()\n\nloaded_models = {}\n\nprint()\nprint(\"=\" * 70)\nprint(\"CHECKPOINT COMPATIBILITY\")\nprint(\"=\" * 70)\n\nfor checkpoint_name, checkpoint_path in checkpoint_paths.items():\n\n checkpoint = torch.load(\n checkpoint_path,\n map_location=inference_device\n )\n\n state_dict = extract_state_dict(\n checkpoint,\n checkpoint_name\n )\n\n # Handle DataParallel-style \"module.\" prefixes if present.\n normalized_state_dict = {}\n\n for key, value in state_dict.items():\n\n normalized_key = key\n\n if normalized_key.startswith(\"module.\"):\n normalized_key = normalized_key[len(\"module.\"):]\n\n normalized_state_dict[normalized_key] = value\n\n checkpoint_keys = set(normalized_state_dict.keys())\n model_keys = set(template_state.keys())\n\n missing_keys = sorted(\n model_keys - checkpoint_keys\n )\n\n unexpected_keys = sorted(\n checkpoint_keys - model_keys\n )\n\n if missing_keys or unexpected_keys:\n raise RuntimeError(\n f\"{checkpoint_name}: state_dict incompatible.\\n\"\n f\"Missing keys: {missing_keys[:10]}\\n\"\n f\"Unexpected keys: {unexpected_keys[:10]}\"\n )\n\n # Verify tensor shapes before loading.\n shape_mismatches = []\n\n for key in model_keys:\n\n model_shape = tuple(\n template_state[key].shape\n )\n\n checkpoint_shape = tuple(\n normalized_state_dict[key].shape\n )\n\n if model_shape != checkpoint_shape:\n shape_mismatches.append(\n (\n key,\n model_shape,\n checkpoint_shape\n )\n )\n\n if shape_mismatches:\n raise RuntimeError(\n f\"{checkpoint_name}: parameter shape mismatch: \"\n + str(shape_mismatches[:5])\n )\n\n # Create an independent model copy.\n candidate_model = copy.deepcopy(model)\n\n candidate_model.load_state_dict(\n normalized_state_dict,\n strict=True\n )\n\n candidate_model = candidate_model.to(\n inference_device\n )\n\n candidate_model.eval()\n\n # Parameter validity.\n invalid_parameters = False\n\n for parameter in candidate_model.parameters():\n\n if not torch.isfinite(\n parameter\n ).all():\n invalid_parameters = True\n break\n\n if invalid_parameters:\n raise RuntimeError(\n f\"{checkpoint_name}: checkpoint contains \"\n \"NaN/Inf parameters.\"\n )\n\n loaded_models[checkpoint_name] = candidate_model\n\n print(\n f\"{checkpoint_name}: \"\n \"state_dict compatibility PASS | \"\n \"model load PASS | \"\n \"parameter validity PASS\"\n )\n\nprint(\"All six checkpoints loaded successfully: PASS\")\n\n# ----------------------------------------------------------------------\n# MODEL ARCHITECTURE VERIFICATION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"ARCHITECTURE VERIFICATION\")\nprint(\"=\" * 70)\n\n# Use one known checkpoint model for a synthetic forward test.\n# This does NOT use any dataset or test image.\n\ntest_model = loaded_models[\"baseline_fold0\"]\n\nwith torch.no_grad():\n\n synthetic_input = torch.zeros(\n 1,\n 21,\n 224,\n 224,\n dtype=torch.float32,\n device=inference_device\n )\n\n synthetic_output = test_model(\n synthetic_input\n )\n\nif not isinstance(\n synthetic_output,\n torch.Tensor\n):\n raise RuntimeError(\n \"Model forward output is not a torch.Tensor.\"\n )\n\nif tuple(synthetic_output.shape) != (1, 12):\n raise RuntimeError(\n \"Unexpected model output shape: \"\n + str(tuple(synthetic_output.shape))\n )\n\nif not torch.isfinite(\n synthetic_output\n).all():\n raise RuntimeError(\n \"Synthetic forward output contains NaN/Inf.\"\n )\n\nprint(\"Synthetic input shape: (1, 21, 224, 224)\")\nprint(\"Synthetic output shape:\", tuple(synthetic_output.shape))\nprint(\"21-channel input: PASS\")\nprint(\"12-target output: PASS\")\nprint(\"Forward pass: PASS\")\nprint(\"Output validity: PASS\")\n\n# ----------------------------------------------------------------------\n# MODEL TYPE SUMMARY\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"LOADED MODEL SUMMARY\")\nprint(\"=\" * 70)\n\nfor model_name, loaded_model in loaded_models.items():\n\n total_parameters = sum(\n parameter.numel()\n for parameter in loaded_model.parameters()\n )\n\n print(\n f\"{model_name}: \"\n f\"parameters={total_parameters:,} | \"\n f\"mode={'eval' if not loaded_model.training else 'train'}\"\n )\n\n# ----------------------------------------------------------------------\n# FINAL PLAN / CHECKPOINT CONSISTENCY\n# ----------------------------------------------------------------------\n\nbaseline_targets = final_plan.loc[\n final_plan[\"selected_model\"] == \"baseline\",\n \"target\"\n].tolist()\n\nfinetuned_targets = final_plan.loc[\n final_plan[\"selected_model\"] == \"fine_tuned\",\n \"target\"\n].tolist()\n\nif len(baseline_targets) != 5:\n raise RuntimeError(\n f\"Expected 5 baseline-selected targets, found \"\n f\"{len(baseline_targets)}.\"\n )\n\nif len(finetuned_targets) != 7:\n raise RuntimeError(\n f\"Expected 7 fine-tuned-selected targets, found \"\n f\"{len(finetuned_targets)}.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL MODEL PLAN CONSISTENCY\")\nprint(\"=\" * 70)\n\nprint(\"Baseline-selected targets: 5/12\")\nprint(\"Fine-tuned-selected targets: 7/12\")\nprint(\"All 12 targets assigned exactly once: PASS\")\nprint(\"OOF thresholds preserved: PASS\")\n\n# ----------------------------------------------------------------------\n# IMPORTANT: NO DATASET INFERENCE\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"DATA SAFETY CHECK\")\nprint(\"=\" * 70)\n\nprint(\"No training performed.\")\nprint(\"No validation dataset inference performed.\")\nprint(\"No test data accessed.\")\nprint(\"No checkpoint modified.\")\nprint(\"No checkpoint overwritten.\")\n\n# ----------------------------------------------------------------------\n# STORE LOADED MODELS FOR NEXT CELL\n# ----------------------------------------------------------------------\n\nfinal_inference_models = loaded_models\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 62 VERDICT\")\nprint(\"=\" * 70)\n\nprint(\"Six validated checkpoints loaded: PASS\")\nprint(\"Baseline checkpoints: PASS\")\nprint(\"Fine-tuned checkpoints: PASS\")\nprint(\"State-dict compatibility: PASS\")\nprint(\"21-channel input: PASS\")\nprint(\"12-target output: PASS\")\nprint(\"Synthetic forward pass: PASS\")\nprint(\"Final OOF-backed target plan: PASS\")\nprint(\"Inference model objects prepared: PASS\")\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 62 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Final inference architecture is ready.\"\n)\n\nprint(\n \"No dataset or test inference was performed.\"\n)\n\nprint(\n \"Send the complete Cell 62 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:35:00.720145Z","iopub.execute_input":"2026-08-11T11:35:00.720949Z","iopub.status.idle":"2026-08-11T11:35:01.430436Z","shell.execute_reply.started":"2026-08-11T11:35:00.720911Z","shell.execute_reply":"2026-08-11T11:35:01.42938Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 63 - FINAL 3-FOLD INFERENCE PLAN AUDIT\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"CELL 63 - FINAL 3-FOLD INFERENCE PLAN AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"TARGETS\",\n \"final_inference_models\"\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\n\n# ----------------------------------------------------------------------\n# TARGET VALIDATION\n# ----------------------------------------------------------------------\n\nif len(TARGETS) != 12:\n raise RuntimeError(\n f\"Expected 12 targets, found {len(TARGETS)}.\"\n )\n\nprint(f\"Targets: {len(TARGETS)}\")\n\n# ----------------------------------------------------------------------\n# LOADED MODEL VALIDATION\n# ----------------------------------------------------------------------\n\nexpected_model_names = {\n \"baseline_fold0\",\n \"baseline_fold1\",\n \"baseline_fold2\",\n \"finetuned_fold0\",\n \"finetuned_fold1\",\n \"finetuned_fold2\"\n}\n\nactual_model_names = set(\n final_inference_models.keys()\n)\n\nif actual_model_names != expected_model_names:\n raise RuntimeError(\n \"Loaded model set mismatch.\\n\"\n f\"Expected: {sorted(expected_model_names)}\\n\"\n f\"Found: {sorted(actual_model_names)}\"\n )\n\nprint(\"Six inference models: PASS\")\n\n# ----------------------------------------------------------------------\n# MODEL MODE VALIDATION\n# ----------------------------------------------------------------------\n\nfor model_name, loaded_model in final_inference_models.items():\n\n if loaded_model.training:\n raise RuntimeError(\n f\"{model_name} is still in training mode.\"\n )\n\nprint(\"All inference models in evaluation mode: PASS\")\n\n# ----------------------------------------------------------------------\n# LOAD FINAL OOF-BACKED PLAN\n# ----------------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nplan_path = os.path.join(\n audit_dir,\n \"cell61_final_oof_backed_model_plan.csv\"\n)\n\nif not os.path.exists(plan_path):\n raise RuntimeError(\n f\"Missing final OOF-backed plan: {plan_path}\"\n )\n\nfinal_plan = pd.read_csv(plan_path)\n\nif len(final_plan) != 12:\n raise RuntimeError(\n \"Final plan must contain exactly 12 targets.\"\n )\n\nrequired_columns = [\n \"target\",\n \"selected_model\",\n \"selected_threshold\",\n \"selected_best_f1\"\n]\n\nmissing_columns = [\n column\n for column in required_columns\n if column not in final_plan.columns\n]\n\nif missing_columns:\n raise RuntimeError(\n \"Final plan missing columns: \"\n + \", \".join(missing_columns)\n )\n\nif list(final_plan[\"target\"]) != list(TARGETS):\n raise RuntimeError(\n \"Final plan target order does not match TARGETS.\"\n )\n\nprint(\"Final OOF-backed plan: PASS\")\nprint(\"Target ordering: PASS\")\n\n# ----------------------------------------------------------------------\n# VALIDATE MODEL ASSIGNMENTS\n# ----------------------------------------------------------------------\n\nvalid_model_assignments = {\n \"baseline\",\n \"fine_tuned\"\n}\n\ninvalid_assignments = [\n value\n for value in final_plan[\"selected_model\"]\n if value not in valid_model_assignments\n]\n\nif invalid_assignments:\n raise RuntimeError(\n \"Invalid model assignments: \"\n + \", \".join(\n sorted(set(invalid_assignments))\n )\n )\n\nprint(\"Target model assignments: PASS\")\n\n# ----------------------------------------------------------------------\n# VALIDATE THRESHOLDS\n# ----------------------------------------------------------------------\n\nthresholds = pd.to_numeric(\n final_plan[\"selected_threshold\"],\n errors=\"coerce\"\n)\n\nif thresholds.isna().any():\n raise RuntimeError(\n \"Selected thresholds contain invalid values.\"\n )\n\nif not thresholds.between(\n 0.10,\n 0.90\n).all():\n raise RuntimeError(\n \"Selected thresholds fall outside \"\n \"the audited 0.10-0.90 range.\"\n )\n\nprint(\"Threshold range: PASS\")\n\n# ----------------------------------------------------------------------\n# BUILD EXPLICIT THREE-FOLD MODEL MAP\n# ----------------------------------------------------------------------\n\nmodel_map = {}\n\nfor _, row in final_plan.iterrows():\n\n target = row[\"target\"]\n selected_model = row[\"selected_model\"]\n threshold = float(\n row[\"selected_threshold\"]\n )\n\n if selected_model == \"baseline\":\n\n fold_models = [\n \"baseline_fold0\",\n \"baseline_fold1\",\n \"baseline_fold2\"\n ]\n\n elif selected_model == \"fine_tuned\":\n\n fold_models = [\n \"finetuned_fold0\",\n \"finetuned_fold1\",\n \"finetuned_fold2\"\n ]\n\n else:\n raise RuntimeError(\n f\"Unsupported selected model: {selected_model}\"\n )\n\n model_map[target] = {\n \"selected_model\": selected_model,\n \"threshold\": threshold,\n \"fold_models\": fold_models\n }\n\n# ----------------------------------------------------------------------\n# VERIFY EVERY TARGET HAS THREE FOLD MODELS\n# ----------------------------------------------------------------------\n\nfor target in TARGETS:\n\n if target not in model_map:\n raise RuntimeError(\n f\"Missing inference plan for target: {target}\"\n )\n\n fold_models = model_map[target][\"fold_models\"]\n\n if len(fold_models) != 3:\n raise RuntimeError(\n f\"{target}: expected 3 fold models.\"\n )\n\n for fold_model in fold_models:\n\n if fold_model not in final_inference_models:\n raise RuntimeError(\n f\"{target}: missing loaded model \"\n f\"{fold_model}.\"\n )\n\nprint(\"Every target has exactly three fold models: PASS\")\n\n# ----------------------------------------------------------------------\n# VERIFY BASELINE / FINE-TUNED COUNTS\n# ----------------------------------------------------------------------\n\nbaseline_targets = [\n target\n for target in TARGETS\n if model_map[target][\"selected_model\"] == \"baseline\"\n]\n\nfinetuned_targets = [\n target\n for target in TARGETS\n if model_map[target][\"selected_model\"] == \"fine_tuned\"\n]\n\nif len(baseline_targets) != 5:\n raise RuntimeError(\n f\"Expected 5 baseline targets, found \"\n f\"{len(baseline_targets)}.\"\n )\n\nif len(finetuned_targets) != 7:\n raise RuntimeError(\n f\"Expected 7 fine-tuned targets, found \"\n f\"{len(finetuned_targets)}.\"\n )\n\nprint(\"Baseline target count: 5\")\nprint(\"Fine-tuned target count: 7\")\nprint(\"Model-family assignment: PASS\")\n\n# ----------------------------------------------------------------------\n# PRINT FINAL INFERENCE PLAN\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL THREE-FOLD INFERENCE PLAN\")\nprint(\"=\" * 70)\n\nplan_rows = []\n\nfor target in TARGETS:\n\n entry = model_map[target]\n\n plan_rows.append({\n \"target\": target,\n \"selected_model\": entry[\"selected_model\"],\n \"threshold\": entry[\"threshold\"],\n \"fold0_model\": entry[\"fold_models\"][0],\n \"fold1_model\": entry[\"fold_models\"][1],\n \"fold2_model\": entry[\"fold_models\"][2]\n })\n\ninference_plan_table = pd.DataFrame(\n plan_rows\n)\n\nprint(\n inference_plan_table.to_string(\n index=False,\n float_format=lambda x: f\"{x:.4f}\"\n )\n)\n\n# ----------------------------------------------------------------------\n# SAVE INFERENCE PLAN\n# ----------------------------------------------------------------------\n\ninference_plan_path = os.path.join(\n audit_dir,\n \"cell63_final_three_fold_inference_plan.csv\"\n)\n\ninference_plan_table.to_csv(\n inference_plan_path,\n index=False\n)\n\nif not os.path.exists(inference_plan_path):\n raise RuntimeError(\n \"Failed to save final three-fold inference plan.