{"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":"from pathlib import Path\nimport os\n\nroot = Path(\"/kaggle/input\")\n\nfor item in root.iterdir():\n print(item)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:45:47.61225Z","iopub.execute_input":"2026-08-17T16:45:47.612654Z","iopub.status.idle":"2026-08-17T16:45:47.624856Z","shell.execute_reply.started":"2026-08-17T16:45:47.612609Z","shell.execute_reply":"2026-08-17T16:45:47.623774Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions\")\n\nfor item in root.iterdir():\n print(item)\n ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:46:34.204884Z","iopub.execute_input":"2026-08-17T16:46:34.205374Z","iopub.status.idle":"2026-08-17T16:46:34.211586Z","shell.execute_reply.started":"2026-08-17T16:46:34.205342Z","shell.execute_reply":"2026-08-17T16:46:34.210716Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\nfor item in sorted(root.iterdir()):\n print(item)\n ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:47:00.460498Z","iopub.execute_input":"2026-08-17T16:47:00.460856Z","iopub.status.idle":"2026-08-17T16:47:00.467487Z","shell.execute_reply.started":"2026-08-17T16:47:00.46083Z","shell.execute_reply":"2026-08-17T16:47:00.466511Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\nfor filename in [\n \"train.csv\",\n \"test.csv\",\n \"train_series.csv\",\n \"test_series.csv\",\n \"sample_submission.csv\"\n]:\n print(\"\\n\" + \"=\" * 70)\n print(filename)\n\n df = pd.read_csv(root / filename)\n\n print(\"Shape:\", df.shape)\n print(\"Columns:\")\n print(df.columns.tolist())\n\n print(\"\\nFirst 3 rows:\")\n display(df.head(3))\n ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:47:31.879914Z","iopub.execute_input":"2026-08-17T16:47:31.880411Z","iopub.status.idle":"2026-08-17T16:47:32.644285Z","shell.execute_reply.started":"2026-08-17T16:47:31.88038Z","shell.execute_reply":"2026-08-17T16:47:32.643421Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\ntrain = pd.read_csv(root / \"train.csv\")\n\nprint(\"Shape:\", train.shape)\nprint(\"\\nColumns:\")\nprint(train.columns.tolist())\n\nprint(\"\\nFirst 5 rows:\")\ndisplay(train.head())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:48:09.518561Z","iopub.execute_input":"2026-08-17T16:48:09.519507Z","iopub.status.idle":"2026-08-17T16:48:09.652582Z","shell.execute_reply.started":"2026-08-17T16:48:09.519438Z","shell.execute_reply":"2026-08-17T16:48:09.651685Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\ntrain = pd.read_csv(root / \"train.csv\")\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(\"Shape:\", train.shape)\n\nprint(\"\\nData types:\")\nprint(train[targets].dtypes)\n\nprint(\"\\nMissing values:\")\nprint(train[targets].isna().sum())\n\nprint(\"\\nUnique values for every target:\")\nfor col in targets:\n print(f\"\\n--- {col} ---\")\n print(train[col].value_counts(dropna=False).sort_index())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:49:14.192162Z","iopub.execute_input":"2026-08-17T16:49:14.192461Z","iopub.status.idle":"2026-08-17T16:49:14.344132Z","shell.execute_reply.started":"2026-08-17T16:49:14.192436Z","shell.execute_reply":"2026-08-17T16:49:14.34313Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\ntrain = pd.read_csv(root / \"train.csv\")\ntest = pd.read_csv(root / \"test.csv\")\ntrain_series = pd.read_csv(root / \"train_series.csv\")\ntest_series = pd.read_csv(root / \"test_series.csv\")\n\nprint(\"TRAIN\")\nprint(train.shape)\n\nprint(\"\\nTEST\")\nprint(test.shape)\n\nprint(\"\\nTRAIN SERIES\")\nprint(train_series.shape)\n\nprint(\"\\nTEST SERIES\")\nprint(test_series.shape)\n\nprint(\"\\nLabeled train rows:\")\nprint(train[\"ACL\"].notna().sum())\n\nprint(\"\\nFirst 5 labeled StudyInstanceUIDs:\")\nprint(train.loc[train[\"ACL\"].notna(), \"StudyInstanceUID\"].head().tolist())\n\nprint(\"\\nFirst 5 unlabeled StudyInstanceUIDs:\")\nprint(train.loc[train[\"ACL\"].isna(), \"StudyInstanceUID\"].head().tolist())\n\nprint(\"\\nLast 5 rows of train:\")\ndisplay(train.tail())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:50:11.726556Z","iopub.execute_input":"2026-08-17T16:50:11.726981Z","iopub.status.idle":"2026-08-17T16:50:11.930252Z","shell.execute_reply.started":"2026-08-17T16:50:11.726951Z","shell.execute_reply":"2026-08-17T16:50:11.929275Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\ntrain_series = pd.read_csv(root / \"train_series.csv\")\n\nprint(\"Shape:\", train_series.shape)\n\nprint(\"\\nColumns:\")\nprint(train_series.columns.tolist())\n\nprint(\"\\nFirst 10 rows:\")\ndisplay(train_series.head(10))\n\nprint(\"\\nData types:\")\nprint(train_series.dtypes)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:50:49.52549Z","iopub.execute_input":"2026-08-17T16:50:49.526182Z","iopub.status.idle":"2026-08-17T16:50:49.609471Z","shell.execute_reply.started":"2026-08-17T16:50:49.526133Z","shell.execute_reply":"2026-08-17T16:50:49.608178Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"\\nNumber of unique studies:\")\nprint(train_series[\"StudyInstanceUID\"].nunique())\n\nprint(\"\\nNumber of unique series:\")\nprint(train_series[\"SeriesInstanceUID\"].nunique())\n\nprint(\"\\nSeries per study:\")\nseries_per_study = train_series.groupby(\"StudyInstanceUID\")[\"SeriesInstanceUID\"].nunique()\n\nprint(series_per_study.describe())\n\nprint(\"\\nFirst 20 studies:\")\ndisplay(series_per_study.head(20))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:51:02.576069Z","iopub.execute_input":"2026-08-17T16:51:02.576665Z","iopub.status.idle":"2026-08-17T16:51:02.619642Z","shell.execute_reply.started":"2026-08-17T16:51:02.576635Z","shell.execute_reply":"2026-08-17T16:51:02.618713Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\nimport pandas as pd\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\ntrain_series = pd.read_csv(root / \"train_series.csv\")\n\nprint(\"Shape:\", train_series.shape)\n\nprint(\"\\nColumns:\")\nprint(train_series.columns.tolist())\n\nprint(\"\\nFirst 20 rows:\")\ndisplay(train_series.head(20))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:51:44.752576Z","iopub.execute_input":"2026-08-17T16:51:44.752889Z","iopub.status.idle":"2026-08-17T16:51:44.829197Z","shell.execute_reply.started":"2026-08-17T16:51:44.752865Z","shell.execute_reply":"2026-08-17T16:51:44.828177Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nfrom pathlib import Path\nimport pandas as pd\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\nstudy_id = pd.read_csv(root / \"train_series.csv\").iloc[0][\"StudyInstanceUID\"]\nseries_id = pd.read_csv(root / \"train_series.csv\").iloc[0][\"SeriesInstanceUID\"]\n\nprint(\"StudyInstanceUID:\")\nprint(study_id)\n\nprint(\"\\nSeriesInstanceUID:\")\nprint(series_id)\n\nprint(\"\\nSearching for DICOM files...