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 63 VERIFICATION\")\nprint(\"=\" * 70)\n\nprint(\"12 targets covered: PASS\")\nprint(\"Three folds assigned per target: PASS\")\nprint(\"Baseline model family mapping: PASS\")\nprint(\"Fine-tuned model family mapping: PASS\")\nprint(\"OOF thresholds preserved: PASS\")\nprint(\"All six loaded checkpoints referenced: PASS\")\nprint(\"No training performed: PASS\")\nprint(\"No validation inference performed: PASS\")\nprint(\"No test data accessed: PASS\")\nprint(\"No checkpoint modified: PASS\")\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 63 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n f\"Saved: {inference_plan_path}\"\n)\n\nprint(\n \"Final three-fold inference plan is ready.\"\n)\n\nprint(\n \"Send the complete Cell 63 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:38:00.227447Z","iopub.execute_input":"2026-08-11T11:38:00.22777Z","iopub.status.idle":"2026-08-11T11:38:00.265398Z","shell.execute_reply.started":"2026-08-11T11:38:00.227741Z","shell.execute_reply":"2026-08-11T11:38:00.264231Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 64R - TEST SOURCE RESTORATION\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"CELL 64R - TEST SOURCE RESTORATION\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"TARGETS\",\n \"KneeStudyDataset\",\n \"test\",\n \"test_series\",\n \"test_slice_audit\",\n \"TEST_SERIES_DIR\",\n \"TEST_SERIES_CSV\",\n]\n\nmissing_objects = [\n name for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\n\n# ----------------------------------------------------------------------\n# ESTABLISHED SCHEMA\n# ----------------------------------------------------------------------\n\nIDENTITY_COLUMN = \"StudyInstanceUID\"\n\nPRIMARY_SERIES_COLUMNS = [\n \"Sagittal_SeriesInstanceUID\",\n \"Coronal_SeriesInstanceUID\",\n \"Axial_SeriesInstanceUID\",\n]\n\nTEST_SERIES_COLUMNS = [\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\",\n \"Anatomical_Plane\",\n]\n\nprint()\nprint(\"=\" * 70)\nprint(\"ESTABLISHED TEST SCHEMA\")\nprint(\"=\" * 70)\n\nprint(\"Identity:\", IDENTITY_COLUMN)\n\nfor column in PRIMARY_SERIES_COLUMNS:\n print(\"Primary series:\", column)\n\n# ----------------------------------------------------------------------\n# VALIDATE ACTUAL TEST SOURCE\n# ----------------------------------------------------------------------\n\nif not isinstance(test, pd.DataFrame):\n raise RuntimeError(\n \"Notebook object 'test' is not a DataFrame.\"\n )\n\nif not isinstance(test_series, pd.DataFrame):\n raise RuntimeError(\n \"Notebook object 'test_series' is not a DataFrame.\"\n )\n\nif not isinstance(test_slice_audit, pd.DataFrame):\n raise RuntimeError(\n \"Notebook object 'test_slice_audit' is not a DataFrame.\"\n )\n\nmissing_test_columns = [\n column\n for column in [IDENTITY_COLUMN]\n if column not in test.columns\n]\n\nif missing_test_columns:\n raise RuntimeError(\n \"test is missing required columns: \"\n + \", \".join(missing_test_columns)\n )\n\nmissing_series_columns = [\n column\n for column in TEST_SERIES_COLUMNS\n if column not in test_series.columns\n]\n\nif missing_series_columns:\n raise RuntimeError(\n \"test_series is missing required columns: \"\n + \", \".join(missing_series_columns)\n )\n\nprint(\"test DataFrame: PASS\")\nprint(\"test_series DataFrame: PASS\")\nprint(\"test_slice_audit DataFrame: PASS\")\n\n# ----------------------------------------------------------------------\n# NORMALIZE IDENTIFIERS\n# ----------------------------------------------------------------------\n\ntest_ids = (\n test[IDENTITY_COLUMN]\n .astype(str)\n .str.strip()\n)\n\ntest_series_work = test_series.copy()\n\ntest_series_work[IDENTITY_COLUMN] = (\n test_series_work[IDENTITY_COLUMN]\n .astype(str)\n .str.strip()\n)\n\ntest_series_work[\"SeriesInstanceUID\"] = (\n test_series_work[\"SeriesInstanceUID\"]\n .astype(str)\n .str.strip()\n)\n\n# Preserve exact test.csv ordering.\ntest_id_order = {\n study_id: index\n for index, study_id in enumerate(test_ids.tolist())\n}\n\ntest_id_set = set(test_ids.tolist())\n\nprint()\nprint(\"=\" * 70)\nprint(\"ACTUAL TEST SOURCE\")\nprint(\"=\" * 70)\n\nprint(\"TEST_SERIES_DIR:\", TEST_SERIES_DIR)\nprint(\"TEST_SERIES_CSV:\", TEST_SERIES_CSV)\nprint(\"Test studies from test:\", len(test_id_set))\nprint(\"Rows in test_series:\", len(test_series_work))\n\n# ----------------------------------------------------------------------\n# ISOLATE ONLY THE ACTUAL TEST STUDIES\n#\n# Do not require test_series to contain only test.csv studies globally.\n# The earlier 64S failure came from over-restrictive source\n# classification. We only need to safely extract the requested test IDs.\n# ----------------------------------------------------------------------\n\ntest_candidates = test_series_work[\n test_series_work[IDENTITY_COLUMN].isin(test_id_set)\n].copy()\n\nif test_candidates.empty:\n raise RuntimeError(\n \"No rows from test_series matched the StudyInstanceUIDs \"\n \"in test.\"\n )\n\nmatched_test_ids = set(\n test_candidates[IDENTITY_COLUMN]\n)\n\nmissing_test_studies = (\n test_id_set - matched_test_ids\n)\n\nprint()\nprint(\"Matched test studies:\", len(matched_test_ids))\n\nif missing_test_studies:\n print(\n \"Studies from test.csv without test_series rows:\",\n sorted(missing_test_studies),\n )\n raise RuntimeError(\n \"Actual test studies are missing from test_series. \"\n \"Do not fabricate test metadata.\"\n )\n\nprint(\"Test study-to-series mapping: PASS\")\n\n# ----------------------------------------------------------------------\n# TEST SLICE-COUNT INFORMATION\n# ----------------------------------------------------------------------\n\nslice_audit = test_slice_audit.copy()\n\nslice_key_columns = [\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n]\n\nmissing_slice_keys = [\n column\n for column in slice_key_columns\n if column not in slice_audit.columns\n]\n\nif missing_slice_keys:\n raise RuntimeError(\n \"test_slice_audit is missing: \"\n + \", \".join(missing_slice_keys)\n )\n\nslice_audit[IDENTITY_COLUMN] = (\n slice_audit[IDENTITY_COLUMN]\n .astype(str)\n .str.strip()\n)\n\nslice_audit[\"SeriesInstanceUID\"] = (\n slice_audit[\"SeriesInstanceUID\"]\n .astype(str)\n .str.strip()\n)\n\nif \"slice_count\" in slice_audit.columns:\n\n slice_counts = slice_audit[\n slice_key_columns + [\"slice_count\"]\n ].copy()\n\n slice_counts[\"slice_count\"] = pd.to_numeric(\n slice_counts[\"slice_count\"],\n errors=\"coerce\",\n )\n\n # Avoid accidental many-to-many merge.\n slice_counts = (\n slice_counts\n .drop_duplicates(\n subset=slice_key_columns,\n keep=\"first\",\n )\n )\n\n test_candidates = test_candidates.merge(\n slice_counts,\n on=slice_key_columns,\n how=\"left\",\n validate=\"one_to_one\",\n )\n\nelse:\n # The test-series metadata itself is still sufficient for\n # identifying the series. Do not fabricate slice counts.\n test_candidates[\"slice_count\"] = np.nan\n\nprint(\"Test slice metadata handling: PASS\")\n\n# ----------------------------------------------------------------------\n# PLANE NORMALIZATION\n# ----------------------------------------------------------------------\n\ntest_candidates[\"Anatomical_Plane\"] = (\n test_candidates[\"Anatomical_Plane\"]\n .astype(str)\n .str.strip()\n)\n\nexpected_planes = [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\",\n]\n\n# ----------------------------------------------------------------------\n# SHOW ACTUAL CANDIDATES\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"ACTUAL TEST SERIES CANDIDATES\")\nprint(\"=\" * 70)\n\ncandidate_display_columns = [\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\",\n \"slice_count\",\n]\n\navailable_display_columns = [\n column\n for column in candidate_display_columns\n if column in test_candidates.columns\n]\n\nprint(\n test_candidates[\n available_display_columns\n ]\n .sort_values(\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\",\n \"SeriesInstanceUID\",\n ]\n )\n .to_string(index=False)\n)\n\n# ----------------------------------------------------------------------\n# VALIDATE THREE-PLANE COVERAGE\n# ----------------------------------------------------------------------\n\ncoverage_rows = []\n\nfor study_id in test_ids.tolist():\n\n study_candidates = test_candidates[\n test_candidates[IDENTITY_COLUMN] == study_id\n ]\n\n observed_planes = set(\n study_candidates[\"Anatomical_Plane\"]\n )\n\n missing_planes = [\n plane\n for plane in expected_planes\n if plane not in observed_planes\n ]\n\n coverage_rows.append(\n {\n \"StudyInstanceUID\": study_id,\n \"Sagittal\": int(\n \"Sagittal\" in observed_planes\n ),\n \"Coronal\": int(\n \"Coronal\" in observed_planes\n ),\n \"Axial\": int(\n \"Axial\" in observed_planes\n ),\n \"missing_planes\": (\n \",\".join(missing_planes)\n if missing_planes\n else \"\"\n ),\n }\n )\n\ntest_plane_coverage = pd.DataFrame(\n coverage_rows\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"TEST THREE-PLANE COVERAGE\")\nprint(\"=\" * 70)\n\nprint(\n test_plane_coverage.to_string(\n index=False\n )\n)\n\nif (\n test_plane_coverage[\n \"missing_planes\"\n ].astype(str).str.len() > 0\n).any():\n\n raise RuntimeError(\n \"At least one test study is missing a required \"\n \"primary plane. Do not fabricate a series.\"\n )\n\nprint(\"Three-plane coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# PRIMARY SERIES SELECTION\n#\n# Use the same type of evidence already established in the notebook:\n# 1. Fluid-sensitive preference\n# 2. Fat-suppression preference\n# 3. Slice count when available\n# 4. SeriesInstanceUID only as a deterministic final ordering key\n#\n# No labels or fold information are involved.\n# ----------------------------------------------------------------------\n\ntest_candidates[\"_fluid_score\"] = pd.to_numeric(\n test_candidates[\"Fluid_Sensitive\"],\n errors=\"coerce\",\n).fillna(0)\n\ntest_candidates[\"_fat_score\"] = pd.to_numeric(\n test_candidates[\"Fat_Suppression\"],\n errors=\"coerce\",\n).fillna(0)\n\ntest_candidates[\"_slice_score\"] = pd.to_numeric(\n test_candidates[\"slice_count\"],\n errors=\"coerce\",\n).fillna(-1)\n\nselected_rows = []\n\nfor study_id in test_ids.tolist():\n\n study_candidates = test_candidates[\n test_candidates[IDENTITY_COLUMN] == study_id\n ].copy()\n\n for plane in expected_planes:\n\n plane_candidates = study_candidates[\n study_candidates[\"Anatomical_Plane\"] == plane\n ].copy()\n\n if plane_candidates.empty:\n raise RuntimeError(\n f\"No {plane} series found for study {study_id}.\"\n )\n\n # Deterministic ranking.\n plane_candidates = plane_candidates.sort_values(\n [\n \"_fluid_score\",\n \"_fat_score\",\n \"_slice_score\",\n \"SeriesInstanceUID\",\n ],\n ascending=[\n False,\n False,\n False,\n True,\n ],\n ).reset_index(drop=True)\n\n selected_rows.append(\n plane_candidates.iloc[0].copy()\n )\n\nselected_test_series = pd.DataFrame(\n selected_rows\n)\n\n# ----------------------------------------------------------------------\n# VERIFY EXACTLY ONE SELECTED SERIES PER STUDY / PLANE\n# ----------------------------------------------------------------------\n\nselection_counts = (\n selected_test_series\n .groupby(\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\",\n ]\n )\n .size()\n)\n\nif not bool(\n (selection_counts == 1).all()\n):\n raise RuntimeError(\n \"Primary-series selection produced duplicate \"\n \"study/plane assignments.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"SELECTED PRIMARY TEST SERIES\")\nprint(\"=\" * 70)\n\nprint(\n selected_test_series[\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\",\n \"SeriesInstanceUID\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\",\n \"slice_count\",\n ]\n ].to_string(index=False)\n)\n\nprint(\"One selected series per study/plane: PASS\")\n\n# ----------------------------------------------------------------------\n# PIVOT INTO THE ESTABLISHED 4-COLUMN TEST METADATA\n# ----------------------------------------------------------------------\n\ntest_metadata = (\n selected_test_series\n .pivot(\n index=\"StudyInstanceUID\",\n columns=\"Anatomical_Plane\",\n values=\"SeriesInstanceUID\",\n )\n .reset_index()\n)\n\nmissing_planes_after_pivot = [\n plane\n for plane in expected_planes\n if plane not in test_metadata.columns\n]\n\nif missing_planes_after_pivot:\n raise RuntimeError(\n \"Pivot failed to produce: \"\n + \", \".join(missing_planes_after_pivot)\n )\n\ntest_metadata = test_metadata.rename(\n columns={\n \"Sagittal\":\n \"Sagittal_SeriesInstanceUID\",\n \"Coronal\":\n \"Coronal_SeriesInstanceUID\",\n \"Axial\":\n \"Axial_SeriesInstanceUID\",\n }\n)\n\ntest_metadata = test_metadata[\n [\n \"StudyInstanceUID\",\n \"Sagittal_SeriesInstanceUID\",\n \"Coronal_SeriesInstanceUID\",\n \"Axial_SeriesInstanceUID\",\n ]\n].copy()\n\n# Preserve test.csv order.\ntest_metadata[\"_order\"] = (\n test_metadata[\"StudyInstanceUID\"]\n .map(test_id_order)\n)\n\nif test_metadata[\"_order\"].isna().any():\n raise RuntimeError(\n \"Could not restore test.csv study ordering.\"\n )\n\ntest_metadata = (\n test_metadata\n .sort_values(\"_order\")\n .drop(columns=\"_order\")\n .reset_index(drop=True)\n)\n\n# ----------------------------------------------------------------------\n# FINAL VALIDATION\n# ----------------------------------------------------------------------\n\nexpected_test_rows = len(test)\n\nif len(test_metadata) != expected_test_rows:\n raise RuntimeError(\n \"Final test metadata row count mismatch. \"\n f\"Expected {expected_test_rows}, \"\n f\"got {len(test_metadata)}.\"\n )\n\nif (\n test_metadata[\"StudyInstanceUID\"]\n .duplicated()\n .any()\n):\n raise RuntimeError(\n \"Duplicate StudyInstanceUID values in final test metadata.\"\n )\n\nfor column in PRIMARY_SERIES_COLUMNS:\n\n if test_metadata[column].isna().any():\n raise RuntimeError(\n f\"Missing values in {column}.\"\n )\n\n if (\n test_metadata[column]\n .astype(str)\n .str.strip()\n .eq(\"\")\n .any()\n ):\n raise RuntimeError(\n f\"Empty values in {column}.\"\n )\n\nif set(\n test_metadata[\"StudyInstanceUID\"]\n) != test_id_set:\n\n raise RuntimeError(\n \"Final test metadata does not contain exactly \"\n \"the test.csv studies.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL RESTORED TEST METADATA\")\nprint(\"=\" * 70)\n\nprint(\n test_metadata.to_string(index=False)\n)\n\nprint()\nprint(\"Test metadata shape:\", test_metadata.shape)\nprint(\"Expected test studies:\", expected_test_rows)\nprint(\"StudyInstanceUID coverage: PASS\")\nprint(\"Sagittal_SeriesInstanceUID: PASS\")\nprint(\"Coronal_SeriesInstanceUID: PASS\")\nprint(\"Axial_SeriesInstanceUID: PASS\")\nprint(\"One row per test study: PASS\")\n\n# ----------------------------------------------------------------------\n# SAVE\n# ----------------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nos.makedirs(\n audit_dir,\n exist_ok=True,\n)\n\ntest_metadata_path = os.path.join(\n audit_dir,\n \"cell64r_test_primary_series_metadata.csv\",\n)\n\ntest_metadata.to_csv(\n test_metadata_path,\n index=False,\n)\n\n# Expose the canonical object for the next cell.\ntest_metadata_df = test_metadata.copy()\n\n# ----------------------------------------------------------------------\n# VERDICT\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 64R VERDICT\")\nprint(\"=\" * 70)\n\nprint(\"Actual test source recovered: PASS\")\nprint(\"Test studies:\", len(test_metadata_df))\nprint(\"Primary-series metadata: PASS\")\nprint(\"Established schema preserved: PASS\")\nprint(\"No labels fabricated: PASS\")\nprint(\"No fold assignments fabricated: PASS\")\nprint(\"No modeling table modified: PASS\")\nprint(\"No training performed: PASS\")\nprint(\"No checkpoint modified: PASS\")\nprint(\"No test inference performed: PASS\")\n\nprint()\nprint(\"Saved:\")\nprint(test_metadata_path)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 64R COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Test metadata restoration completed.\"\n)\nprint(\n \"Next step: construct the test Dataset using the \"\n \"established KneeStudyDataset pipeline.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:52:27.190234Z","iopub.execute_input":"2026-08-11T11:52:27.190561Z","iopub.status.idle":"2026-08-11T11:52:27.294007Z","shell.execute_reply.started":"2026-08-11T11:52:27.190534Z","shell.execute_reply":"2026-08-11T11:52:27.293037Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 65 - TEST SERIES SELECTION AUDIT\n# ======================================================================\n\nimport inspect\nimport pandas as pd\nimport numpy as np\n\nprint(\"=\" * 70)\nprint(\"CELL 65 - TEST SERIES SELECTION AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"TARGETS\",\n \"KneeStudyDataset\",\n \"test\",\n \"test_series\",\n \"sample_submission\",\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"TARGETS: AVAILABLE\")\nprint(\"KneeStudyDataset: AVAILABLE\")\nprint(\"test: AVAILABLE\")\nprint(\"test_series: AVAILABLE\")\nprint(\"sample_submission: AVAILABLE\")\n\n# ----------------------------------------------------------------------\n# TEST TABLE SHAPES\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TEST SOURCE TABLES\")\nprint(\"=\" * 70)\n\nprint(\"test shape:\", test.shape)\nprint(\"test_series shape:\", test_series.shape)\nprint(\"sample_submission shape:\", sample_submission.shape)\n\n# ----------------------------------------------------------------------\n# EXACT TEST STUDY IDS\n# ----------------------------------------------------------------------\n\nif \"StudyInstanceUID\" not in test.columns:\n raise RuntimeError(\n \"test is missing StudyInstanceUID.\"\n )\n\nif \"StudyInstanceUID\" not in test_series.columns:\n raise RuntimeError(\n \"test_series is missing StudyInstanceUID.\"\n )\n\ntest_ids_from_test = set(\n test[\"StudyInstanceUID\"].astype(str)\n)\n\ntest_ids_from_series = set(\n test_series[\"StudyInstanceUID\"].astype(str)\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"TEST STUDY COVERAGE\")\nprint(\"=\" * 70)\n\nprint(\n \"Studies in test.csv:\",\n len(test_ids_from_test)\n)\n\nprint(\n \"Studies in test_series.csv:\",\n len(test_ids_from_series)\n)\n\nmissing_series_studies = (\n test_ids_from_test\n - test_ids_from_series\n)\n\nextra_series_studies = (\n test_ids_from_series\n - test_ids_from_test\n)\n\nprint(\n \"Test studies missing from test_series:\",\n len(missing_series_studies)\n)\n\nprint(\n \"Extra studies in test_series:\",\n len(extra_series_studies)\n)\n\nif missing_series_studies:\n raise RuntimeError(\n \"Some test.csv studies have no corresponding test_series rows.\"\n )\n\nif extra_series_studies:\n raise RuntimeError(\n \"test_series contains studies not present in test.csv.\"\n )\n\nprint(\"Test study coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# EXACT SERIES SCHEMA\n# ----------------------------------------------------------------------\n\nrequired_series_columns = [\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\",\n \"Anatomical_Plane\",\n]\n\nmissing_series_columns = [\n column\n for column in required_series_columns\n if column not in test_series.columns\n]\n\nif missing_series_columns:\n raise RuntimeError(\n \"test_series missing columns: \"\n + \", \".join(missing_series_columns)\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"TEST SERIES SCHEMA\")\nprint(\"=\" * 70)\n\nfor column in required_series_columns:\n print(\n f\"{column}: PASS\"\n )\n\n# ----------------------------------------------------------------------\n# SERIES ID VALIDATION\n# ----------------------------------------------------------------------\n\nif test_series[\"SeriesInstanceUID\"].isna().any():\n raise RuntimeError(\n \"test_series contains missing SeriesInstanceUID values.\"\n )\n\nif test_series[\"SeriesInstanceUID\"].astype(str).str.strip().eq(\"\").any():\n raise RuntimeError(\n \"test_series contains empty SeriesInstanceUID values.\"\n )\n\nif test_series[\"StudyInstanceUID\"].isna().any():\n raise RuntimeError(\n \"test_series contains missing StudyInstanceUID values.\"\n )\n\nprint(\"SeriesInstanceUID validity: PASS\")\nprint(\"StudyInstanceUID validity: PASS\")\n\n# ----------------------------------------------------------------------\n# DUPLICATE ROW CHECK\n# ----------------------------------------------------------------------\n\nduplicate_series_rows = test_series[\n test_series.duplicated(\n subset=[\n \"StudyInstanceUID\",\n \"SeriesInstanceUID\",\n ],\n keep=False,\n )\n]\n\nprint(\n \"Duplicate StudyInstanceUID/SeriesInstanceUID rows:\",\n len(duplicate_series_rows)\n)\n\nif len(duplicate_series_rows) > 0:\n raise RuntimeError(\n \"Duplicate study/series mappings found in test_series.\"\n )\n\nprint(\"Study/series uniqueness: PASS\")\n\n# ----------------------------------------------------------------------\n# PLANE VALIDATION\n# ----------------------------------------------------------------------\n\nexpected_planes = {\n \"Sagittal\",\n \"Coronal\",\n \"Axial\",\n}\n\nobserved_planes = set(\n test_series[\"Anatomical_Plane\"]\n .dropna()\n .astype(str)\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"ANATOMICAL PLANE VALIDATION\")\nprint(\"=\" * 70)\n\nprint(\n \"Observed planes:\",\n sorted(observed_planes)\n)\n\nunexpected_planes = (\n observed_planes\n - expected_planes\n)\n\nif unexpected_planes:\n raise RuntimeError(\n \"Unexpected anatomical planes: \"\n + \", \".join(sorted(unexpected_planes))\n )\n\nmissing_planes = (\n expected_planes\n - observed_planes\n)\n\nif missing_planes:\n raise RuntimeError(\n \"Missing anatomical planes globally: \"\n + \", \".join(sorted(missing_planes))\n )\n\nprint(\"Anatomical planes: PASS\")\n\n# ----------------------------------------------------------------------\n# PER-STUDY PLANE COVERAGE\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"PER-STUDY PLANE COVERAGE\")\nprint(\"=\" * 70)\n\ncoverage_rows = []\n\nfor study_id, group in test_series.groupby(\n \"StudyInstanceUID\",\n sort=False,\n):\n\n planes = set(\n group[\"Anatomical_Plane\"]\n .astype(str)\n )\n\n coverage_rows.append(\n {\n \"StudyInstanceUID\": study_id,\n \"series_count\": len(group),\n \"sagittal\": \"Sagittal\" in planes,\n \"coronal\": \"Coronal\" in planes,\n \"axial\": \"Axial\" in planes,\n }\n )\n\ntest_plane_coverage = pd.DataFrame(\n coverage_rows\n)\n\nprint(test_plane_coverage.to_string(index=False))\n\ncoverage_ok = bool(\n test_plane_coverage[\n [\n \"sagittal\",\n \"coronal\",\n \"axial\",\n ]\n ].all(axis=1).all()\n)\n\nif not coverage_ok:\n raise RuntimeError(\n \"At least one test study is missing a required anatomical plane.\"\n )\n\nprint()\nprint(\"Three-plane coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# PRINT ALL TEST SERIES\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"ALL TEST SERIES\")\nprint(\"=\" * 70)\n\nprint(\n test_series[\n required_series_columns\n ].sort_values(\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\",\n \"SeriesInstanceUID\",\n ]\n ).to_string(index=False)\n)\n\n# ----------------------------------------------------------------------\n# PER-PLANE CANDIDATE AUDIT\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"PER-STUDY / PER-PLANE CANDIDATE AUDIT\")\nprint(\"=\" * 70)\n\ncandidate_rows = []\n\nfor study_id, study_group in test_series.groupby(\n \"StudyInstanceUID\",\n sort=False,\n):\n\n for plane in [\n \"Sagittal\",\n \"Coronal\",\n \"Axial\",\n ]:\n\n plane_group = study_group[\n study_group[\"Anatomical_Plane\"].astype(str)\n == plane\n ].copy()\n\n if plane_group.empty:\n raise RuntimeError(\n f\"Study {study_id} has no {plane} series.\"\n )\n\n for _, row in plane_group.iterrows():\n\n candidate_rows.append(\n {\n \"StudyInstanceUID\": study_id,\n \"Anatomical_Plane\": plane,\n \"SeriesInstanceUID\": row[\n \"SeriesInstanceUID\"\n ],\n \"Fluid_Sensitive\": row[\n \"Fluid_Sensitive\"\n ],\n \"Fat_Suppression\": row[\n \"Fat_Suppression\"\n ],\n }\n )\n\ntest_series_candidates = pd.DataFrame(\n candidate_rows\n)\n\nprint(\n test_series_candidates.to_string(\n index=False\n )\n)\n\n# ----------------------------------------------------------------------\n# EXISTING NOTEBOOK SELECTION OBJECTS\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"EXISTING SERIES-SELECTION OBJECTS\")\nprint(\"=\" * 70)\n\nfor object_name in [\n \"selected_series\",\n \"series_audit\",\n \"test_summary\",\n \"test_slice_audit\",\n \"metadata_source\",\n \"test_mapping_counts\",\n \"test_bad_series_mapping\",\n]:\n\n if object_name in globals():\n\n value = globals()[object_name]\n\n print()\n print(\n f\"{object_name}: \"\n f\"type={type(value).__name__}\"\n )\n\n if isinstance(value, pd.DataFrame):\n\n print(\n \"shape:\",\n value.shape\n )\n\n print(\n \"columns:\",\n list(value.columns)\n )\n\n elif isinstance(value, (list, tuple, set)):\n\n print(\n \"length:\",\n len(value)\n )\n\n# ----------------------------------------------------------------------\n# EXISTING FUNCTION INSPECTION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"EXISTING FUNCTION SIGNATURES\")\nprint(\"=\" * 70)\n\nfunctions_to_inspect = [\n \"get_series_directory\",\n \"get_series_dir\",\n \"get_dicom_files\",\n \"get_sorted_dicom_files\",\n \"load_series\",\n \"load_series_33\",\n \"normalize_dicom_image\",\n \"read_dicom_series\",\n]\n\nfor function_name in functions_to_inspect:\n\n if function_name in globals():\n\n function_object = globals()[function_name]\n\n try:\n\n signature = inspect.signature(\n function_object\n )\n\n print(\n f\"{function_name}{signature}\"\n )\n\n except Exception as exc:\n\n print(\n f\"{function_name}: \"\n f\"signature unavailable \"\n f\"({type(exc).__name__})\"\n )\n\n# ----------------------------------------------------------------------\n# DATASET CONSTRUCTOR INSPECTION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"KNEESTUDYDATASET CONSTRUCTOR\")\nprint(\"=\" * 70)\n\ntry:\n\n dataset_signature = inspect.signature(\n KneeStudyDataset\n )\n\n print(\n \"KneeStudyDataset:\",\n dataset_signature\n )\n\nexcept Exception as exc:\n\n print(\n \"Could not inspect KneeStudyDataset signature:\",\n type(exc).__name__,\n str(exc)\n )\n\n# ----------------------------------------------------------------------\n# EXISTING TEST LOADER CANDIDATES\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"EXISTING TEST LOADER CANDIDATES\")\nprint(\"=\" * 70)\n\nfor object_name in [\n \"test_loader_candidates\",\n \"available_test_loaders\",\n]:\n\n if object_name in globals():\n\n value = globals()[object_name]\n\n print(\n f\"{object_name}:\",\n value\n )\n\n# ----------------------------------------------------------------------\n# NO SELECTION / NO DATASET CREATION\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 65 VERDICT\")\nprint(\"=\" * 70)\n\nprint(\n f\"Test studies: {len(test_ids_from_test)}\"\n)\n\nprint(\n f\"Test series rows: {len(test_series)}\"\n)\n\nprint(\n \"Test study coverage: PASS\"\n)\n\nprint(\n \"Test series schema: PASS\"\n)\n\nprint(\n \"Three-plane coverage: PASS\"\n)\n\nprint(\n \"Per-plane candidate audit: PASS\"\n)\n\nprint(\n \"Existing selection objects inspected: PASS\"\n)\n\nprint(\n \"Existing dataset constructor inspected: PASS\"\n)\n\nprint()\nprint(\n \"NO TEST DATASET CREATED.\"\n)\n\nprint(\n \"NO TEST DATALOADER CREATED.\"\n)\n\nprint(\n \"NO TEST INFERENCE PERFORMED.\"\n)\n\nprint(\n \"NO CHECKPOINT MODIFIED.\"\n)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 65 COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Send the complete Cell 65 output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:53:55.353582Z","iopub.execute_input":"2026-08-11T11:53:55.353962Z","iopub.status.idle":"2026-08-11T11:53:55.415534Z","shell.execute_reply.started":"2026-08-11T11:53:55.353926Z","shell.execute_reply":"2026-08-11T11:53:55.414635Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 66A - KNEESTUDYDATASET TEST-MODE COMPATIBILITY AUDIT\n# ======================================================================\n\nimport inspect\n\nprint(\"=\" * 70)\nprint(\"CELL 66A - KNEESTUDYDATASET TEST-MODE COMPATIBILITY AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"TARGETS\",\n \"KneeStudyDataset\",\n \"test_metadata\",\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\n\n# ----------------------------------------------------------------------\n# TEST METADATA VALIDATION\n# ----------------------------------------------------------------------\n\nrequired_columns = [\n \"StudyInstanceUID\",\n \"Sagittal_SeriesInstanceUID\",\n \"Coronal_SeriesInstanceUID\",\n \"Axial_SeriesInstanceUID\",\n]\n\nmissing_columns = [\n column\n for column in required_columns\n if column not in test_metadata.columns\n]\n\nif missing_columns:\n raise RuntimeError(\n \"Test metadata missing required columns: \"\n + \", \".join(missing_columns)\n )\n\nif len(test_metadata) != 3:\n raise RuntimeError(\n f\"Expected 3 test studies, got {len(test_metadata)}.