\")\n\nseries_path = root / \"train_series\" / str(study_id) / str(series_id)\n\nprint(\"Expected series path:\")\nprint(series_path)\n\nif series_path.exists():\n files = sorted([p for p in series_path.rglob(\"*\") if p.is_file()])\n print(\"\\nNumber of files:\", len(files))\n\n print(\"\\nFirst 10 files:\")\n for f in files[:10]:\n print(f)\nelse:\n print(\"\\nSeries path does not exist at this exact structure.\")\n print(\"We will inspect the directory structure next.\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:52:12.963049Z","iopub.execute_input":"2026-08-17T16:52:12.963738Z","iopub.status.idle":"2026-08-17T16:52:13.106415Z","shell.execute_reply.started":"2026-08-17T16:52:12.963704Z","shell.execute_reply":"2026-08-17T16:52:13.105355Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pydicom\nimport matplotlib.pyplot as plt\nfrom pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\nstudy_id = \"1.2.826.0.1.3680043.8.498.10004873229099053869093324292195817260\"\nseries_id = \"1.2.826.0.1.3680043.8.498.12343110195036213483454091715412333772\"\n\nseries_path = root / \"train_series\" / study_id / series_id\n\nfiles = sorted(series_path.glob(\"*.dcm\"))\n\nprint(\"Number of DICOM files:\", len(files))\n\n# Load all slices\nslices = []\n\nfor f in files:\n ds = pydicom.dcmread(f)\n slices.append((ds, ds.pixel_array))\n\n# Show metadata from first slice\nds = slices[0][0]\n\nprint(\"\\nDICOM metadata\")\nprint(\"Rows:\", ds.Rows)\nprint(\"Columns:\", ds.Columns)\nprint(\"Modality:\", getattr(ds, \"Modality\", \"N/A\"))\nprint(\"Series Description:\", getattr(ds, \"SeriesDescription\", \"N/A\"))\nprint(\"Instance Number:\", getattr(ds, \"InstanceNumber\", \"N/A\"))\nprint(\"Slice Thickness:\", getattr(ds, \"SliceThickness\", \"N/A\"))\nprint(\"Pixel Spacing:\", getattr(ds, \"PixelSpacing\", \"N/A\"))\n\n# Show 6 evenly spaced slices\nindices = [0, 4, 8, 12, 16, 21]\n\nfig, axes = plt.subplots(2, 3, figsize=(15, 10))\n\nfor ax, idx in zip(axes.ravel(), indices):\n image = slices[idx][1]\n ax.imshow(image, cmap=\"gray\")\n ax.set_title(f\"Slice {idx + 1}/{len(slices)}\")\n ax.axis(\"off\")\n\nplt.tight_layout()\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:52:47.734696Z","iopub.execute_input":"2026-08-17T16:52:47.735785Z","iopub.status.idle":"2026-08-17T16:52:50.158393Z","shell.execute_reply.started":"2026-08-17T16:52:47.735741Z","shell.execute_reply":"2026-08-17T16:52:50.157161Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pydicom\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\nstudy_id = \"1.2.826.0.1.3680043.8.498.10004873229099053869093324292195817260\"\nseries_id = \"1.2.826.0.1.3680043.8.498.12343110195036213483454091715412333772\"\n\nseries_path = root / \"train_series\" / study_id / series_id\n\nrecords = []\n\nfor f in series_path.glob(\"*.dcm\"):\n ds = pydicom.dcmread(f)\n\n position = getattr(ds, \"ImagePositionPatient\", None)\n orientation = getattr(ds, \"ImageOrientationPatient\", None)\n instance = getattr(ds, \"InstanceNumber\", None)\n\n records.append({\n \"file\": f,\n \"instance\": instance,\n \"position\": position,\n \"orientation\": orientation,\n })\n\nprint(\"Number of slices:\", len(records))\n\nprint(\"\\nFirst 10 metadata records:\")\nfor r in records[:10]:\n print(\n \"Instance:\", r[\"instance\"],\n \"| Position:\", r[\"position\"],\n \"| File:\", r[\"file\"].name\n )\n\n# Sort using ImagePositionPatient when available.\nif all(r[\"position\"] is not None for r in records):\n records = sorted(\n records,\n key=lambda r: tuple(float(x) for x in r[\"position\"])\n )\n sorting_method = \"ImagePositionPatient\"\nelse:\n records = sorted(\n records,\n key=lambda r: (\n r[\"instance\"] if r[\"instance\"] is not None else 0\n )\n )\n sorting_method = \"InstanceNumber\"\n\nprint(\"\\nSorting method:\", sorting_method)\n\nprint(\"\\nOrdered slices:\")\nfor i, r in enumerate(records):\n print(\n i,\n \"| Instance:\", r[\"instance\"],\n \"| Position:\", r[\"position\"]\n )\n\n# Visualize ordered slices\nimages = []\n\nfor r in records:\n ds = pydicom.dcmread(r[\"file\"])\n images.append(ds.pixel_array.astype(np.float32))\n\nindices = np.linspace(\n 0,\n len(images) - 1,\n min(6, len(images)),\n dtype=int\n)\n\nfig, axes = plt.subplots(2, 3, figsize=(12, 8))\n\nfor ax, idx in zip(axes.ravel(), indices):\n ax.imshow(images[idx], cmap=\"gray\")\n ax.set_title(f\"Ordered slice {idx + 1}/{len(images)}\")\n ax.axis(\"off\")\n\nplt.tight_layout()\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:55:20.740211Z","iopub.execute_input":"2026-08-17T16:55:20.741468Z","iopub.status.idle":"2026-08-17T16:55:21.859428Z","shell.execute_reply.started":"2026-08-17T16:55:20.741432Z","shell.execute_reply":"2026-08-17T16:55:21.858231Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pydicom\nimport numpy as np\nfrom pathlib import Path\n\n\ndef load_dicom_series(series_path):\n \"\"\"\n Load a DICOM MRI series and order slices using\n ImageOrientationPatient + ImagePositionPatient.\n \"\"\"\n\n series_path = Path(series_path)\n\n dicom_files = list(series_path.glob(\"*.dcm\"))\n\n if not dicom_files:\n raise FileNotFoundError(f\"No DICOM files found in: {series_path}\")\n\n records = []\n\n for file_path in dicom_files:\n ds = pydicom.dcmread(file_path)\n\n if not hasattr(ds, \"ImagePositionPatient\"):\n raise ValueError(\n f\"ImagePositionPatient missing: {file_path.name}\"\n )\n\n if not hasattr(ds, \"ImageOrientationPatient\"):\n raise ValueError(\n f\"ImageOrientationPatient missing: {file_path.name}\"\n )\n\n position = np.asarray(\n ds.ImagePositionPatient,\n dtype=np.float64\n )\n\n orientation = np.asarray(\n ds.ImageOrientationPatient,\n dtype=np.float64\n )\n\n row_direction = orientation[:3]\n column_direction = orientation[3:]\n\n slice_normal = np.cross(\n row_direction,\n column_direction\n )\n\n slice_position = np.dot(\n position,\n slice_normal\n )\n\n records.append({\n \"file\": file_path,\n \"dataset\": ds,\n \"position\": position,\n \"slice_position\": slice_position,\n \"instance\": getattr(ds, \"InstanceNumber\", None),\n })\n\n # Sort by physical slice position\n records.sort(key=lambda x: x[\"slice_position\"])\n\n images = []\n\n for record in records:\n image = record[\"dataset\"].pixel_array.astype(np.float32)\n images.append(image)\n\n volume = np.stack(images, axis=0)\n\n metadata = {\n \"num_slices\": len(records),\n \"height\": volume.shape[1],\n \"width\": volume.shape[2],\n \"modality\": getattr(records[0][\"dataset\"], \"Modality\", None),\n \"series_description\": getattr(\n records[0][\"dataset\"],\n \"SeriesDescription\",\n None\n ),\n \"pixel_spacing\": getattr(\n records[0][\"dataset\"],\n \"PixelSpacing\",\n None\n ),\n \"slice_thickness\": getattr(\n records[0][\"dataset\"],\n \"SliceThickness\",\n None\n ),\n \"slice_positions\": [\n r[\"slice_position\"] for r in records\n ],\n \"instance_numbers\": [\n r[\"instance\"] for r in records\n ],\n }\n\n return volume, metadata","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:57:54.212002Z","iopub.execute_input":"2026-08-17T16:57:54.212895Z","iopub.status.idle":"2026-08-17T16:57:54.224823Z","shell.execute_reply.started":"2026-08-17T16:57:54.212856Z","shell.execute_reply":"2026-08-17T16:57:54.223898Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"root = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\nstudy_id = (\n \"1.2.826.0.1.3680043.8.498.