\"\n )\n\nif test_metadata[\"StudyInstanceUID\"].duplicated().any():\n raise RuntimeError(\n \"Duplicate StudyInstanceUID values in test metadata.\"\n )\n\nprint(\"Test metadata shape:\", test_metadata.shape)\nprint(\"Test metadata schema: PASS\")\nprint(\"Test study count: 3\")\nprint(\"One row per study: PASS\")\n\n# ----------------------------------------------------------------------\n# TARGET VALIDATION\n# ----------------------------------------------------------------------\n\nif len(TARGETS) != 12:\n raise RuntimeError(\n f\"Expected 12 TARGETS, got {len(TARGETS)}.\"\n )\n\nprint(\"TARGETS: PASS\")\nprint(\"Target count:\", len(TARGETS))\n\n# ----------------------------------------------------------------------\n# CLASS SIGNATURE\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"DATASET CLASS SIGNATURE\")\nprint(\"=\" * 70)\n\ntry:\n class_signature = inspect.signature(\n KneeStudyDataset\n )\n print(\n \"KneeStudyDataset:\",\n class_signature\n )\nexcept Exception as exc:\n raise RuntimeError(\n \"Could not inspect KneeStudyDataset signature.\"\n ) from exc\n\n# ----------------------------------------------------------------------\n# METHOD SIGNATURES\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"DATASET METHOD SIGNATURES\")\nprint(\"=\" * 70)\n\nmethods_to_check = [\n \"__init__\",\n \"__len__\",\n \"__getitem__\",\n]\n\nmethod_signatures = {}\n\nfor method_name in methods_to_check:\n\n if not hasattr(KneeStudyDataset, method_name):\n raise RuntimeError(\n f\"KneeStudyDataset missing method: {method_name}\"\n )\n\n method_object = getattr(\n KneeStudyDataset,\n method_name\n )\n\n try:\n signature = inspect.signature(\n method_object\n )\n\n method_signatures[method_name] = signature\n\n print(\n f\"{method_name}{signature}\"\n )\n\n except Exception as exc:\n\n print(\n f\"{method_name}: \"\n f\"signature unavailable \"\n f\"({type(exc).__name__})\"\n )\n\n# ----------------------------------------------------------------------\n# SOURCE INSPECTION\n#\n# We inspect the existing implementation rather than creating a second\n# Dataset class or guessing how test samples should be represented.\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"DATASET IMPLEMENTATION SOURCE\")\nprint(\"=\" * 70)\n\nsource_results = {}\n\nfor method_name in methods_to_check:\n\n method_object = getattr(\n KneeStudyDataset,\n method_name\n )\n\n try:\n source = inspect.getsource(\n method_object\n )\n\n source_results[method_name] = source\n\n print()\n print(\n f\"----- {method_name} -----\"\n )\n print(source)\n\n except Exception as exc:\n\n source_results[method_name] = None\n\n print()\n print(\n f\"{method_name}: source unavailable \"\n f\"({type(exc).__name__})\"\n )\n\n# ----------------------------------------------------------------------\n# SOURCE AVAILABILITY CHECK\n# ----------------------------------------------------------------------\n\nunavailable_methods = [\n name\n for name, source in source_results.items()\n if source is None\n]\n\nif unavailable_methods:\n raise RuntimeError(\n \"Could not inspect Dataset implementation for: \"\n + \", \".join(unavailable_methods)\n + \". \"\n \"Do not guess the test Dataset behavior.\"\n )\n\nprint()\nprint(\"Dataset implementation inspection: PASS\")\n\n# ----------------------------------------------------------------------\n# LABEL DEPENDENCY AUDIT\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"LABEL DEPENDENCY AUDIT\")\nprint(\"=\" * 70)\n\nall_source = \"\\n\".join(\n source_results.values()\n)\n\nlabel_related_terms = [\n \"label\",\n \"targets\",\n \"TARGETS\",\n \"dataframe[target\",\n \"dataframe[targets\",\n]\n\ndetected_terms = []\n\nfor term in label_related_terms:\n\n if term in all_source:\n detected_terms.append(term)\n\nprint(\n \"Label-related implementation references:\",\n detected_terms\n)\n\n# ----------------------------------------------------------------------\n# TEST METADATA COLUMN ACCESS AUDIT\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TEST METADATA ACCESS AUDIT\")\nprint(\"=\" * 70)\n\nfor column in required_columns:\n\n if column in all_source:\n print(\n f\"{column}: referenced by Dataset implementation\"\n )\n else:\n print(\n f\"{column}: not directly referenced by Dataset implementation\"\n )\n\n# ----------------------------------------------------------------------\n# IMPORTANT SAFETY STOP\n#\n# We do NOT instantiate the Dataset yet. The next cell will be written\n# from the actual implementation discovered here.\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 66A VERDICT\")\nprint(\"=\" * 70)\n\nprint(\"Test metadata available: PASS\")\nprint(\"Test metadata schema: PASS\")\nprint(\"12-target configuration: PASS\")\nprint(\"KneeStudyDataset signature: PASS\")\nprint(\"Dataset methods located: PASS\")\nprint(\"Dataset implementation inspected: PASS\")\nprint()\nprint(\"No test Dataset constructed.\")\nprint(\"No test DataLoader constructed.\")\nprint(\"No test inference performed.\")\nprint(\"No checkpoint modified.\")\nprint(\"No training performed.\")\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 66A COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Existing Dataset implementation has been audited.\"\n)\nprint(\n \"Send the complete CELL 66A output before constructing \"\n \"the test Dataset.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T11:57:09.285862Z","iopub.execute_input":"2026-08-11T11:57:09.287157Z","iopub.status.idle":"2026-08-11T11:57:09.311817Z","shell.execute_reply.started":"2026-08-11T11:57:09.287118Z","shell.execute_reply":"2026-08-11T11:57:09.310862Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 66C-R - FINALIZE TEST DATALOADER\n# ======================================================================\n\nimport os\nimport torch\nfrom torch.utils.data import DataLoader\n\nprint(\"=\" * 70)\nprint(\"CELL 66C-R - FINALIZE TEST DATALOADER\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"TARGETS\",\n \"test_metadata\",\n \"test_dataset\",\n \"TRAIN_SERIES_DIR\",\n \"TEST_SERIES_DIR\",\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\n\n# ----------------------------------------------------------------------\n# TEST METADATA VALIDATION\n# ----------------------------------------------------------------------\n\nrequired_test_columns = [\n \"StudyInstanceUID\",\n \"Sagittal_SeriesInstanceUID\",\n \"Coronal_SeriesInstanceUID\",\n \"Axial_SeriesInstanceUID\",\n]\n\nmissing_columns = [\n column\n for column in required_test_columns\n if column not in test_metadata.columns\n]\n\nif missing_columns:\n raise RuntimeError(\n \"Test metadata missing required columns: \"\n + \", \".join(missing_columns)\n )\n\nif len(test_metadata) != 3:\n raise RuntimeError(\n f\"Expected 3 test studies, got {len(test_metadata)}.\"\n )\n\nprint(\"Test metadata schema: PASS\")\nprint(\"Test metadata shape:\", test_metadata.shape)\nprint(\"Test studies:\", len(test_metadata))\n\n# ----------------------------------------------------------------------\n# DATASET VALIDATION\n# ----------------------------------------------------------------------\n\nif len(test_dataset) != 3:\n raise RuntimeError(\n f\"Expected test Dataset length 3, got {len(test_dataset)}.\"\n )\n\nif not isinstance(\n test_dataset,\n TestKneeStudyDataset,\n):\n raise RuntimeError(\n \"test_dataset is not the expected TestKneeStudyDataset.\"\n )\n\nprint(\"Test Dataset: PASS\")\nprint(\"Test Dataset length:\", len(test_dataset))\nprint(\n \"Test Dataset class:\",\n type(test_dataset).__name__\n)\n\n# ----------------------------------------------------------------------\n# PATH SAFETY\n#\n# The original training directory must already be restored after\n# __getitem__ completes. We simply verify the current global value.\n# No local variable from __getitem__ is referenced here.\n# ----------------------------------------------------------------------\n\nif not os.path.isdir(TRAIN_SERIES_DIR):\n raise RuntimeError(\n \"TRAIN_SERIES_DIR is not a valid directory:\\n\"\n + str(TRAIN_SERIES_DIR)\n )\n\nif not os.path.isdir(TEST_SERIES_DIR):\n raise RuntimeError(\n \"TEST_SERIES_DIR is not a valid directory:\\n\"\n + str(TEST_SERIES_DIR)\n )\n\nprint(\"TRAIN_SERIES_DIR restored: PASS\")\nprint(\"TEST_SERIES_DIR available: PASS\")\n\n# ----------------------------------------------------------------------\n# CREATE TEST DATALOADER\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CREATING TEST DATALOADER\")\nprint(\"=\" * 70)\n\ntest_loader = DataLoader(\n test_dataset,\n batch_size=2,\n shuffle=False,\n num_workers=0,\n pin_memory=False,\n)\n\nif len(test_loader) != 2:\n raise RuntimeError(\n f\"Expected 2 test batches, got {len(test_loader)}.\"\n )\n\nprint(\"Test DataLoader: PASS\")\nprint(\"Test batches:\", len(test_loader))\n\n# ----------------------------------------------------------------------\n# DATALOADER SMOKE TEST\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TEST DATALOADER SMOKE TEST\")\nprint(\"=\" * 70)\n\nfirst_batch = next(iter(test_loader))\n\nrequired_batch_keys = [\n \"image\",\n \"label\",\n \"study_id\",\n]\n\nmissing_batch_keys = [\n key\n for key in required_batch_keys\n if key not in first_batch\n]\n\nif missing_batch_keys:\n raise RuntimeError(\n \"Test batch missing keys: \"\n + \", \".join(missing_batch_keys)\n )\n\nimages = first_batch[\"image\"]\nlabels = first_batch[\"label\"]\nstudy_ids = first_batch[\"study_id\"]\n\nif images.ndim != 4:\n raise RuntimeError(\n \"Unexpected image tensor dimensions: \"\n + str(tuple(images.shape))\n )\n\nif images.shape[1:] != (\n 21,\n 224,\n 224,\n):\n raise RuntimeError(\n \"Unexpected image batch shape: \"\n + str(tuple(images.shape))\n )\n\nif labels.ndim != 2:\n raise RuntimeError(\n \"Unexpected label tensor dimensions: \"\n + str(tuple(labels.shape))\n )\n\nif labels.shape[1] != 12:\n raise RuntimeError(\n \"Unexpected label batch shape: \"\n + str(tuple(labels.shape))\n )\n\nif not torch.isfinite(images).all():\n raise RuntimeError(\n \"NaN/Inf detected in test image batch.\"\n )\n\nprint(\"Batch keys:\", list(first_batch.keys()))\nprint(\"Image batch shape:\", tuple(images.shape))\nprint(\"Label structural shape:\", tuple(labels.shape))\nprint(\"Study IDs:\", list(study_ids))\nprint(\"Image min:\", float(images.min()))\nprint(\"Image max:\", float(images.max()))\nprint(\"Image mean:\", float(images.mean()))\nprint(\"Image std:\", float(images.std()))\n\nprint(\"Batch structure: PASS\")\nprint(\"21-channel input: PASS\")\nprint(\"12-target structural format: PASS\")\nprint(\"NaN/Inf check: PASS\")\n\n# ----------------------------------------------------------------------\n# COMPLETE TEST COVERAGE\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"TEST DATALOADER COVERAGE\")\nprint(\"=\" * 70)\n\nloader_study_ids = []\n\nfor batch in test_loader:\n\n batch_images = batch[\"image\"]\n batch_labels = batch[\"label\"]\n batch_ids = batch[\"study_id\"]\n\n if batch_images.shape[1:] != (\n 21,\n 224,\n 224,\n ):\n raise RuntimeError(\n \"Invalid image shape encountered during full \"\n \"test DataLoader iteration: \"\n + str(tuple(batch_images.shape))\n )\n\n if batch_labels.shape[1] != 12:\n raise RuntimeError(\n \"Invalid label structural shape encountered \"\n \"during full test DataLoader iteration: \"\n + str(tuple(batch_labels.shape))\n )\n\n if not torch.isfinite(batch_images).all():\n raise RuntimeError(\n \"NaN/Inf detected during full test DataLoader iteration.\"\n )\n\n loader_study_ids.extend(\n [\n str(study_id)\n for study_id in batch_ids\n ]\n )\n\nexpected_study_ids = (\n test_metadata[\n \"StudyInstanceUID\"\n ]\n .astype(str)\n .tolist()\n)\n\nif len(loader_study_ids) != 3:\n raise RuntimeError(\n \"Expected exactly 3 studies from test DataLoader, \"\n f\"got {len(loader_study_ids)}.\"\n )\n\nif len(set(loader_study_ids)) != 3:\n raise RuntimeError(\n \"Duplicate StudyInstanceUID detected in test DataLoader.\"\n )\n\nif set(loader_study_ids) != set(\n expected_study_ids\n):\n raise RuntimeError(\n \"Test DataLoader study coverage does not match \"\n \"test metadata.\"\n )\n\nprint(\"Studies loaded:\", len(loader_study_ids))\nprint(\"Unique studies:\", len(set(loader_study_ids)))\nprint(\"Expected studies:\", len(expected_study_ids))\nprint(\"3/3 test studies covered exactly once: PASS\")\n\n# ----------------------------------------------------------------------\n# FINAL GLOBAL PATH SAFETY CHECK\n# ----------------------------------------------------------------------\n\nif TRAIN_SERIES_DIR != (\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection/\"\n \"train_series\"\n):\n raise RuntimeError(\n \"TRAIN_SERIES_DIR was not restored to the \"\n \"established training-series directory.\"\n )\n\nif TEST_SERIES_DIR != (\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection/\"\n \"test_series\"\n):\n raise RuntimeError(\n \"TEST_SERIES_DIR changed unexpectedly.\"\n )\n\nprint(\"Established TRAIN_SERIES_DIR restored: PASS\")\nprint(\"Established TEST_SERIES_DIR preserved: PASS\")\n\n# ----------------------------------------------------------------------\n# FINAL VERDICT\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 66C-R VERDICT\")\nprint(\"=\" * 70)\n\nprint(\"Test metadata: PASS\")\nprint(\"Test Dataset: PASS\")\nprint(\"Test DataLoader: PASS\")\nprint(\"3-plane test loading: PASS\")\nprint(\"21-channel input: PASS\")\nprint(\"12-target structural format: PASS\")\nprint(\"NaN/Inf validation: PASS\")\nprint(\"3/3 test studies covered exactly once: PASS\")\nprint(\"Training-series path restored: PASS\")\nprint(\"Test-series path preserved: PASS\")\nprint()\nprint(\"No labels fabricated.\")\nprint(\"No training performed.\")\nprint(\"No checkpoint modified.\")\nprint(\"No predictions generated.\")\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 66C-R COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Test Dataset and DataLoader are ready for final inference.\"\n)\nprint(\n \"Send the complete CELL 66C-R output before proceeding.