\"\n \"10004873229099053869093324292195817260\"\n)\n\nseries_id = (\n \"1.2.826.0.1.3680043.8.498.\"\n \"12343110195036213483454091715412333772\"\n)\n\nseries_path = (\n root\n / \"train_series\"\n / study_id\n / series_id\n)\n\nvolume, metadata = load_dicom_series(series_path)\n\nprint(\"Volume shape:\", volume.shape)\n\nprint(\"\\nMetadata:\")\nfor key, value in metadata.items():\n if key not in [\"slice_positions\"]:\n print(f\"{key}: {value}\")\n\nprint(\"\\nFirst 5 physical positions:\")\nprint(metadata[\"slice_positions\"][:5])\n\nprint(\"\\nLast 5 physical positions:\")\nprint(metadata[\"slice_positions\"][-5:])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T16:58:11.315199Z","iopub.execute_input":"2026-08-17T16:58:11.315505Z","iopub.status.idle":"2026-08-17T16:58:11.44761Z","shell.execute_reply.started":"2026-08-17T16:58:11.31548Z","shell.execute_reply":"2026-08-17T16:58:11.446789Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\ntrain_series = pd.read_csv(root / \"train_series.csv\")\ntrain = pd.read_csv(root / \"train.csv\")\n\nprint(\"Total series rows:\", len(train_series))\nprint(\"Unique studies:\", train_series[\"StudyInstanceUID\"].nunique())\nprint(\"Unique series:\", train_series[\"SeriesInstanceUID\"].nunique())\n\nprint(\"\\n\" + \"=\"*60)\nprint(\"ANATOMICAL PLANE\")\nprint(\"=\"*60)\ndisplay(\n train_series[\"Anatomical_Plane\"]\n .value_counts(dropna=False)\n .to_frame(\"count\")\n)\n\nprint(\"\\n\" + \"=\"*60)\nprint(\"FLUID SENSITIVE\")\nprint(\"=\"*60)\ndisplay(\n train_series[\"Fluid_Sensitive\"]\n .value_counts(dropna=False)\n .to_frame(\"count\")\n)\n\nprint(\"\\n\" + \"=\"*60)\nprint(\"FAT SUPPRESSION\")\nprint(\"=\"*60)\ndisplay(\n train_series[\"Fat_Suppression\"]\n .value_counts(dropna=False)\n .to_frame(\"count\")\n)\n\nprint(\"\\n\" + \"=\"*60)\nprint(\"PLANE × FLUID SENSITIVE × FAT SUPPRESSION\")\nprint(\"=\"*60)\n\ndisplay(\n train_series\n .groupby(\n [\"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\"],\n dropna=False\n )\n .size()\n .reset_index(name=\"number_of_series\")\n .sort_values(\"number_of_series\", ascending=False)\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:03:15.506826Z","iopub.execute_input":"2026-08-17T17:03:15.507129Z","iopub.status.idle":"2026-08-17T17:03:15.730904Z","shell.execute_reply.started":"2026-08-17T17:03:15.507103Z","shell.execute_reply":"2026-08-17T17:03:15.730075Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# How many series of each type does a typical study contain?\n\nstudy_series_summary = (\n train_series\n .groupby([\"StudyInstanceUID\", \"Anatomical_Plane\"])\n .size()\n .unstack(fill_value=0)\n)\n\nprint(\"Series counts per study by anatomical plane:\")\ndisplay(study_series_summary.describe())\n\nprint(\"\\nFirst 20 studies:\")\ndisplay(study_series_summary.head(20))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:03:39.056513Z","iopub.execute_input":"2026-08-17T17:03:39.056856Z","iopub.status.idle":"2026-08-17T17:03:39.109732Z","shell.execute_reply.started":"2026-08-17T17:03:39.056828Z","shell.execute_reply":"2026-08-17T17:03:39.108761Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Create one row per study showing which series types are present\n\nseries_types = (\n train_series\n .groupby(\n [\n \"StudyInstanceUID\",\n \"Anatomical_Plane\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\"\n ]\n )\n .size()\n .reset_index(name=\"count\")\n)\n\n# Turn the six sequence combinations into columns\nstudy_matrix = (\n series_types\n .assign(\n sequence_type=lambda x:\n x[\"Anatomical_Plane\"].astype(str)\n + \"_F\"\n + x[\"Fluid_Sensitive\"].astype(str)\n + \"_FS\"\n + x[\"Fat_Suppression\"].astype(str)\n )\n .pivot_table(\n index=\"StudyInstanceUID\",\n columns=\"sequence_type\",\n values=\"count\",\n fill_value=0\n )\n)\n\nprint(\"Number of studies:\", len(study_matrix))\n\nprint(\"\\nHow many sequence types does each study contain?\")\n\nnum_sequence_types = (study_matrix > 0).sum(axis=1)\n\ndisplay(\n num_sequence_types\n .value_counts()\n .sort_index()\n .to_frame(\"number_of_studies\")\n)\n\nprint(\"\\nSequence presence across studies:\")\n\npresence = (study_matrix > 0).mean().sort_values(ascending=False)\n\ndisplay(\n (presence * 100)\n .round(2)\n .to_frame(\"percentage_of_studies\")\n)\n\nprint(\"\\nFirst 20 studies:\")\ndisplay(study_matrix.head(20))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:04:07.72678Z","iopub.execute_input":"2026-08-17T17:04:07.727254Z","iopub.status.idle":"2026-08-17T17:04:07.821536Z","shell.execute_reply.started":"2026-08-17T17:04:07.727224Z","shell.execute_reply":"2026-08-17T17:04:07.820599Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Exact study-level sequence completeness\n\nnum_sequence_types = (study_matrix > 0).sum(axis=1)\n\nprint(\"=\" * 60)\nprint(\"NUMBER OF DIFFERENT SEQUENCE TYPES PER STUDY\")\nprint(\"=\" * 60)\n\ndisplay(\n num_sequence_types\n .value_counts()\n .sort_index()\n .rename(\"number_of_studies\")\n .to_frame()\n)\n\nprint(\"\\n\" + \"=\" * 60)\nprint(\"PERCENTAGE OF STUDIES\")\nprint(\"=\" * 60)\n\ndisplay(\n (\n num_sequence_types\n .value_counts(normalize=True)\n .sort_index()\n .mul(100)\n .round(2)\n .rename(\"percentage\")\n .to_frame()\n )\n)\n\nprint(\"\\n\" + \"=\" * 60)\nprint(\"TOTAL NUMBER OF STUDIES\")\nprint(\"=\" * 60)\n\nprint(len(study_matrix))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:05:21.686671Z","iopub.execute_input":"2026-08-17T17:05:21.687945Z","iopub.status.idle":"2026-08-17T17:05:21.711186Z","shell.execute_reply.started":"2026-08-17T17:05:21.687909Z","shell.execute_reply":"2026-08-17T17:05:21.710339Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"=\" * 60)\nprint(\"PRESENCE OF EACH SEQUENCE TYPE\")\nprint(\"=\" * 60)\n\npresence = (study_matrix > 0).sum(axis=0)\ntotal_studies = len(study_matrix)\n\nsummary = pd.DataFrame({\n \"studies_with_sequence\": presence,\n \"percentage\": (presence / total_studies * 