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T12:01:01.521563Z","iopub.execute_input":"2026-08-11T12:01:01.521929Z","iopub.status.idle":"2026-08-11T12:01:06.152789Z","shell.execute_reply.started":"2026-08-11T12:01:01.52188Z","shell.execute_reply":"2026-08-11T12:01:06.151784Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 67A-R - RESTORE FINAL INFERENCE OBJECTS\n# ======================================================================\n\nimport os\nimport copy\nimport torch\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"CELL 67A-R - RESTORE FINAL INFERENCE OBJECTS\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED NOTEBOOK OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"TARGETS\",\n \"test_loader\",\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\n\n# ----------------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------------\n\nif \"device\" not in globals():\n device = torch.device(\n \"cuda\"\n if torch.cuda.is_available()\n else \"cpu\"\n )\n\nprint(\"Device:\", device)\n\n# ----------------------------------------------------------------------\n# AUDIT DIRECTORY\n# ----------------------------------------------------------------------\n\naudit_dir = \"/kaggle/working/rsna_knee_audit\"\n\nif not os.path.isdir(audit_dir):\n raise RuntimeError(\n \"Audit directory not found: \"\n + audit_dir\n )\n\nprint(\"Audit directory: PASS\")\n\n# ----------------------------------------------------------------------\n# LOAD CELL 61 AUTHORITATIVE OOF-BACKED MODEL PLAN\n#\n# Cell 61 contains:\n# target\n# selected_model\n# selected_threshold\n# selected_best_f1\n# ----------------------------------------------------------------------\n\ncell61_path = os.path.join(\n audit_dir,\n \"cell61_final_oof_backed_model_plan.csv\",\n)\n\nif not os.path.isfile(cell61_path):\n raise RuntimeError(\n \"Cell-61 final OOF-backed model plan not found: \"\n + cell61_path\n )\n\ncell61_plan = pd.read_csv(\n cell61_path\n)\n\nprint(\n \"Cell-61 model plan:\",\n cell61_plan.shape,\n)\n\nrequired_cell61_columns = [\n \"target\",\n \"selected_model\",\n \"selected_threshold\",\n]\n\nmissing_cell61_columns = [\n column\n for column in required_cell61_columns\n if column not in cell61_plan.columns\n]\n\nif missing_cell61_columns:\n raise RuntimeError(\n \"Cell-61 model plan missing required columns: \"\n + \", \".join(missing_cell61_columns)\n )\n\n# ----------------------------------------------------------------------\n# LOAD CELL 63 AUTHORITATIVE THREE-FOLD INFERENCE PLAN\n#\n# Cell 63 contains:\n# target\n# selected_model\n# threshold\n# fold0_model\n# fold1_model\n# fold2_model\n# ----------------------------------------------------------------------\n\ncell63_path = os.path.join(\n audit_dir,\n \"cell63_final_three_fold_inference_plan.csv\",\n)\n\nif not os.path.isfile(cell63_path):\n raise RuntimeError(\n \"Cell-63 final three-fold inference plan not found: \"\n + cell63_path\n )\n\ncell63_plan = pd.read_csv(\n cell63_path\n)\n\nprint(\n \"Cell-63 inference plan:\",\n cell63_plan.shape,\n)\n\nrequired_cell63_columns = [\n \"target\",\n \"selected_model\",\n \"threshold\",\n \"fold0_model\",\n \"fold1_model\",\n \"fold2_model\",\n]\n\nmissing_cell63_columns = [\n column\n for column in required_cell63_columns\n if column not in cell63_plan.columns\n]\n\nif missing_cell63_columns:\n raise RuntimeError(\n \"Cell-63 inference plan missing required columns: \"\n + \", \".join(missing_cell63_columns)\n )\n\n# ----------------------------------------------------------------------\n# BASIC PLAN VALIDATION\n# ----------------------------------------------------------------------\n\nif len(cell61_plan) != 12:\n raise RuntimeError(\n \"Cell-61 plan must contain exactly 12 targets.\"\n )\n\nif len(cell63_plan) != 12:\n raise RuntimeError(\n \"Cell-63 plan must contain exactly 12 targets.\"\n )\n\nexpected_targets = list(TARGETS)\n\nif cell61_plan[\n \"target\"\n].tolist() != expected_targets:\n\n raise RuntimeError(\n \"Cell-61 target ordering does not match TARGETS.\"\n )\n\nif cell63_plan[\n \"target\"\n].tolist() != expected_targets:\n\n raise RuntimeError(\n \"Cell-63 target ordering does not match TARGETS.\"\n )\n\nif cell61_plan[\n \"target\"\n].duplicated().any():\n\n raise RuntimeError(\n \"Duplicate targets found in Cell-61 plan.\"\n )\n\nif cell63_plan[\n \"target\"\n].duplicated().any():\n\n raise RuntimeError(\n \"Duplicate targets found in Cell-63 plan.\"\n )\n\nprint(\"Cell-61 target ordering: PASS\")\nprint(\"Cell-63 target ordering: PASS\")\nprint(\"12-target coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# VERIFY CELL 61 AND CELL 63 MODEL ASSIGNMENTS AGREE\n# ----------------------------------------------------------------------\n\ncell61_models = dict(\n zip(\n cell61_plan[\"target\"],\n cell61_plan[\"selected_model\"],\n )\n)\n\ncell63_models = dict(\n zip(\n cell63_plan[\"target\"],\n cell63_plan[\"selected_model\"],\n )\n)\n\nmodel_disagreements = []\n\nfor target in expected_targets:\n\n if cell61_models[target] != cell63_models[target]:\n\n model_disagreements.append(\n (\n target,\n cell61_models[target],\n cell63_models[target],\n )\n )\n\nif model_disagreements:\n raise RuntimeError(\n \"Cell-61 and Cell-63 selected-model disagreement: \"\n + str(model_disagreements)\n )\n\nprint(\n \"Cell-61 vs Cell-63 selected-model consistency: PASS\"\n)\n\n# ----------------------------------------------------------------------\n# VERIFY CELL 61 THRESHOLDS AGAINST CELL 63 THRESHOLDS\n#\n# Cell 61 uses selected_threshold.\n# Cell 63 uses threshold.\n#\n# They must represent the same OOF-derived thresholds.\n# ----------------------------------------------------------------------\n\ncell61_thresholds = dict(\n zip(\n cell61_plan[\"target\"],\n cell61_plan[\"selected_threshold\"],\n )\n)\n\ncell63_thresholds = dict(\n zip(\n cell63_plan[\"target\"],\n cell63_plan[\"threshold\"],\n )\n)\n\nthreshold_disagreements = []\n\nfor target in expected_targets:\n\n threshold61 = float(\n cell61_thresholds[target]\n )\n\n threshold63 = float(\n cell63_thresholds[target]\n )\n\n if abs(\n threshold61 - threshold63\n ) > 1e-12:\n\n threshold_disagreements.append(\n (\n target,\n threshold61,\n threshold63,\n )\n )\n\nif threshold_disagreements:\n raise RuntimeError(\n \"Cell-61 and Cell-63 threshold disagreement: \"\n + str(threshold_disagreements)\n )\n\nprint(\n \"Cell-61 vs Cell-63 threshold consistency: PASS\"\n)\n\n# ----------------------------------------------------------------------\n# CONSTRUCT THE AUTHORITATIVE FINAL INFERENCE PLAN\n#\n# Threshold comes from Cell 61.\n# Fold mappings come from Cell 63.\n#\n# No threshold is recomputed.\n# No model assignment is changed.\n# ----------------------------------------------------------------------\n\nfinal_inference_plan = pd.DataFrame(\n {\n \"target\": expected_targets,\n \"selected_model\": [\n cell61_models[target]\n for target in expected_targets\n ],\n \"selected_threshold\": [\n float(cell61_thresholds[target])\n for target in expected_targets\n ],\n \"fold0_model\": [\n cell63_plan.loc[\n cell63_plan[\"target\"] == target,\n \"fold0_model\",\n ].iloc[0]\n for target in expected_targets\n ],\n \"fold1_model\": [\n cell63_plan.loc[\n cell63_plan[\"target\"] == target,\n \"fold1_model\",\n ].iloc[0]\n for target in expected_targets\n ],\n \"fold2_model\": [\n cell63_plan.loc[\n cell63_plan[\"target\"] == target,\n \"fold2_model\",\n ].iloc[0]\n for target in expected_targets\n ],\n }\n)\n\n# ----------------------------------------------------------------------\n# FINAL PLAN SCHEMA VALIDATION\n# ----------------------------------------------------------------------\n\nrequired_final_plan_columns = [\n \"target\",\n \"selected_model\",\n \"selected_threshold\",\n \"fold0_model\",\n \"fold1_model\",\n \"fold2_model\",\n]\n\nmissing_final_plan_columns = [\n column\n for column in required_final_plan_columns\n if column not in final_inference_plan.columns\n]\n\nif missing_final_plan_columns:\n raise RuntimeError(\n \"Final inference plan missing columns: \"\n + \", \".join(missing_final_plan_columns)\n )\n\nif len(final_inference_plan) != 12:\n raise RuntimeError(\n \"Final inference plan must contain 12 targets.\"\n )\n\nif final_inference_plan[\n \"target\"\n].tolist() != expected_targets:\n\n raise RuntimeError(\n \"Final inference plan target ordering invalid.\"\n )\n\n# ----------------------------------------------------------------------\n# VERIFY SELECTED MODEL VALUES\n# ----------------------------------------------------------------------\n\nallowed_models = {\n \"baseline\",\n \"fine_tuned\",\n}\n\ninvalid_models = sorted(\n set(\n final_inference_plan[\n \"selected_model\"\n ]\n ) - allowed_models\n)\n\nif invalid_models:\n raise RuntimeError(\n \"Invalid selected model values: \"\n + \", \".join(invalid_models)\n )\n\n# ----------------------------------------------------------------------\n# VERIFY THRESHOLDS\n# ----------------------------------------------------------------------\n\nif (\n final_inference_plan[\n \"selected_threshold\"\n ].isna().any()\n):\n\n raise RuntimeError(\n \"Missing selected threshold.\"\n )\n\nif (\n (final_inference_plan[\"selected_threshold\"] < 0.0)\n | (final_inference_plan[\"selected_threshold\"] > 1.0)\n).any():\n\n raise RuntimeError(\n \"Selected threshold outside [0, 1].\"\n )\n\nprint(\"Final inference plan schema: PASS\")\nprint(\"Selected model values: PASS\")\nprint(\"Selected thresholds: PASS\")\nprint()\nprint(final_inference_plan.to_string(index=False))\n\n# ----------------------------------------------------------------------\n# CHECKPOINT PATHS\n#\n# These are the exact validated checkpoints from the completed\n# baseline and controlled fine-tuning experiments.\n# ----------------------------------------------------------------------\n\ncheckpoint_paths = {\n \"baseline_fold0\": os.path.join(\n audit_dir,\n \"cell37_fold0_best_model.pt\",\n ),\n \"baseline_fold1\": os.path.join(\n audit_dir,\n \"cell40_fold1_best_model.pt\",\n ),\n \"baseline_fold2\": os.path.join(\n audit_dir,\n \"cell44_fold2_best_model.pt\",\n ),\n \"finetuned_fold0\": os.path.join(\n audit_dir,\n \"cell52_fold0_finetuned_experiment.pt\",\n ),\n \"finetuned_fold1\": os.path.join(\n audit_dir,\n \"cell54_fold1_finetuned_experiment.pt\",\n ),\n \"finetuned_fold2\": os.path.join(\n audit_dir,\n \"cell56_fold2_finetuned_experiment.pt\",\n ),\n}\n\nfor model_name, checkpoint_path in checkpoint_paths.items():\n\n if not os.path.isfile(\n checkpoint_path\n ):\n raise RuntimeError(\n f\"Missing validated checkpoint: \"\n f\"{checkpoint_path}\"\n )\n\nprint(\"Six checkpoint files: PASS\")\n\n# ----------------------------------------------------------------------\n# FIND EXISTING VALIDATED MODEL ARCHITECTURE\n#\n# Cell 62 established:\n# total parameters = 11,367,372\n# input = 21 channels\n# output = 12 targets\n#\n# Do not invent a new constructor here.\n# ----------------------------------------------------------------------\n\narchitecture_model = None\narchitecture_name = None\n\npreferred_model_names = [\n \"test_model\",\n \"model\",\n]\n\nfor name in preferred_model_names:\n\n if name not in globals():\n continue\n\n candidate = globals()[name]\n\n if not isinstance(\n candidate,\n torch.nn.Module,\n ):\n continue\n\n parameter_count = sum(\n parameter.numel()\n for parameter in candidate.parameters()\n )\n\n if parameter_count == 11_367_372:\n\n architecture_model = candidate\n architecture_name = name\n break\n\nif architecture_model is None:\n\n for name, candidate in globals().items():\n\n if not isinstance(\n candidate,\n torch.nn.Module,\n ):\n continue\n\n try:\n parameter_count = sum(\n parameter.numel()\n for parameter in candidate.parameters()\n )\n except Exception:\n continue\n\n if parameter_count == 11_367_372:\n\n architecture_model = candidate\n architecture_name = name\n break\n\nif architecture_model is None:\n raise RuntimeError(\n \"No existing validated 11,367,372-parameter \"\n \"model architecture is available. \"\n \"Do not invent a model constructor.\"\n )\n\nprint(\n \"Existing validated architecture:\",\n architecture_name,\n)\n\n# ----------------------------------------------------------------------\n# ARCHITECTURE VALIDATION\n# ----------------------------------------------------------------------\n\narchitecture_model = architecture_model.to(\n device\n)\n\narchitecture_model.eval()\n\nsynthetic_input = torch.zeros(\n 1,\n 21,\n 224,\n 224,\n dtype=torch.float32,\n device=device,\n)\n\nwith torch.no_grad():\n\n synthetic_output = architecture_model(\n synthetic_input\n )\n\nif synthetic_output.shape != (\n 1,\n 12,\n):\n\n raise RuntimeError(\n \"Architecture output shape is \"\n + str(tuple(synthetic_output.shape))\n )\n\nif not torch.isfinite(\n synthetic_output\n).all():\n\n raise RuntimeError(\n \"Architecture synthetic output contains NaN/Inf.\"\n )\n\nprint(\"Architecture: PASS\")\nprint(\"21-channel input: PASS\")\nprint(\"12-target output: PASS\")\nprint(\"Synthetic forward pass: PASS\")\n\n# ----------------------------------------------------------------------\n# CHECKPOINT STATE-DICT EXTRACTION\n# ----------------------------------------------------------------------\n\ndef get_state_dict_from_validated_checkpoint(\n checkpoint,\n model_name,\n):\n\n if not isinstance(\n checkpoint,\n dict,\n ):\n\n raise RuntimeError(\n f\"Unexpected checkpoint type for {model_name}: \"\n f\"{type(checkpoint)}\"\n )\n\n if \"state_dict\" in checkpoint:\n\n state_dict = checkpoint[\n \"state_dict\"\n ]\n\n elif \"model_state_dict\" in checkpoint:\n\n state_dict = checkpoint[\n \"model_state_dict\"\n ]\n\n else:\n\n # A raw PyTorch state_dict is itself a dictionary\n # whose values are tensors.\n if checkpoint and all(\n isinstance(value, torch.Tensor)\n for value in checkpoint.values()\n ):\n\n state_dict = checkpoint\n\n else:\n\n raise RuntimeError(\n f\"Checkpoint {model_name} does not contain \"\n \"a recognizable state_dict.\"\n )\n\n if not state_dict:\n raise RuntimeError(\n f\"Empty state_dict for {model_name}.\"\n )\n\n return state_dict\n\n# ----------------------------------------------------------------------\n# LOAD ALL SIX VALIDATED CHECKPOINTS\n# ----------------------------------------------------------------------\n\nmodel_registry = {}\n\nfor model_name, checkpoint_path in checkpoint_paths.items():\n\n checkpoint = torch.load(\n checkpoint_path,\n map_location=device,\n )\n\n state_dict = (\n get_state_dict_from_validated_checkpoint(\n checkpoint,\n model_name,\n )\n )\n\n restored_model = copy.deepcopy(\n architecture_model\n )\n\n try:\n\n restored_model.load_state_dict(\n state_dict,\n strict=True,\n )\n\n except RuntimeError as exc:\n\n raise RuntimeError(\n f\"State-dict compatibility failed for \"\n f\"{model_name}.\"\n ) from exc\n\n restored_model = restored_model.to(\n device\n )\n\n restored_model.eval()\n\n # Parameter validity\n for parameter in restored_model.parameters():\n\n if not torch.isfinite(\n parameter\n ).all():\n\n raise RuntimeError(\n f\"NaN/Inf parameter found in \"\n f\"{model_name}.\"\n )\n\n # Forward validation\n with torch.no_grad():\n\n output = restored_model(\n synthetic_input\n )\n\n if output.shape != (\n 1,\n 12,\n ):\n\n raise RuntimeError(\n f\"Invalid output shape for \"\n f\"{model_name}: \"\n f\"{tuple(output.shape)}\"\n )\n\n if not torch.isfinite(\n output\n ).all():\n\n raise RuntimeError(\n f\"NaN/Inf output for \"\n f\"{model_name}.\"\n )\n\n model_registry[\n model_name\n ] = restored_model\n\n print(\n f\"{model_name}: LOAD PASS\"\n )\n\n# ----------------------------------------------------------------------\n# CREATE THE EXACT OBJECT NAMES REQUIRED BY CELL 67\n# ----------------------------------------------------------------------\n\nbaseline_fold0 = model_registry[\n \"baseline_fold0\"\n]\n\nbaseline_fold1 = model_registry[\n \"baseline_fold1\"\n]\n\nbaseline_fold2 = model_registry[\n \"baseline_fold2\"\n]\n\nfinetuned_fold0 = model_registry[\n \"finetuned_fold0\"\n]\n\nfinetuned_fold1 = model_registry[\n \"finetuned_fold1\"\n]\n\nfinetuned_fold2 = model_registry[\n \"finetuned_fold2\"\n]\n\n# ----------------------------------------------------------------------\n# VERIFY ALL MODELS ARE IN EVALUATION MODE\n# ----------------------------------------------------------------------\n\nfor model_name, model_object in model_registry.items():\n\n if model_object.training:\n\n raise RuntimeError(\n f\"{model_name} is not in evaluation mode.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 67A-R VERDICT\")\nprint(\"=\" * 70)\n\nprint(\"Cell-61 model plan: PASS\")\nprint(\"Cell-63 fold mapping plan: PASS\")\nprint(\"Model-selection consistency: PASS\")\nprint(\"Threshold consistency: PASS\")\nprint(\"Final combined inference plan: PASS\")\nprint(\"12 targets covered: PASS\")\nprint(\"Existing architecture: PASS\")\nprint(\"21-channel input: PASS\")\nprint(\"12-target output: PASS\")\nprint(\"Six validated checkpoints: PASS\")\nprint(\"All six checkpoint loads: PASS\")\nprint(\"All six models in evaluation mode: PASS\")\n\nprint()\nprint(\"No training performed.\")\nprint(\"No checkpoint modified.\")\nprint(\"No test inference performed.\")\n\nprint(\"=\" * 70)\nprint(\"CELL 67A-R COMPLETE\")\nprint(\"=\" * 70)\n\nprint(\n \"Final inference objects are restored and validated.\"\n)\nprint(\n \"Existing test_loader preserved.\"\n)\nprint(\n \"Proceed to Cell 67 final test inference.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T12:04:45.162672Z","iopub.execute_input":"2026-08-11T12:04:45.163074Z","iopub.status.idle":"2026-08-11T12:04:46.296501Z","shell.execute_reply.started":"2026-08-11T12:04:45.163041Z","shell.execute_reply":"2026-08-11T12:04:46.295491Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 67 - FINAL THREE-FOLD TEST INFERENCE\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport torch\n\nprint(\"=\" * 70)\nprint(\"CELL 67 - FINAL THREE-FOLD TEST INFERENCE\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED OBJECTS\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"TARGETS\",\n \"test_loader\",\n \"final_inference_plan\",\n \"baseline_fold0\",\n \"baseline_fold1\",\n \"baseline_fold2\",\n \"finetuned_fold0\",\n \"finetuned_fold1\",\n \"finetuned_fold2\",\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n + \". Do not continue.\"\n )\n\nprint(\"Required notebook objects: PASS\")\n\n# ----------------------------------------------------------------------\n# DEVICE\n# ----------------------------------------------------------------------\n\nif \"device\" not in globals():\n device = torch.device(\n \"cuda\"\n if torch.cuda.is_available()\n else \"cpu\"\n )\n\nprint(\"Device:\", device)\n\n# ----------------------------------------------------------------------\n# VALIDATE FINAL PLAN\n# ----------------------------------------------------------------------\n\nrequired_plan_columns = [\n \"target\",\n \"selected_model\",\n \"selected_threshold\",\n \"fold0_model\",\n \"fold1_model\",\n \"fold2_model\",\n]\n\nmissing_plan_columns = [\n column\n for column in required_plan_columns\n if column not in final_inference_plan.columns\n]\n\nif missing_plan_columns:\n raise RuntimeError(\n \"Final inference plan missing columns: \"\n + \", \".join(missing_plan_columns)\n )\n\nif len(final_inference_plan) != len(TARGETS):\n raise RuntimeError(\n \"Final inference plan does not contain exactly \"\n f\"{len(TARGETS)} targets.\"\n )\n\nif final_inference_plan[\n \"target\"\n].tolist() != list(TARGETS):\n\n raise RuntimeError(\n \"Final inference plan target ordering does not \"\n \"match TARGETS.\"\n )\n\nprint(\"Final inference plan: PASS\")\nprint(\"Target ordering: PASS\")\nprint(\"12-target configuration: PASS\")\n\n# ----------------------------------------------------------------------\n# MODEL REGISTRY\n# ----------------------------------------------------------------------\n\nmodel_registry = {\n \"baseline_fold0\": baseline_fold0,\n \"baseline_fold1\": baseline_fold1,\n \"baseline_fold2\": baseline_fold2,\n \"finetuned_fold0\": finetuned_fold0,\n \"finetuned_fold1\": finetuned_fold1,\n \"finetuned_fold2\": finetuned_fold2,\n}\n\n# ----------------------------------------------------------------------\n# MODEL VALIDATION\n# ----------------------------------------------------------------------\n\nfor model_name, model_object in model_registry.items():\n\n if not isinstance(\n model_object,\n torch.nn.Module,\n ):\n raise RuntimeError(\n f\"{model_name} is not a PyTorch model.\"\n )\n\n if model_object.training:\n raise RuntimeError(\n f\"{model_name} is not in evaluation mode.\"\n )\n\nprint(\"Six inference models: PASS\")\nprint(\"Evaluation mode: PASS\")\n\n# ----------------------------------------------------------------------\n# VALIDATE TARGET MODEL MAPPING\n# ----------------------------------------------------------------------\n\nfor _, row in final_inference_plan.iterrows():\n\n target = row[\"target\"]\n\n for fold_column in [\n \"fold0_model\",\n \"fold1_model\",\n \"fold2_model\",\n ]:\n\n model_name = row[fold_column]\n\n if model_name not in model_registry:\n raise RuntimeError(\n f\"Invalid model '{model_name}' for \"\n f\"{target} in {fold_column}.\"\n )\n\n selected_model = row[\"selected_model\"]\n\n if selected_model not in [\n \"baseline\",\n \"fine_tuned\",\n ]:\n raise RuntimeError(\n f\"Invalid selected_model '{selected_model}' \"\n f\"for target {target}.\"\n )\n\n threshold = float(\n row[\"selected_threshold\"]\n )\n\n if not 0.0 <= threshold <= 1.0:\n raise RuntimeError(\n f\"Invalid threshold {threshold} \"\n f\"for target {target}.\"\n )\n\nprint(\"Target model assignments: PASS\")\nprint(\"Threshold validation: PASS\")\n\n# ----------------------------------------------------------------------\n# TEST LOADER VALIDATION\n# ----------------------------------------------------------------------\n\nexpected_test_count = 3\n\nif len(test_loader.dataset) != expected_test_count:\n raise RuntimeError(\n \"Unexpected test dataset size: \"\n + str(len(test_loader.dataset))\n )\n\nprint(\n \"Test Dataset size:\",\n len(test_loader.dataset),\n)\n\nprint(\"Test DataLoader: PASS\")\n\n# ----------------------------------------------------------------------\n# EXPECTED TEST STUDY IDS\n#\n# Prefer the already-established test metadata object.\n# Do not guess IDs.\n# ----------------------------------------------------------------------\n\nif \"test_metadata\" in globals():\n\n test_metadata_ids = [\n str(value)\n for value in test_metadata[\n \"StudyInstanceUID\"\n ].tolist()\n ]\n\nelif \"test\" in globals() and (\n \"StudyInstanceUID\" in test.columns\n):\n\n test_metadata_ids = [\n str(value)\n for value in test[\n \"StudyInstanceUID\"\n ].tolist()\n ]\n\nelif \"test_ids\" in globals():\n\n test_metadata_ids = sorted(\n str(value)\n for value in test_ids\n )\n\nelse:\n raise RuntimeError(\n \"No established test StudyInstanceUID source \"\n \"is available.\"\n )\n\nif len(test_metadata_ids) != expected_test_count:\n raise RuntimeError(\n \"Expected exactly 3 test StudyInstanceUIDs, \"\n f\"found {len(test_metadata_ids)}.\"\n )\n\nif len(set(test_metadata_ids)) != expected_test_count:\n raise RuntimeError(\n \"Duplicate test StudyInstanceUIDs detected.\"\n )\n\nprint(\"Expected test studies:\", len(test_metadata_ids))\nprint(\"Test StudyInstanceUID source: PASS\")\n\n# ----------------------------------------------------------------------\n# INFERENCE STORAGE\n#\n# Store probabilities by study ID and target.\n# ----------------------------------------------------------------------\n\nprobability_rows = []\n\nseen_studies = []\n\n# ----------------------------------------------------------------------\n# FINAL TEST INFERENCE\n# ----------------------------------------------------------------------\n\nfor batch_index, batch in enumerate(test_loader):\n\n if not isinstance(batch, dict):\n raise RuntimeError(\n f\"Unexpected test batch type at batch \"\n f\"{batch_index}: {type(batch)}\"\n )\n\n required_batch_keys = [\n \"image\",\n \"study_id\",\n ]\n\n missing_batch_keys = [\n key\n for key in required_batch_keys\n if key not in batch\n ]\n\n if missing_batch_keys:\n raise RuntimeError(\n f\"Test batch {batch_index} missing keys: \"\n + \", \".join(missing_batch_keys)\n )\n\n images = batch[\"image\"]\n\n study_ids = [\n str(value)\n for value in batch[\"study_id\"]\n ]\n\n if images.ndim != 4:\n raise RuntimeError(\n f\"Unexpected image batch shape: \"\n f\"{tuple(images.shape)}\"\n )\n\n if images.shape[1:] != (\n 21,\n 224,\n 224,\n ):\n raise RuntimeError(\n \"Unexpected test image shape: \"\n + str(tuple(images.shape))\n )\n\n if not torch.isfinite(\n images\n ).all():\n\n raise RuntimeError(\n f\"NaN/Inf detected in test batch \"\n f\"{batch_index}.\"\n )\n\n if len(study_ids) != images.shape[0]:\n raise RuntimeError(\n f\"Study ID count does not match batch size \"\n f\"for batch {batch_index}.\"\n )\n\n images = images.to(\n device,\n non_blocking=True,\n )\n\n # --------------------------------------------------------------\n # Run every model required by the final target plan.\n # This computes each model once per test batch.\n # --------------------------------------------------------------\n\n batch_model_probabilities = {}\n\n with torch.no_grad():\n\n for model_name, model_object in model_registry.items():\n\n model_object.eval()\n\n logits = model_object(\n images\n )\n\n if logits.shape != (\n images.shape[0],\n len(TARGETS),\n ):\n\n raise RuntimeError(\n f\"{model_name} produced unexpected \"\n f\"output shape: {tuple(logits.shape)}\"\n )\n\n if not torch.isfinite(\n logits\n ).all():\n\n raise RuntimeError(\n f\"NaN/Inf logits from \"\n f\"{model_name}.\"\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n if not torch.isfinite(\n probabilities\n ).all():\n\n raise RuntimeError(\n f\"NaN/Inf probabilities from \"\n f\"{model_name}.\"\n )\n\n batch_model_probabilities[\n model_name\n ] = (\n probabilities\n .detach()\n .cpu()\n .numpy()\n )\n\n # --------------------------------------------------------------\n # TARGET-WISE THREE-FOLD ENSEMBLE\n # --------------------------------------------------------------\n\n for sample_index, study_id in enumerate(\n study_ids\n ):\n\n row = {\n \"StudyInstanceUID\": study_id\n }\n\n for target_index, target in enumerate(\n TARGETS\n ):\n\n plan_row = final_inference_plan[\n final_inference_plan[\"target\"]\n == target\n ]\n\n if len(plan_row) != 1:\n raise RuntimeError(\n f\"Expected exactly one plan row \"\n f\"for target {target}.\"\n )\n\n plan_row = plan_row.iloc[0]\n\n fold_model_names = [\n plan_row[\"fold0_model\"],\n plan_row[\"fold1_model\"],\n plan_row[\"fold2_model\"],\n ]\n\n fold_probabilities = np.asarray(\n [\n batch_model_probabilities[\n model_name\n ][\n sample_index,\n target_index\n ]\n for model_name in fold_model_names\n ],\n dtype=np.float64,\n )\n\n if fold_probabilities.shape != (\n 3,\n ):\n raise RuntimeError(\n f\"Expected three fold probabilities \"\n f\"for {target}.\"\n )\n\n if not np.isfinite(\n fold_probabilities\n ).all():\n\n raise RuntimeError(\n f\"Invalid fold probabilities \"\n f\"for {target}.\"\n )\n\n ensemble_probability = float(\n np.mean(\n fold_probabilities\n )\n )\n\n if not np.isfinite(\n ensemble_probability\n ):\n raise RuntimeError(\n f\"Invalid ensemble probability \"\n f\"for {target}.\"\n )\n\n row[\n target\n ] = ensemble_probability\n\n probability_rows.append(row)\n seen_studies.append(study_id)\n\n print(\n f\"Batch {batch_index + 1}: \"\n f\"{len(study_ids)} studies inferred\"\n )\n\n# ----------------------------------------------------------------------\n# BUILD PROBABILITY TABLE\n# ----------------------------------------------------------------------\n\nif len(probability_rows) != expected_test_count:\n raise RuntimeError(\n \"Unexpected number of test prediction rows: \"\n + str(len(probability_rows))\n )\n\ntest_probability_df = pd.DataFrame(\n probability_rows\n)\n\nexpected_probability_columns = [\n \"StudyInstanceUID\"\n] + list(TARGETS)\n\nif test_probability_df.columns.tolist() != (\n expected_probability_columns\n):\n\n raise RuntimeError(\n \"Test probability column ordering is invalid.\"\n )\n\n# ----------------------------------------------------------------------\n# STUDY COVERAGE VALIDATION\n# ----------------------------------------------------------------------\n\nif len(test_probability_df) != expected_test_count:\n raise RuntimeError(\n \"Test probability table must contain exactly 3 rows.\"\n )\n\nactual_ids = [\n str(value)\n for value in test_probability_df[\n \"StudyInstanceUID\"\n ].tolist()\n]\n\nif len(set(actual_ids)) != expected_test_count:\n raise RuntimeError(\n \"Duplicate StudyInstanceUIDs in predictions.\"\n )\n\nif set(actual_ids) != set(\n test_metadata_ids\n):\n\n raise RuntimeError(\n \"Predicted StudyInstanceUIDs do not exactly \"\n \"match the established test studies.