100).round(2)\n})\n\ndisplay(summary.sort_values(\"percentage\", ascending=False))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:05:34.011166Z","iopub.execute_input":"2026-08-17T17:05:34.011576Z","iopub.status.idle":"2026-08-17T17:05:34.025568Z","shell.execute_reply.started":"2026-08-17T17:05:34.011545Z","shell.execute_reply":"2026-08-17T17:05:34.024486Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\nimport numpy as np\n\nroot = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\ntrain_series = pd.read_csv(root / \"train_series.csv\")\n\n# Only the four common sequence types for now\ncommon_types = [\n (\"Axial\", 1, 1),\n (\"Sagittal\", 0, 0),\n (\"Coronal\", 1, 1),\n (\"Sagittal\", 1, 1),\n]\n\nrows = []\n\nfor plane, fluid, fat in common_types:\n \n subset = train_series[\n (train_series[\"Anatomical_Plane\"] == plane) &\n (train_series[\"Fluid_Sensitive\"] == fluid) &\n (train_series[\"Fat_Suppression\"] == fat)\n ]\n\n # Count actual DICOM files for up to 100 series first\n for _, r in subset.head(100).iterrows():\n\n series_path = (\n root\n / \"train_series\"\n / str(r[\"StudyInstanceUID\"])\n / str(r[\"SeriesInstanceUID\"])\n )\n\n n_files = len(list(series_path.glob(\"*.dcm\")))\n\n rows.append({\n \"sequence_type\": f\"{plane}_F{fluid}_FS{fat}\",\n \"StudyInstanceUID\": r[\"StudyInstanceUID\"],\n \"SeriesInstanceUID\": r[\"SeriesInstanceUID\"],\n \"num_slices\": n_files\n })\n\nslice_summary = pd.DataFrame(rows)\n\nprint(\"Sampled series:\", len(slice_summary))\n\ndisplay(\n slice_summary\n .groupby(\"sequence_type\")[\"num_slices\"]\n .describe()\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:06:00.310393Z","iopub.execute_input":"2026-08-17T17:06:00.310748Z","iopub.status.idle":"2026-08-17T17:06:05.84834Z","shell.execute_reply.started":"2026-08-17T17:06:00.310719Z","shell.execute_reply":"2026-08-17T17:06:05.847404Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\ntrain = pd.read_csv(root / \"train.csv\")\ntrain_series = pd.read_csv(root / \"train_series.csv\")\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# These are the studies with known labels\nlabelled = train[train[targets].notna().all(axis=1)].copy()\n\nprint(\"Total studies:\", len(train))\nprint(\"Fully labelled studies:\", len(labelled))\n\n# Sequence availability among ONLY labelled studies\ncommon_types = [\n (\"Axial\", 1, 1),\n (\"Sagittal\", 0, 0),\n (\"Coronal\", 1, 1),\n (\"Sagittal\", 1, 1),\n]\n\nlabelled_series = train_series[\n train_series[\"StudyInstanceUID\"].isin(labelled[\"StudyInstanceUID\"])\n].copy()\n\nprint(\"\\nLabelled-study sequence availability:\")\n\nfor plane, fluid, fat in common_types:\n mask = (\n (labelled_series[\"Anatomical_Plane\"] == plane) &\n (labelled_series[\"Fluid_Sensitive\"] == fluid) &\n (labelled_series[\"Fat_Suppression\"] == fat)\n )\n\n n = labelled_series.loc[mask, \"StudyInstanceUID\"].nunique()\n pct = n / len(labelled) * 100\n\n print(\n f\"{plane}_F{fluid}_FS{fat}: \"\n f\"{n}/{len(labelled)} studies ({pct:.2f}%)\"\n )\n\nprint(\"\\nLabelled studies:\")\ndisplay(labelled[[\"StudyInstanceUID\"] + targets].head())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:07:06.223931Z","iopub.execute_input":"2026-08-17T17:07:06.224621Z","iopub.status.idle":"2026-08-17T17:07:06.415635Z","shell.execute_reply.started":"2026-08-17T17:07:06.22459Z","shell.execute_reply":"2026-08-17T17:07:06.414882Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\"/kaggle/input/competitions/rsna-knee-abnormality-detection\")\n\ntrain = pd.read_csv(root / \"train.csv\")\ntrain_series = pd.read_csv(root / \"train_series.csv\")\n\ntargets = [\n \"ACL\", \"MCL\", \"Medial Meniscus\", \"Lateral Meniscus\",\n \"Medial OA\", \"Lateral OA\", \"PF OA\", \"Effusion\",\n \"Synovitis\", \"Baker's\", \"Contusion\", \"Fracture\"\n]\n\n# Exactly the studies with complete labels\nlabelled = train[train[targets].notna().all(axis=1)].copy()\n\nprint(\"Total studies:\", len(train))\nprint(\"Fully labelled studies:\", len(labelled))\n\nlabelled_series = train_series[\n train_series[\"StudyInstanceUID\"].isin(labelled[\"StudyInstanceUID\"])\n].copy()\n\nsequence_types = [\n (\"Axial\", 1, 1),\n (\"Sagittal\", 0, 0),\n (\"Coronal\", 1, 1),\n (\"Sagittal\", 1, 1),\n (\"Coronal\", 0, 0),\n (\"Axial\", 0, 0),\n]\n\nprint(\"\\nSequence availability among the 58 labelled studies:\\n\")\n\nfor plane, fluid, fat in sequence_types:\n mask = (\n (labelled_series[\"Anatomical_Plane\"] == plane) &\n (labelled_series[\"Fluid_Sensitive\"] == fluid) &\n (labelled_series[\"Fat_Suppression\"] == fat)\n )\n\n n_studies = labelled_series.loc[\n mask, \"StudyInstanceUID\"\n ].nunique()\n\n percentage = 100 * n_studies / len(labelled)\n\n print(\n f\"{plane}_F{fluid}_FS{fat}: \"\n f\"{n_studies}/58 ({percentage:.2f}%)\"\n )\n\nprint(\"\\nLabel distribution in the 58 labelled studies:\")\n\ndistribution = pd.DataFrame({\n \"positive\": labelled[targets].sum(),\n \"negative\": (1 - labelled[targets]).sum()\n})\n\ndistribution[\"positive_rate_%\"] = (\n distribution[\"positive\"] / len(labelled) * 100\n).round(2)\n\ndisplay(distribution)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:07:39.041432Z","iopub.execute_input":"2026-08-17T17:07:39.041896Z","iopub.status.idle":"2026-08-17T17:07:39.234654Z","shell.execute_reply.started":"2026-08-17T17:07:39.041866Z","shell.execute_reply":"2026-08-17T17:07:39.23384Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"\\nSequence availability among the 58 labelled studies:\\n\")\n\nfor plane, fluid, fat in [\n (\"Axial\", 1, 1),\n (\"Sagittal\", 0, 0),\n (\"Coronal\", 1, 1),\n (\"Sagittal\", 1, 1),\n (\"Coronal\", 0, 0),\n (\"Axial\", 0, 0),\n]:\n mask = (\n (labelled_series[\"Anatomical_Plane\"] == plane) &\n (labelled_series[\"Fluid_Sensitive\"] == fluid) &\n (labelled_series[\"Fat_Suppression\"] == fat)\n )\n\n n = labelled_series.loc[\n mask, \"StudyInstanceUID\"\n ].nunique()\n\n print(\n f\"{plane}_F{fluid}_FS{fat}: \"\n f\"{n}/58 ({100*n/58:.2f}%)\"\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:08:44.543968Z","iopub.execute_input":"2026-08-17T17:08:44.544501Z","iopub.status.idle":"2026-08-17T17:08:44.559723Z","shell.execute_reply.started":"2026-08-17T17:08:44.544467Z","shell.execute_reply":"2026-08-17T17:08:44.558535Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"sequence_cols = [\n (\"Axial\", 1, 1, \"Axial_F1_FS1\"),\n (\"Sagittal\", 0, 0, \"Sagittal_F0_FS0\"),\n (\"Coronal\", 1, 1, \"Coronal_F1_FS1\"),\n (\"Sagittal\", 1, 1, \"Sagittal_F1_FS1\"),\n]\n\nprint(\"Duplicate series counts within labelled studies:\\n\")\n\nfor plane, fluid, fat, name in sequence_cols:\n\n subset = labelled_series[\n (labelled_series[\"Anatomical_Plane\"] == plane) &\n (labelled_series[\"Fluid_Sensitive\"] == fluid) &\n (labelled_series[\"Fat_Suppression\"] == fat)\n ]\n\n counts = (\n subset.groupby(\"StudyInstanceUID\")\n .size()\n )\n\n print(f\"\\n{name}\")\n print(\"Studies with sequence:\", counts.size)\n print(\"Studies with >1 series:\", int((counts > 1).sum()))\n print(\"Maximum series in one study:\", int(counts.max()))\n print(\"\\nDistribution:\")\n print(counts.value_counts().sort_index())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:09:30.15378Z","iopub.execute_input":"2026-08-17T17:09:30.154923Z","iopub.status.idle":"2026-08-17T17:09:30.176175Z","shell.execute_reply.started":"2026-08-17T17:09:30.154883Z","shell.execute_reply":"2026-08-17T17:09:30.174942Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\ntrain = pd.read_csv(root / \"train.csv\")\ntrain_series = pd.read_csv(root / \"train_series.csv\")\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# 1. Keep only completely labelled studies\n# ---------------------------------------------------------\n\nlabelled = train[\n train[targets].notna().all(axis=1)\n].copy()\n\nprint(\"Labelled studies:\", len(labelled))\n\n# ---------------------------------------------------------\n# 2. Define the four sequence types for Baseline V1\n# ---------------------------------------------------------\n\nsequence_types = [\n (\"Axial\", 1, 1, \"axial_f1_fs1\"),\n (\"Sagittal\", 0, 0, \"sagittal_f0_fs0\"),\n (\"Coronal\", 1, 1, \"coronal_f1_fs1\"),\n (\"Sagittal\", 1, 1, \"sagittal_f1_fs1\"),\n]\n\n# ---------------------------------------------------------\n# 3. Find candidate series for every study/sequence type\n# ---------------------------------------------------------\n\nrows = []\n\nfor plane, fluid, fat, name in sequence_types:\n\n subset = train_series[\n (train_series[\"Anatomical_Plane\"] == plane) &\n (train_series[\"Fluid_Sensitive\"] == fluid) &\n (train_series[\"Fat_Suppression\"] == fat)\n ].copy()\n\n for study_id, group in subset.groupby(\"StudyInstanceUID\"):\n\n # Only studies in our labelled set\n if study_id not in set(labelled[\"StudyInstanceUID\"]):\n continue\n\n candidates = []\n\n for _, series_row in group.iterrows():\n\n series_path = (\n root\n / \"train_series\"\n / str(study_id)\n / str(series_row[\"SeriesInstanceUID\"])\n )\n\n n_slices = len(\n list(series_path.glob(\"*.dcm\"))\n )\n\n candidates.append({\n \"SeriesInstanceUID\":\n series_row[\"SeriesInstanceUID\"],\n \"num_slices\":\n n_slices\n })\n\n # Pick longest series\n candidates = sorted(\n candidates,\n key=lambda x: x[\"num_slices\"],\n reverse=True\n )\n\n if candidates:\n selected = candidates[0]\n\n rows.append({\n \"StudyInstanceUID\": study_id,\n \"sequence_type\": name,\n \"SeriesInstanceUID\":\n selected[\"SeriesInstanceUID\"],\n \"num_slices\":\n selected[\"num_slices\"],\n \"num_candidate_series\":\n len(candidates)\n })\n\nselected_series = pd.DataFrame(rows)\n\nprint(\"\\nSelected series:\")\ndisplay(selected_series.head(20))\n\nprint(\"\\nRows:\", len(selected_series))\n\nprint(\"\\nSelected series by sequence type:\")\ndisplay(\n selected_series[\"sequence_type\"]\n .value_counts()\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:10:20.536712Z","iopub.execute_input":"2026-08-17T17:10:20.537628Z","iopub.status.idle":"2026-08-17T17:10:24.207425Z","shell.execute_reply.started":"2026-08-17T17:10:20.537595Z","shell.execute_reply":"2026-08-17T17:10:24.206526Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ---------------------------------------------------------\n# 4. Pivot into one row per study\n# ---------------------------------------------------------\n\nstudy_metadata = labelled[\n [\"StudyInstanceUID\", \"Report\"] + targets\n].copy()\n\nseries_pivot = selected_series.pivot(\n index=\"StudyInstanceUID\",\n columns=\"sequence_type\",\n values=\"SeriesInstanceUID\"\n).reset_index()\n\nstudy_metadata = study_metadata.merge(\n series_pivot,\n on=\"StudyInstanceUID\",\n how=\"left\"\n)\n\nprint(\"Final metadata table shape:\", study_metadata.shape)\n\ndisplay(study_metadata.head())\n\nprint(\"\\nMissing selected series:\")\ndisplay(\n study_metadata[\n [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\"\n ]\n ].isna().sum()\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:10:35.446993Z","iopub.execute_input":"2026-08-17T17:10:35.447692Z","iopub.status.idle":"2026-08-17T17:10:35.484066Z","shell.execute_reply.started":"2026-08-17T17:10:35.447657Z","shell.execute_reply":"2026-08-17T17:10:35.483114Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"output_path = Path(\"/kaggle/working/study_metadata_v1.csv\")\n\nstudy_metadata.to_csv(\n output_path,\n index=False\n)\n\nprint(\"Saved:\", output_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:11:16.909264Z","iopub.execute_input":"2026-08-17T17:11:16.910449Z","iopub.status.idle":"2026-08-17T17:11:16.922311Z","shell.execute_reply.started":"2026-08-17T17:11:16.910406Z","shell.execute_reply":"2026-08-17T17:11:16.921367Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nroot = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\nmetadata_path = Path(\"/kaggle/working/study_metadata_v1.csv\")\n\nstudy_metadata = pd.read_csv(metadata_path)\n\nprint(\"Shape:\", study_metadata.shape)\nprint(\"\\nColumns:\")\nprint(study_metadata.columns.tolist())\n\nprint(\"\\nMissing values:\")\ndisplay(\n study_metadata.isna().sum().to_frame(\"missing\")\n)\n\nprint(\"\\nSelected series columns:\")\ndisplay(\n study_metadata[\n [\n \"StudyInstanceUID\",\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\"\n ]\n ].head(10)\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:11:42.677048Z","iopub.execute_input":"2026-08-17T17:11:42.677797Z","iopub.status.idle":"2026-08-17T17:11:42.703275Z","shell.execute_reply.started":"2026-08-17T17:11:42.677764Z","shell.execute_reply":"2026-08-17T17:11:42.702358Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\nimport pandas as pd\nimport pydicom\n\nroot = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\nstudy_metadata = pd.read_csv(\n \"/kaggle/working/study_metadata_v1.csv\"\n)\n\nsequence_columns = [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\",\n]\n\nresults = []\n\nfor _, row in study_metadata.iterrows():\n\n study_id = row[\"StudyInstanceUID\"]\n\n for seq in sequence_columns:\n\n series_id = row[seq]\n\n if pd.isna(series_id):\n results.append({\n \"StudyInstanceUID\": study_id,\n \"sequence\": seq,\n \"exists\": False,\n \"num_dicom\": 0,\n \"readable\": False,\n })\n continue\n\n series_path = (\n root\n / \"train_series\"\n / str(study_id)\n / str(series_id)\n )\n\n dcm_files = list(series_path.glob(\"*.dcm\"))\n\n readable = False\n\n if dcm_files:\n try:\n ds = pydicom.dcmread(\n dcm_files[0],\n stop_before_pixels=True\n )\n readable = True\n except Exception:\n readable = False\n\n results.append({\n \"StudyInstanceUID\": study_id,\n \"sequence\": seq,\n \"exists\": series_path.exists(),\n \"num_dicom\": len(dcm_files),\n \"readable\": readable,\n })\n\nintegrity = pd.DataFrame(results)\n\nprint(\"Total checks:\", len(integrity))\n\nprint(\"\\nDirectory existence:\")\ndisplay(\n integrity[\"exists\"].value_counts(dropna=False)\n)\n\nprint(\"\\nReadable DICOM:\")\ndisplay(\n integrity[\"readable\"].value_counts(dropna=False)\n)\n\nprint(\"\\nDICOM slice-count summary:\")\ndisplay(\n integrity\n .groupby(\"sequence\")[\"num_dicom\"]\n .describe()\n)\n\nprint(\"\\nAny problematic entries:\")\ndisplay(\n integrity[\n (~integrity[\"exists\"]) |\n (~integrity[\"readable\"]) |\n (integrity[\"num_dicom\"] == 0)\n ]\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:12:23.264032Z","iopub.execute_input":"2026-08-17T17:12:23.264517Z","iopub.status.idle":"2026-08-17T17:12:26.004128Z","shell.execute_reply.started":"2026-08-17T17:12:23.264487Z","shell.execute_reply":"2026-08-17T17:12:26.003138Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Pick the first problematic entry automatically\n\nbad = integrity[\n (~integrity[\"exists\"]) |\n (~integrity[\"readable\"]) |\n (integrity[\"num_dicom\"] == 0)\n].iloc[0]\n\nprint(\"Problematic study:\")\nprint(\"StudyInstanceUID:\", bad[\"StudyInstanceUID\"])\nprint(\"Sequence:\", bad[\"sequence\"])\n\nstudy_id = bad[\"StudyInstanceUID\"]\n\nseries_id = study_metadata.loc[\n study_metadata[\"StudyInstanceUID\"] == study_id,\n bad[\"sequence\"]\n].iloc[0]\n\nprint(\"\\nSelected SeriesInstanceUID:\")\nprint(series_id)\n\nstudy_path = root / \"train_series\" / str(study_id)\n\nprint(\"\\nStudy directory exists:\", study_path.exists())\nprint(\"Study directory:\", study_path)\n\nif study_path.exists():\n print(\"\\nActual contents:\")\n for item in sorted(study_path.iterdir()):\n print(item)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:14:38.771239Z","iopub.execute_input":"2026-08-17T17:14:38.772155Z","iopub.status.idle":"2026-08-17T17:14:38.790041Z","shell.execute_reply.started":"2026-08-17T17:14:38.77201Z","shell.execute_reply":"2026-08-17T17:14:38.788996Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"study_id = \"1.2.826.0.1.3680043.8.498.25695966186129395049687747156570466645\"\n\nrow = study_metadata[\n study_metadata[\"StudyInstanceUID\"] == study_id\n].iloc[0]\n\nsequence_columns = [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\",\n]\n\nprint(\"Selected series for this study:\\n\")\n\nfor col in sequence_columns:\n print(f\"{col}: {row[col]}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:15:19.422214Z","iopub.execute_input":"2026-08-17T17:15:19.422692Z","iopub.status.idle":"2026-08-17T17:15:19.430368Z","shell.execute_reply.started":"2026-08-17T17:15:19.422664Z","shell.execute_reply":"2026-08-17T17:15:19.429328Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"\\nActual series from train_series.csv:\\n\")\n\nactual = train_series[\n train_series[\"StudyInstanceUID\"] == study_id\n]\n\ndisplay(\n actual[\n [\n \"SeriesInstanceUID\",\n \"Fluid_Sensitive\",\n \"Fat_Suppression\",\n \"Anatomical_Plane\"\n ]\n ]\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:15:29.825986Z","iopub.execute_input":"2026-08-17T17:15:29.826405Z","iopub.status.idle":"2026-08-17T17:15:29.841549Z","shell.execute_reply.started":"2026-08-17T17:15:29.826377Z","shell.execute_reply":"2026-08-17T17:15:29.840534Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nfrom pathlib import Path\n\nmetadata_path = Path(\"/kaggle/working/study_metadata_v1.csv\")\n\nstudy_metadata = pd.read_csv(metadata_path)\n\nsequence_columns = [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\",\n]\n\nfor col in sequence_columns:\n study_metadata[col + \"_available\"] = (\n study_metadata[col].notna().astype(int)\n )\n\navailability_columns = [\n col + \"_available\"\n for col in sequence_columns\n]\n\nprint(\"Sequence availability:\")\ndisplay(\n study_metadata[availability_columns]\n .value_counts()\n .reset_index(name=\"number_of_studies\")\n)\n\nprint(\"\\nMissing sequence count:\")\nfor col in sequence_columns:\n print(\n f\"{col}: \"\n f\"{study_metadata[col].isna().sum()} missing\"\n )\n\nnew_path = Path(\n \"/kaggle/working/study_metadata_v2.csv\"\n)\n\nstudy_metadata.to_csv(\n new_path,\n index=False\n)\n\nprint(\"\\nSaved:\", new_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:16:07.656891Z","iopub.execute_input":"2026-08-17T17:16:07.657312Z","iopub.status.idle":"2026-08-17T17:16:07.690749Z","shell.execute_reply.started":"2026-08-17T17:16:07.65728Z","shell.execute_reply":"2026-08-17T17:16:07.689972Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport pydicom\nimport torch\n\nfrom pathlib import Path\nfrom torch.utils.data import Dataset\nimport torch.nn.functional as F\n\n\nROOT = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\nMETADATA_PATH = Path(\n \"/kaggle/working/study_metadata_v2.csv\"\n)\n\nmetadata = pd.read_csv(METADATA_PATH)\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\nSEQUENCES = [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\",\n]\n\n\ndef load_dicom_volume(series_path):\n \"\"\"\n Load one DICOM series.\n\n Slice ordering:\n ImageOrientationPatient\n +\n ImagePositionPatient\n\n Returns:\n volume: numpy array [N, H, W]\n \"\"\"\n\n series_path = Path(series_path)\n\n files = list(series_path.glob(\"*.dcm\"))\n\n if not files:\n raise FileNotFoundError(\n f\"No DICOM files found: {series_path}\"\n )\n\n records = []\n\n for file_path in files:\n\n ds = pydicom.dcmread(file_path)\n\n if not hasattr(ds, \"ImagePositionPatient\"):\n raise ValueError(\n f\"Missing ImagePositionPatient: {file_path}\"\n )\n\n if not hasattr(ds, \"ImageOrientationPatient\"):\n raise ValueError(\n f\"Missing ImageOrientationPatient: {file_path}\"\n )\n\n position = np.asarray(\n ds.ImagePositionPatient,\n dtype=np.float64\n )\n\n orientation = np.asarray(\n ds.ImageOrientationPatient,\n dtype=np.float64\n )\n\n row_direction = orientation[:3]\n column_direction = orientation[3:]\n\n normal = np.cross(\n row_direction,\n column_direction\n )\n\n physical_position = np.dot(\n position,\n normal\n )\n\n records.append(\n (\n physical_position,\n file_path\n )\n )\n\n records.sort(key=lambda x: x[0])\n\n images = []\n\n for _, file_path in records:\n\n ds = pydicom.dcmread(file_path)\n\n image = ds.pixel_array.astype(\n np.float32\n )\n\n images.append(image)\n\n volume = np.stack(\n images,\n axis=0\n )\n\n return volume\n\n\ndef sample_slices(volume, num_slices=16):\n \"\"\"\n Uniformly sample a fixed number of slices.\n \"\"\"\n\n n = volume.shape[0]\n\n if n == num_slices:\n return volume\n\n indices = np.linspace(\n 0,\n n - 1,\n num_slices\n ).round().astype(int)\n\n return volume[indices]\n\n\ndef normalize_slice(image):\n \"\"\"\n Robust per-slice percentile normalization.