\"\n )\n\nprint(\"Test study coverage: PASS\")\nprint(\"3/3 test studies predicted exactly once: PASS\")\n\n# ----------------------------------------------------------------------\n# PROBABILITY VALIDATION\n# ----------------------------------------------------------------------\n\nprobability_values = test_probability_df[\n TARGETS\n].to_numpy(\n dtype=np.float64\n)\n\nif not np.isfinite(\n probability_values\n).all():\n\n raise RuntimeError(\n \"Final probabilities contain NaN/Inf.\"\n )\n\nif (\n (probability_values < 0.0)\n | (probability_values > 1.0)\n).any():\n\n raise RuntimeError(\n \"Final probabilities outside [0, 1].\"\n )\n\nprint(\"Probability validity: PASS\")\n\n# ----------------------------------------------------------------------\n# APPLY OOF-DERIVED TARGET-SPECIFIC THRESHOLDS\n# ----------------------------------------------------------------------\n\ntest_submission_df = pd.DataFrame()\n\ntest_submission_df[\n \"StudyInstanceUID\"\n] = test_probability_df[\n \"StudyInstanceUID\"\n]\n\nfor target in TARGETS:\n\n plan_row = final_inference_plan[\n final_inference_plan[\"target\"]\n == target\n ]\n\n if len(plan_row) != 1:\n raise RuntimeError(\n f\"Missing unique plan row for {target}.\"\n )\n\n threshold = float(\n plan_row.iloc[0][\n \"selected_threshold\"\n ]\n )\n\n probabilities = test_probability_df[\n target\n ].to_numpy(\n dtype=np.float64\n )\n\n test_submission_df[\n target\n ] = (\n probabilities >= threshold\n ).astype(\n np.int64\n )\n\n# ----------------------------------------------------------------------\n# SUBMISSION SCHEMA\n# ----------------------------------------------------------------------\n\nexpected_submission_columns = [\n \"StudyInstanceUID\"\n] + list(TARGETS)\n\nif test_submission_df.columns.tolist() != (\n expected_submission_columns\n):\n\n raise RuntimeError(\n \"Final submission column ordering does not \"\n \"match the established 13-column schema.\"\n )\n\nif test_submission_df.shape != (\n expected_test_count,\n 1 + len(TARGETS),\n):\n\n raise RuntimeError(\n \"Unexpected final submission shape: \"\n + str(test_submission_df.shape)\n )\n\n# ----------------------------------------------------------------------\n# BINARY OUTPUT VALIDATION\n# ----------------------------------------------------------------------\n\nprediction_values = test_submission_df[\n TARGETS\n].to_numpy()\n\nunique_prediction_values = set(\n np.unique(\n prediction_values\n ).tolist()\n)\n\nif not unique_prediction_values.issubset(\n {0, 1}\n):\n\n raise RuntimeError(\n \"Final predictions are not binary.\"\n )\n\nprint(\"Threshold application: PASS\")\nprint(\"Binary prediction validation: PASS\")\nprint(\"Submission schema: PASS\")\n\n# ----------------------------------------------------------------------\n# DISPLAY FINAL PROBABILITIES\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL TEST PROBABILITIES\")\nprint(\"=\" * 70)\n\nprint(\n test_probability_df.to_string(\n index=False\n )\n)\n\n# ----------------------------------------------------------------------\n# DISPLAY FINAL BINARY PREDICTIONS\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL TEST PREDICTIONS\")\nprint(\"=\" * 70)\n\nprint(\n test_submission_df.to_string(\n index=False\n )\n)\n\n# ----------------------------------------------------------------------\n# SAVE FINAL PROBABILITY TABLE\n# ----------------------------------------------------------------------\n\nprobability_path = os.path.join(\n audit_dir,\n \"cell67_final_test_probabilities.csv\",\n)\n\ntest_probability_df.to_csv(\n probability_path,\n index=False,\n)\n\nif not os.path.isfile(\n probability_path\n):\n\n raise RuntimeError(\n \"Final probability file was not created.\"\n )\n\n# ----------------------------------------------------------------------\n# SAVE FINAL SUBMISSION\n# ----------------------------------------------------------------------\n\nsubmission_path = os.path.join(\n audit_dir,\n \"cell67_final_submission.csv\",\n)\n\ntest_submission_df.to_csv(\n submission_path,\n index=False,\n)\n\nif not os.path.isfile(\n submission_path\n):\n\n raise RuntimeError(\n \"Final submission file was not created.\"\n )\n\n# ----------------------------------------------------------------------\n# FINAL VALIDATION\n# ----------------------------------------------------------------------\n\nsaved_submission = pd.read_csv(\n submission_path\n)\n\nif saved_submission.shape != (\n expected_test_count,\n 1 + len(TARGETS),\n):\n\n raise RuntimeError(\n \"Saved submission shape validation failed.\"\n )\n\nif saved_submission.columns.tolist() != (\n expected_submission_columns\n):\n\n raise RuntimeError(\n \"Saved submission column validation failed.\"\n )\n\nsaved_ids = [\n str(value)\n for value in saved_submission[\n \"StudyInstanceUID\"\n ].tolist()\n]\n\nif set(saved_ids) != set(\n test_metadata_ids\n):\n\n raise RuntimeError(\n \"Saved submission StudyInstanceUID coverage failed.\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 67 VERDICT\")\nprint(\"=\" * 70)\n\nprint(\"Existing test DataLoader used: PASS\")\nprint(\"3 test studies loaded: PASS\")\nprint(\"21-channel input: PASS\")\nprint(\"12-target output: PASS\")\nprint(\"Three-fold inference: PASS\")\nprint(\"Target-wise model selection preserved: PASS\")\nprint(\"OOF-derived thresholds preserved: PASS\")\nprint(\"Probability validity: PASS\")\nprint(\"Study coverage: PASS\")\nprint(\"Submission schema: PASS\")\nprint(\"Binary predictions: PASS\")\nprint(\"Final submission saved: PASS\")\nprint(\"No training performed: PASS\")\nprint(\"No checkpoint modified: PASS\")\nprint(\"No validation data used: PASS\")\n\nprint()\nprint(\"Probability file:\")\nprint(probability_path)\n\nprint()\nprint(\"Final submission:\")\nprint(submission_path)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 67 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T12:07:10.603403Z","iopub.execute_input":"2026-08-11T12:07:10.603766Z","iopub.status.idle":"2026-08-11T12:07:13.82911Z","shell.execute_reply.started":"2026-08-11T12:07:10.603726Z","shell.execute_reply":"2026-08-11T12:07:13.827999Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 68R - FINAL SUBMISSION INTEGRITY AUDIT\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"CELL 68R - FINAL SUBMISSION INTEGRITY AUDIT\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# Established paths\n# ----------------------------------------------------------------------\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nFINAL_SUBMISSION_PATH = os.path.join(\n AUDIT_DIR,\n \"cell67_final_submission.csv\"\n)\n\nPROBABILITY_PATH = os.path.join(\n AUDIT_DIR,\n \"cell67_final_test_probabilities.csv\"\n)\n\nCELL61_PLAN_PATH = os.path.join(\n AUDIT_DIR,\n \"cell61_final_oof_backed_model_plan.csv\"\n)\n\nCELL63_PLAN_PATH = os.path.join(\n AUDIT_DIR,\n \"cell63_final_three_fold_inference_plan.csv\"\n)\n\n# ----------------------------------------------------------------------\n# Established target ordering\n# ----------------------------------------------------------------------\n\nEXPECTED_TARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\",\n]\n\nEXPECTED_COLUMNS = [\n \"StudyInstanceUID\"\n] + EXPECTED_TARGETS\n\n# ----------------------------------------------------------------------\n# Required notebook objects\n# ----------------------------------------------------------------------\n\nrequired_objects = [\n \"TARGETS\",\n \"test\",\n \"sample_submission\",\n]\n\nmissing_objects = [\n name\n for name in required_objects\n if name not in globals()\n]\n\nif missing_objects:\n raise RuntimeError(\n \"Missing notebook objects: \"\n + \", \".join(missing_objects)\n )\n\nif list(TARGETS) != EXPECTED_TARGETS:\n raise RuntimeError(\n \"Notebook TARGETS ordering does not match the established \"\n \"12-target configuration.\"\n )\n\nprint(\"Required notebook objects: PASS\")\nprint(\"Target count:\", len(TARGETS))\nprint(\"Target ordering: PASS\")\n\n# ----------------------------------------------------------------------\n# Required files\n# ----------------------------------------------------------------------\n\nrequired_files = {\n \"Final submission\": FINAL_SUBMISSION_PATH,\n \"Final probability file\": PROBABILITY_PATH,\n \"Cell-61 model plan\": CELL61_PLAN_PATH,\n \"Cell-63 inference plan\": CELL63_PLAN_PATH,\n}\n\nfor name, path in required_files.items():\n if not os.path.isfile(path):\n raise RuntimeError(\n f\"{name} not found: {path}\"\n )\n\n print(f\"{name}: EXISTS\")\n\n# ----------------------------------------------------------------------\n# Load final submission\n# ----------------------------------------------------------------------\n\nsubmission = pd.read_csv(\n FINAL_SUBMISSION_PATH\n)\n\nprint()\nprint(\"Final submission shape:\", submission.shape)\n\n# ----------------------------------------------------------------------\n# Submission schema\n# ----------------------------------------------------------------------\n\nif list(submission.columns) != EXPECTED_COLUMNS:\n raise RuntimeError(\n \"Final submission schema mismatch.\\n\"\n f\"Expected: {EXPECTED_COLUMNS}\\n\"\n f\"Found: {list(submission.columns)}\"\n )\n\nif list(sample_submission.columns) != EXPECTED_COLUMNS:\n raise RuntimeError(\n \"sample_submission schema mismatch.\"\n )\n\nprint(\"Submission schema: PASS\")\nprint(\"Column ordering: PASS\")\nprint(\"Column count:\", len(submission.columns))\nprint(\"sample_submission schema: PASS\")\n\n# ----------------------------------------------------------------------\n# Row count\n# ----------------------------------------------------------------------\n\nif len(submission) != len(test):\n raise RuntimeError(\n f\"Submission row count {len(submission)} does not match \"\n f\"test study count {len(test)}.\"\n )\n\nif len(submission) != 3:\n raise RuntimeError(\n f\"Expected 3 test studies, found {len(submission)}.\"\n )\n\nprint(\"Submission row count:\", len(submission))\nprint(\"Expected test studies: 3\")\nprint(\"Row count: PASS\")\n\n# ----------------------------------------------------------------------\n# StudyInstanceUID validation\n# ----------------------------------------------------------------------\n\nsubmission_ids = submission[\n \"StudyInstanceUID\"\n].astype(str)\n\ntest_ids = test[\n \"StudyInstanceUID\"\n].astype(str)\n\nif submission_ids.isna().any():\n raise RuntimeError(\n \"Submission contains missing StudyInstanceUID values.\"\n )\n\nif submission_ids.duplicated().any():\n raise RuntimeError(\n \"Submission contains duplicate StudyInstanceUID values.\"\n )\n\nif set(submission_ids) != set(test_ids):\n raise RuntimeError(\n \"Submission StudyInstanceUID set does not exactly match \"\n \"the recovered test studies.\"\n )\n\nprint(\"StudyInstanceUID completeness: PASS\")\nprint(\"StudyInstanceUID uniqueness: PASS\")\nprint(\"Test-study coverage: PASS\")\nprint(\"3/3 expected test studies present exactly once: PASS\")\n\n# ----------------------------------------------------------------------\n# Target validation\n# ----------------------------------------------------------------------\n\nfor target in EXPECTED_TARGETS:\n\n if submission[target].isna().any():\n raise RuntimeError(\n f\"Target '{target}' contains missing values.\"\n )\n\n numeric_values = pd.to_numeric(\n submission[target],\n errors=\"raise\"\n )\n\n unique_values = set(\n numeric_values.tolist()\n )\n\n if not unique_values.issubset({0, 1}):\n raise RuntimeError(\n f\"Target '{target}' contains non-binary values: \"\n f\"{sorted(unique_values)}\"\n )\n\n if not np.isfinite(\n numeric_values.to_numpy(dtype=np.float32)\n ).all():\n raise RuntimeError(\n f\"Target '{target}' contains NaN or Inf values.\"\n )\n\nprint(\"All target columns present: PASS\")\nprint(\"No missing target values: PASS\")\nprint(\"Binary target values: PASS\")\nprint(\"Target numeric validity: PASS\")\nprint(\"NaN/Inf validation: PASS\")\n\n# ----------------------------------------------------------------------\n# Sample submission study coverage\n# ----------------------------------------------------------------------\n\nsample_ids = (\n sample_submission[\"StudyInstanceUID\"]\n .astype(str)\n)\n\nif set(sample_ids) != set(submission_ids):\n raise RuntimeError(\n \"Submission studies do not match sample_submission studies.\"\n )\n\nprint(\"Sample-submission study coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# Probability file validation\n# ----------------------------------------------------------------------\n\nprobabilities = pd.read_csv(\n PROBABILITY_PATH\n)\n\nexpected_probability_columns = [\n \"StudyInstanceUID\"\n] + EXPECTED_TARGETS\n\nif list(probabilities.columns) != expected_probability_columns:\n raise RuntimeError(\n \"Probability file schema mismatch.\"\n )\n\nif len(probabilities) != 3:\n raise RuntimeError(\n \"Probability file must contain exactly 3 studies.\"\n )\n\nprobability_ids = (\n probabilities[\"StudyInstanceUID\"]\n .astype(str)\n)\n\nif probability_ids.duplicated().any():\n raise RuntimeError(\n \"Probability file contains duplicate StudyInstanceUID values.\"\n )\n\nif set(probability_ids) != set(test_ids):\n raise RuntimeError(\n \"Probability file study coverage does not match test studies.\"\n )\n\nprobability_matrix = probabilities[\n EXPECTED_TARGETS\n].to_numpy(dtype=np.float32)\n\nif not np.isfinite(probability_matrix).all():\n raise RuntimeError(\n \"Probability file contains NaN or Inf values.\"\n )\n\nif (\n (probability_matrix < 0.0)\n | (probability_matrix > 1.0)\n).any():\n raise RuntimeError(\n \"Probability file contains values outside [0, 1].\"\n )\n\nprint(\"Probability file schema: PASS\")\nprint(\"Probability validity: PASS\")\nprint(\"Probability range [0,1]: PASS\")\nprint(\"Probability study coverage: PASS\")\n\n# ======================================================================\n# CRITICAL PLAN VALIDATION\n#\n# Cell 61 = authoritative OOF-backed model + threshold selection\n# Cell 63 = authoritative three-fold model mapping\n#\n# Cell 63 DOES NOT contain selected_threshold.\n# Therefore thresholds MUST come from Cell 61.\n# ======================================================================\n\ncell61_plan = pd.read_csv(\n CELL61_PLAN_PATH\n)\n\ncell63_plan = pd.read_csv(\n CELL63_PLAN_PATH\n)\n\n# ----------------------------------------------------------------------\n# Cell 61 schema\n# ----------------------------------------------------------------------\n\nrequired_cell61_columns = [\n \"target\",\n \"selected_model\",\n \"selected_threshold\",\n \"selected_best_f1\",\n]\n\nmissing_cell61_columns = [\n column\n for column in required_cell61_columns\n if column not in cell61_plan.columns\n]\n\nif missing_cell61_columns:\n raise RuntimeError(\n \"Cell-61 model plan missing required columns: \"\n + \", \".join(missing_cell61_columns)\n )\n\nif len(cell61_plan) != 12:\n raise RuntimeError(\n f\"Cell-61 model plan must contain 12 targets, \"\n f\"found {len(cell61_plan)}.\"\n )\n\nif list(cell61_plan[\"target\"]) != EXPECTED_TARGETS:\n raise RuntimeError(\n \"Cell-61 target ordering does not match TARGETS.\"\n )\n\nprint(\"Cell-61 model plan schema: PASS\")\nprint(\"Cell-61 target ordering: PASS\")\nprint(\"Cell-61 target coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# Cell 63 schema\n#\n# Deliberately do NOT require selected_threshold here.\n# ----------------------------------------------------------------------\n\nrequired_cell63_columns = [\n \"target\",\n \"selected_model\",\n \"fold0_model\",\n \"fold1_model\",\n \"fold2_model\",\n]\n\nmissing_cell63_columns = [\n column\n for column in required_cell63_columns\n if column not in cell63_plan.columns\n]\n\nif missing_cell63_columns:\n raise RuntimeError(\n \"Cell-63 inference plan missing required columns: \"\n + \", \".join(missing_cell63_columns)\n )\n\nif len(cell63_plan) != 12:\n raise RuntimeError(\n f\"Cell-63 inference plan must contain 12 targets, \"\n f\"found {len(cell63_plan)}.\"\n )\n\nif list(cell63_plan[\"target\"]) != EXPECTED_TARGETS:\n raise RuntimeError(\n \"Cell-63 target ordering does not match TARGETS.\"\n )\n\nprint(\"Cell-63 model-mapping schema: PASS\")\nprint(\"Cell-63 target ordering: PASS\")\nprint(\"Cell-63 target coverage: PASS\")\n\n# ----------------------------------------------------------------------\n# Cell 61 vs Cell 63 model-selection consistency\n# ----------------------------------------------------------------------\n\ncell61_models = dict(\n zip(\n cell61_plan[\"target\"],\n cell61_plan[\"selected_model\"]\n )\n)\n\ncell63_models = dict(\n zip(\n cell63_plan[\"target\"],\n cell63_plan[\"selected_model\"]\n )\n)\n\nif cell61_models != cell63_models:\n raise RuntimeError(\n \"Cell-61 and Cell-63 selected-model assignments disagree.\"\n )\n\nprint(\"Cell-61 vs Cell-63 selected-model consistency: PASS\")\n\n# ----------------------------------------------------------------------\n# Thresholds come ONLY from Cell 61\n# ----------------------------------------------------------------------\n\nthreshold_map = dict(\n zip(\n cell61_plan[\"target\"],\n pd.to_numeric(\n cell61_plan[\"selected_threshold\"],\n errors=\"raise\"\n )\n )\n)\n\nif set(threshold_map.keys()) != set(EXPECTED_TARGETS):\n raise RuntimeError(\n \"Cell-61 threshold coverage does not match TARGETS.\"\n )\n\nfor target in EXPECTED_TARGETS:\n\n threshold = float(\n threshold_map[target]\n )\n\n if not np.isfinite(threshold):\n raise RuntimeError(\n f\"Threshold for '{target}' is not finite.\"\n )\n\n if not 0.0 <= threshold <= 1.0:\n raise RuntimeError(\n f\"Threshold for '{target}' is outside [0,1]: \"\n f\"{threshold}\"\n )\n\nprint(\"OOF-derived thresholds loaded from Cell 61: PASS\")\nprint(\"Selected threshold range: PASS\")\n\n# ----------------------------------------------------------------------\n# Validate the exact three-fold mapping from Cell 63\n# ----------------------------------------------------------------------\n\nexpected_model_names = {\n \"baseline\": [\n \"baseline_fold0\",\n \"baseline_fold1\",\n \"baseline_fold2\",\n ],\n \"fine_tuned\": [\n \"finetuned_fold0\",\n \"finetuned_fold1\",\n \"finetuned_fold2\",\n ],\n}\n\nfor _, row in cell63_plan.iterrows():\n\n selected_model = row[\"selected_model\"]\n\n if selected_model not in expected_model_names:\n raise RuntimeError(\n f\"Unexpected selected_model '{selected_model}' \"\n f\"for target '{row['target']}'.\"\n )\n\n expected_folds = expected_model_names[\n selected_model\n ]\n\n actual_folds = [\n row[\"fold0_model\"],\n row[\"fold1_model\"],\n row[\"fold2_model\"],\n ]\n\n if actual_folds != expected_folds:\n raise RuntimeError(\n f\"Invalid fold mapping for target \"\n f\"'{row['target']}'.\\n\"\n f\"Expected: {expected_folds}\\n\"\n f\"Found: {actual_folds}\"\n )\n\nprint(\"Three-fold model mapping: PASS\")\nprint(\"Baseline/fine-tuned family mapping: PASS\")\n\n# ======================================================================\n# FINAL PROBABILITY -> BINARY SUBMISSION CONSISTENCY\n# ======================================================================\n\nfor target in EXPECTED_TARGETS:\n\n threshold = float(\n threshold_map[target]\n )\n\n expected_binary = (\n probabilities[target]\n .to_numpy(dtype=np.float32)\n >= threshold\n ).astype(int)\n\n actual_binary = (\n submission[target]\n .to_numpy(dtype=int)\n )\n\n if not np.array_equal(\n expected_binary,\n actual_binary\n ):\n raise RuntimeError(\n f\"Final submission mismatch for target \"\n f\"'{target}'. \"\n f\"Submission values do not equal the Cell-61 \"\n f\"OOF-derived threshold applied to Cell-67 probabilities.\"\n )\n\nprint(\"Probability-to-binary consistency: PASS\")\nprint(\"OOF-derived threshold application: PASS\")\n\n# ----------------------------------------------------------------------\n# Final submission preview\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL SUBMISSION\")\nprint(\"=\" * 70)\n\nprint(submission.to_string(index=False))\n\n# ----------------------------------------------------------------------\n# Final verdict\n# ----------------------------------------------------------------------\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 68R VERDICT\")\nprint(\"=\" * 70)\n\nprint(\"Final submission exists: PASS\")\nprint(\"Final submission schema: PASS\")\nprint(\"Target ordering: PASS\")\nprint(\"3/3 test studies covered: PASS\")\nprint(\"StudyInstanceUID uniqueness: PASS\")\nprint(\"All 12 targets present: PASS\")\nprint(\"No missing target values: PASS\")\nprint(\"Binary prediction validity: PASS\")\nprint(\"Probability file validity: PASS\")\nprint(\"Cell-61 OOF-backed model selection: PASS\")\nprint(\"Cell-61 OOF-derived thresholds: PASS\")\nprint(\"Cell-63 three-fold mapping: PASS\")\nprint(\"Probability-to-submission consistency: PASS\")\n\nprint()\nprint(\"No training performed.\")\nprint(\"No checkpoint modified.\")\nprint(\"No test labels used.\")\nprint(\"No additional inference performed.\")\n\nprint()\nprint(\"Final submission:\")\nprint(FINAL_SUBMISSION_PATH)\n\nprint()\nprint(\"Final probability file:\")\nprint(PROBABILITY_PATH)\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 68R COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T12:10:52.882128Z","iopub.execute_input":"2026-08-11T12:10:52.882479Z","iopub.status.idle":"2026-08-11T12:10:52.94176Z","shell.execute_reply.started":"2026-08-11T12:10:52.882449Z","shell.execute_reply":"2026-08-11T12:10:52.940912Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======================================================================\n# CELL 69 - FINAL KAGGLE PROBABILITY SUBMISSION\n# ======================================================================\n\nimport os\nimport numpy as np\nimport pandas as pd\n\nprint(\"=\" * 70)\nprint(\"CELL 69 - FINAL KAGGLE PROBABILITY SUBMISSION\")\nprint(\"=\" * 70)\n\n# ----------------------------------------------------------------------\n# REQUIRED CONFIGURATION\n# ----------------------------------------------------------------------\n\nTARGETS = [\n \"ACL\",\n \"MCL\",\n \"Medial Meniscus\",\n \"Lateral Meniscus\",\n \"Medial OA\",\n \"Lateral OA\",\n \"PF OA\",\n \"Effusion\",\n \"Synovitis\",\n \"Baker's\",\n \"Contusion\",\n \"Fracture\",\n]\n\nEXPECTED_COLUMNS = [\n \"StudyInstanceUID\"\n] + TARGETS\n\nAUDIT_DIR = \"/kaggle/working/rsna_knee_audit\"\n\nPROBABILITY_FILE = os.path.join(\n AUDIT_DIR,\n \"cell67_final_test_probabilities.csv\"\n)\n\nFINAL_SUBMISSION = \"/kaggle/working/submission.csv\"\n\n# ----------------------------------------------------------------------\n# FILE EXISTENCE\n# ----------------------------------------------------------------------\n\nif not os.path.isfile(PROBABILITY_FILE):\n raise FileNotFoundError(\n \"Validated probability file not found: \"\n + PROBABILITY_FILE\n )\n\nprint(\"Validated probability file: EXISTS\")\n\n# ----------------------------------------------------------------------\n# LOAD CONTINUOUS PROBABILITIES\n# ----------------------------------------------------------------------\n\nprob_df = pd.read_csv(PROBABILITY_FILE)\n\nprint(\"Probability input shape:\", prob_df.shape)\n\n# ----------------------------------------------------------------------\n# COLUMN VALIDATION\n# ----------------------------------------------------------------------\n\nif list(prob_df.columns) != EXPECTED_COLUMNS:\n raise RuntimeError(\n \"Probability file column ordering/schema mismatch.\\n\"\n \"Expected: \"\n + str(EXPECTED_COLUMNS)\n + \"\\nFound: \"\n + str(list(prob_df.columns))\n )\n\nprint(\"Kaggle submission schema: PASS\")\nprint(\"Column ordering: PASS\")\nprint(\"Column count:\", len(prob_df.columns))\n\n# ----------------------------------------------------------------------\n# ROW VALIDATION\n# ----------------------------------------------------------------------\n\nif len(prob_df) != 3:\n raise RuntimeError(\n \"Expected 3 example test studies, found \"\n + str(len(prob_df))\n )\n\nprint(\"Submission row count:\", len(prob_df))\nprint(\"Expected test studies: 3\")\nprint(\"Row count: PASS\")\n\n# ----------------------------------------------------------------------\n# STUDY ID VALIDATION\n# ----------------------------------------------------------------------\n\nif prob_df[\"StudyInstanceUID\"].isna().any():\n raise RuntimeError(\n \"StudyInstanceUID contains missing values.\"\n )\n\nif prob_df[\"StudyInstanceUID\"].duplicated().any():\n raise RuntimeError(\n \"Duplicate StudyInstanceUID detected.\"\n )\n\nprint(\"StudyInstanceUID completeness: PASS\")\nprint(\"StudyInstanceUID uniqueness: PASS\")\n\n# ----------------------------------------------------------------------\n# TARGET NUMERIC VALIDATION\n# ----------------------------------------------------------------------\n\nfor target in TARGETS:\n\n if not pd.api.types.is_numeric_dtype(prob_df[target]):\n raise RuntimeError(\n f\"Target column is not numeric: {target}\"\n )\n\nprint(\"Target numeric types: PASS\")\n\n# ----------------------------------------------------------------------\n# PROBABILITY VALIDATION\n# ----------------------------------------------------------------------\n\ntarget_values = prob_df[TARGETS].to_numpy(\n dtype=np.float64\n)\n\nif not np.isfinite(target_values).all():\n raise RuntimeError(\n \"Probability file contains NaN or Inf values.\"\n )\n\nprint(\"NaN/Inf validation: PASS\")\n\nif (target_values < 0).any():\n raise RuntimeError(\n \"Probability below 0 detected.\"\n )\n\nif (target_values > 1).any():\n raise RuntimeError(\n \"Probability above 1 detected.\"\n )\n\nprint(\"Probability range [0,1]: PASS\")\n\n# ----------------------------------------------------------------------\n# IMPORTANT:\n# DO NOT THRESHOLD THE PROBABILITIES.\n# ROC-AUC REQUIRES CONTINUOUS CONFIDENCE SCORES.\n# ----------------------------------------------------------------------\n\nif np.all(\n np.isin(\n target_values,\n [0.0, 1.0]\n )\n):\n raise RuntimeError(\n \"All target values are binary 0/1. \"\n \"Kaggle requires continuous confidence scores. \"\n \"Use the Cell 67 probability file, not the thresholded submission.\"\n )\n\nprint(\"Continuous confidence scores: PASS\")\n\n# ----------------------------------------------------------------------\n# SAVE EXACT KAGGLE FILENAME\n# ----------------------------------------------------------------------\n\nprob_df.to_csv(\n FINAL_SUBMISSION,\n index=False\n)\n\nif not os.path.isfile(FINAL_SUBMISSION):\n raise RuntimeError(\n \"Final submission.csv was not created.\"\n )\n\nprint(\"Final submission file: CREATED\")\nprint(\"Path:\", FINAL_SUBMISSION)\n\n# ----------------------------------------------------------------------\n# RELOAD AND FINAL AUDIT\n# ----------------------------------------------------------------------\n\nfinal_df = pd.read_csv(FINAL_SUBMISSION)\n\nif list(final_df.columns) != EXPECTED_COLUMNS:\n raise RuntimeError(\n \"Reloaded submission schema mismatch.\"\n )\n\nif final_df.shape != (3, 13):\n raise RuntimeError(\n \"Unexpected final submission shape: \"\n + str(final_df.shape)\n )\n\nfinal_values = final_df[TARGETS].to_numpy(\n dtype=np.float64\n)\n\nif not np.isfinite(final_values).all():\n raise RuntimeError(\n \"Final submission contains NaN/Inf.\"\n )\n\nif (final_values < 0).any() or (final_values > 1).any():\n raise RuntimeError(\n \"Final submission contains values outside [0,1].\"\n )\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL KAGGLE SUBMISSION AUDIT\")\nprint(\"=\" * 70)\n\nprint(\"Shape:\", final_df.shape)\nprint(\"Schema: PASS\")\nprint(\"Column ordering: PASS\")\nprint(\"StudyInstanceUID completeness: PASS\")\nprint(\"StudyInstanceUID uniqueness: PASS\")\nprint(\"12 target columns: PASS\")\nprint(\"Continuous confidence scores: PASS\")\nprint(\"Probability range [0,1]: PASS\")\nprint(\"NaN/Inf validation: PASS\")\n\nprint()\nprint(\"Final submission preview:\")\nprint(final_df.to_string(index=False))\n\nprint()\nprint(\"=\" * 70)\nprint(\"CELL 69 VERDICT\")\nprint(\"=\" * 70)\n\nprint(\"Kaggle ROC-AUC probability format: PASS\")\nprint(\"Continuous probabilities preserved: PASS\")\nprint(\"Thresholding NOT applied: PASS\")\nprint(\"Submission filename submission.csv: PASS\")\nprint(\"Final submission schema: PASS\")\nprint(\"Final submission ready for Kaggle upload: PASS\")\n\nprint()\nprint(\"Final submission:\")\nprint(FINAL_SUBMISSION)\n\nprint()\nprint(\"No training performed.\")\nprint(\"No inference performed.\")\nprint(\"No checkpoint modified.\")\nprint(\"No test labels used.\")\nprint(\"=\" * 70)\nprint(\"CELL 69 COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-11T12:20:47.380633Z","iopub.execute_input":"2026-08-11T12:20:47.381041Z","iopub.status.idle":"2026-08-11T12:20:47.418958Z","shell.execute_reply.started":"2026-08-11T12:20:47.381009Z","shell.execute_reply":"2026-08-11T12:20:47.417629Z"}},"outputs":[],"execution_count":null}]}