\n\n Output approximately [0, 1].\n \"\"\"\n\n image = image.astype(\n np.float32\n )\n\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 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\ndef preprocess_volume(\n volume,\n num_slices=16,\n image_size=224\n):\n \"\"\"\n Convert [N,H,W] volume into\n [N,1,image_size,image_size].\n \"\"\"\n\n volume = sample_slices(\n volume,\n num_slices=num_slices\n )\n\n processed = []\n\n for image in volume:\n\n image = normalize_slice(\n image\n )\n\n tensor = torch.from_numpy(\n image\n ).unsqueeze(0).unsqueeze(0)\n\n tensor = F.interpolate(\n tensor,\n size=(image_size, image_size),\n mode=\"bilinear\",\n align_corners=False\n )\n\n processed.append(\n tensor.squeeze(0)\n )\n\n return torch.stack(\n processed,\n dim=0\n )\n\n\nclass KneeStudyDataset(Dataset):\n\n def __init__(\n self,\n dataframe,\n root,\n num_slices=16,\n image_size=224,\n ):\n\n self.df = dataframe.reset_index(\n drop=True\n )\n\n self.root = Path(root)\n\n self.num_slices = num_slices\n self.image_size = image_size\n\n def __len__(self):\n return len(self.df)\n\n def __getitem__(self, idx):\n\n row = self.df.iloc[idx]\n\n study_id = row[\n \"StudyInstanceUID\"\n ]\n\n sequence_tensors = []\n availability = []\n\n for sequence in SEQUENCES:\n\n series_id = row[\n sequence\n ]\n\n if pd.isna(series_id):\n\n # Missing sequence\n tensor = torch.zeros(\n self.num_slices,\n 1,\n self.image_size,\n self.image_size,\n dtype=torch.float32\n )\n\n availability.append(0)\n\n else:\n\n series_path = (\n self.root\n / \"train_series\"\n / str(study_id)\n / str(series_id)\n )\n\n volume = load_dicom_volume(\n series_path\n )\n\n tensor = preprocess_volume(\n volume,\n num_slices=self.num_slices,\n image_size=self.image_size\n )\n\n availability.append(1)\n\n sequence_tensors.append(\n tensor\n )\n\n labels = torch.tensor(\n row[TARGETS].values.astype(\n np.float32\n ),\n dtype=torch.float32\n )\n\n return {\n \"study_id\": study_id,\n\n \"images\": torch.stack(\n sequence_tensors,\n dim=0\n ),\n\n \"availability\": torch.tensor(\n availability,\n dtype=torch.float32\n ),\n\n \"labels\": labels,\n }","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:16:49.319966Z","iopub.execute_input":"2026-08-17T17:16:49.32031Z","iopub.status.idle":"2026-08-17T17:16:52.659303Z","shell.execute_reply.started":"2026-08-17T17:16:49.320283Z","shell.execute_reply":"2026-08-17T17:16:52.658261Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"dataset = KneeStudyDataset(\n dataframe=metadata,\n root=ROOT,\n num_slices=16,\n image_size=224\n)\n\nsample = dataset[0]\n\nprint(\"Study ID:\")\nprint(sample[\"study_id\"])\n\nprint(\"\\nImages shape:\")\nprint(sample[\"images\"].shape)\n\nprint(\"\\nAvailability:\")\nprint(sample[\"availability\"])\n\nprint(\"\\nLabels:\")\nprint(sample[\"labels\"])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:17:08.853417Z","iopub.execute_input":"2026-08-17T17:17:08.853774Z","iopub.status.idle":"2026-08-17T17:17:11.049668Z","shell.execute_reply.started":"2026-08-17T17:17:08.853747Z","shell.execute_reply":"2026-08-17T17:17:11.048446Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import matplotlib.pyplot as plt\nimport numpy as np\n\nsample = dataset[0]\n\nimages = sample[\"images\"]\navailability = sample[\"availability\"]\n\nsequence_names = [\n \"Axial F1/FS1\",\n \"Sagittal F0/FS0\",\n \"Coronal F1/FS1\",\n \"Sagittal F1/FS1\",\n]\n\n# Pick representative slices\nslice_indices = [2, 6, 10, 13]\n\nfig, axes = plt.subplots(\n 4,\n 4,\n figsize=(12, 12)\n)\n\nfor row_idx, sequence_name in enumerate(sequence_names):\n\n for col_idx, slice_idx in enumerate(slice_indices):\n\n ax = axes[row_idx, col_idx]\n\n if availability[row_idx] == 0:\n ax.text(\n 0.5,\n 0.5,\n \"MISSING\",\n ha=\"center\",\n va=\"center\",\n fontsize=14\n )\n ax.axis(\"off\")\n continue\n\n image = (\n images[row_idx, slice_idx, 0]\n .numpy()\n )\n\n ax.imshow(\n image,\n cmap=\"gray\"\n )\n\n ax.set_title(\n f\"{sequence_name}\\nSlice {slice_idx + 1}\"\n )\n\n ax.axis(\"off\")\n\nplt.tight_layout()\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:17:50.12865Z","iopub.execute_input":"2026-08-17T17:17:50.129181Z","iopub.status.idle":"2026-08-17T17:17:52.115524Z","shell.execute_reply.started":"2026-08-17T17:17:50.129146Z","shell.execute_reply":"2026-08-17T17:17:52.112966Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Full dataset preprocessing validation\n\nprint(\"Testing all 58 labelled studies...\\n\")\n\nfailed = []\n\nfor idx in range(len(dataset)):\n\n try:\n sample = dataset[idx]\n\n images = sample[\"images\"]\n availability = sample[\"availability\"]\n labels = sample[\"labels\"]\n\n # Expected tensor structure\n assert images.shape == (\n 4, 16, 1, 224, 224\n )\n\n assert availability.shape == (4,)\n assert labels.shape == (12,)\n\n # Check numerical validity\n assert torch.isfinite(images).all()\n assert torch.isfinite(labels).all()\n\n except Exception as e:\n\n failed.append({\n \"index\": idx,\n \"study_id\": metadata.iloc[idx][\"StudyInstanceUID\"],\n \"error\": str(e)\n })\n\n if (idx + 1) % 10 == 0:\n print(f\"Processed {idx + 1}/{len(dataset)}\")\n\nprint(\"\\n\" + \"=\" * 60)\nprint(\"VALIDATION RESULT\")\nprint(\"=\" * 60)\n\nprint(\"Total studies:\", len(dataset))\nprint(\"Successful:\", len(dataset) - len(failed))\nprint(\"Failed:\", len(failed))\n\nif failed:\n print(\"\\nFailed studies:\")\n display(pd.DataFrame(failed))\nelse:\n print(\"\\nALL STUDIES PASSED.\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:19:10.572723Z","iopub.execute_input":"2026-08-17T17:19:10.573271Z","iopub.status.idle":"2026-08-17T17:21:22.758676Z","shell.execute_reply.started":"2026-08-17T17:19:10.573236Z","shell.execute_reply":"2026-08-17T17:21:22.757493Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import numpy as np\nimport pandas as pd\n\n# ---------------------------------------------------------\n# Multilabel-aware greedy 5-fold split\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\ndf = metadata.copy().reset_index(drop=True)\n\nY = df[targets].values.astype(int)\n\nN_FOLDS = 5\nRANDOM_STATE = 42\n\nrng = np.random.default_rng(RANDOM_STATE)\n\nn_samples = len(df)\n\n# Random starting order, then prioritize rare labels\nsample_order = np.arange(n_samples)\nrng.shuffle(sample_order)\n\nlabel_frequency = Y.sum(axis=0)\n\n# Rare-label priority\nsample_priority = np.zeros(n_samples)\n\nfor i in range(n_samples):\n positive_labels = np.where(Y[i] == 1)[0]\n\n if len(positive_labels) > 0:\n sample_priority[i] = sum(\n 1.0 / max(label_frequency[j], 1)\n for j in positive_labels\n )\n\n# Sort from rarest / most informative samples first\nsample_order = sorted(\n sample_order,\n key=lambda i: sample_priority[i],\n reverse=True\n)\n\nfold_indices = [[] for _ in range(N_FOLDS)]\nfold_label_counts = 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# Assign each sample to the fold currently most in need\nfor idx in sample_order:\n\n labels = Y[idx]\n\n best_fold = None\n best_score = None\n\n for fold in range(N_FOLDS):\n\n # Prefer folds with fewer samples\n size_penalty = fold_sizes[fold]\n\n # Prefer folds with fewer examples of this sample's labels\n label_penalty = np.sum(\n labels * fold_label_counts[fold]\n )\n\n score = (\n label_penalty * 10\n + size_penalty\n )\n\n if best_score is None or score < best_score:\n best_score = score\n best_fold = fold\n\n fold_indices[best_fold].append(idx)\n\n fold_label_counts[best_fold] += labels\n fold_sizes[best_fold] += 1\n\n\n# Add fold assignments to metadata\nfold_assignment = np.full(\n n_samples,\n -1,\n dtype=int\n)\n\nfor fold, indices in enumerate(fold_indices):\n fold_assignment[indices] = fold\n\ndf[\"fold\"] = fold_assignment\n\nprint(\"Fold sizes:\")\nprint(\n df[\"fold\"]\n .value_counts()\n .sort_index()\n)\n\nprint(\"\\nLabel distribution by fold:\")\n\nfor fold in range(N_FOLDS):\n\n subset = df[df[\"fold\"] == fold]\n\n distribution = pd.DataFrame({\n \"positive\": subset[targets].sum(),\n \"total\": len(subset),\n \"positive_rate_%\":\n (subset[targets].mean() * 100).round(1)\n })\n\n print(f\"\\n{'='*50}\")\n print(f\"FOLD {fold}\")\n print(f\"{'='*50}\")\n\n display(distribution)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:22:14.916982Z","iopub.execute_input":"2026-08-17T17:22:14.917863Z","iopub.status.idle":"2026-08-17T17:22:14.987978Z","shell.execute_reply.started":"2026-08-17T17:22:14.917831Z","shell.execute_reply":"2026-08-17T17:22:14.98725Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Compact 5-fold validation summary\n\nfold_summary = []\n\nfor fold in range(N_FOLDS):\n subset = df[df[\"fold\"] == fold]\n\n row = {\n \"fold\": fold,\n \"n_studies\": len(subset)\n }\n\n for target in targets:\n row[target] = int(subset[target].sum())\n\n fold_summary.append(row)\n\nfold_summary = pd.DataFrame(fold_summary)\n\nprint(\"Positive cases in each validation fold:\")\ndisplay(fold_summary)\n\nprint(\"\\nMinimum positives available for each target across folds:\")\n\nmin_positive = fold_summary[targets].min().sort_values()\n\ndisplay(\n min_positive.to_frame(\"minimum_positive_cases_in_any_fold\")\n)\n\nprint(\"\\nFold sizes:\")\ndisplay(\n fold_summary[[\"fold\", \"n_studies\"]]\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:22:46.767705Z","iopub.execute_input":"2026-08-17T17:22:46.768037Z","iopub.status.idle":"2026-08-17T17:22:46.802938Z","shell.execute_reply.started":"2026-08-17T17:22:46.768005Z","shell.execute_reply":"2026-08-17T17:22:46.802105Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"Positive cases in each fold:\")\n\nfor fold in range(5):\n subset = df[df[\"fold\"] == fold]\n\n print(f\"\\nFOLD {fold} — {len(subset)} studies\")\n\n for target in targets:\n print(\n f\"{target:20s}: \"\n f\"{int(subset[target].sum())}\"\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:23:10.599779Z","iopub.execute_input":"2026-08-17T17:23:10.600562Z","iopub.status.idle":"2026-08-17T17:23:10.614709Z","shell.execute_reply.started":"2026-08-17T17:23:10.60053Z","shell.execute_reply":"2026-08-17T17:23:10.613724Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Freeze Baseline V1 folds\n\nfold_path = \"/kaggle/working/study_metadata_v2_folds.csv\"\n\ndf.to_csv(\n fold_path,\n index=False\n)\n\nprint(\"Saved frozen CV split:\")\nprint(fold_path)\n\nprint(\"\\nFold counts:\")\nprint(\n df[\"fold\"]\n .value_counts()\n .sort_index()\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:23:36.799507Z","iopub.execute_input":"2026-08-17T17:23:36.79999Z","iopub.status.idle":"2026-08-17T17:23:36.81419Z","shell.execute_reply.started":"2026-08-17T17:23:36.799953Z","shell.execute_reply":"2026-08-17T17:23:36.81315Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\nimport torch\nimport pandas as pd\nimport time\n\n# ---------------------------------------------------------\n# Paths\n# ---------------------------------------------------------\n\ncache_root = Path(\"/kaggle/working/knee_cache\")\ncache_root.mkdir(parents=True, exist_ok=True)\n\nfolded_metadata = pd.read_csv(\n \"/kaggle/working/study_metadata_v2_folds.csv\"\n)\n\nprint(\"Studies to cache:\", len(folded_metadata))\nprint(\"Cache directory:\", cache_root)\n\n\n# ---------------------------------------------------------\n# Cache every labelled study\n# ---------------------------------------------------------\n\ncache_results = []\n\nstart_time = time.time()\n\nfor idx in range(len(folded_metadata)):\n\n row = folded_metadata.iloc[idx]\n study_id = row[\"StudyInstanceUID\"]\n\n output_path = cache_root / f\"{study_id}.pt\"\n\n # Skip if already cached\n if output_path.exists():\n\n cache_results.append({\n \"study_id\": study_id,\n \"cached\": True,\n \"status\": \"already_exists\"\n })\n\n continue\n\n try:\n\n # Use our verified Dataset implementation\n sample = dataset[idx]\n\n torch.save(\n {\n \"study_id\": sample[\"study_id\"],\n \"images\": sample[\"images\"].to(\n torch.float32\n ),\n \"availability\": sample[\"availability\"].to(\n torch.float32\n ),\n \"labels\": sample[\"labels\"].to(\n torch.float32\n ),\n \"fold\": int(row[\"fold\"]),\n },\n output_path\n )\n\n cache_results.append({\n \"study_id\": study_id,\n \"cached\": True,\n \"status\": \"created\"\n })\n\n except Exception as e:\n\n cache_results.append({\n \"study_id\": study_id,\n \"cached\": False,\n \"status\": str(e)\n })\n\n if (idx + 1) % 10 == 0:\n elapsed = time.time() - start_time\n print(\n f\"Processed {idx + 1}/{len(folded_metadata)} \"\n f\"| elapsed: {elapsed:.1f}s\"\n )\n\n\ncache_results = pd.DataFrame(cache_results)\n\nprint(\"\\n\" + \"=\" * 60)\nprint(\"CACHE RESULT\")\nprint(\"=\" * 60)\n\ndisplay(\n cache_results[\"status\"]\n .value_counts()\n .to_frame(\"count\")\n)\n\nfailed = cache_results[\n ~cache_results[\"cached\"]\n]\n\nprint(\"\\nFailed:\", len(failed))\n\nif len(failed):\n display(failed)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:24:05.984908Z","iopub.execute_input":"2026-08-17T17:24:05.985617Z","iopub.status.idle":"2026-08-17T17:25:10.275031Z","shell.execute_reply.started":"2026-08-17T17:24:05.985553Z","shell.execute_reply":"2026-08-17T17:25:10.274181Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Verify cached tensors\n\nfiles = sorted(cache_root.glob(\"*.pt\"))\n\nprint(\"Cached studies:\", len(files))\n\nsample_file = files[0]\n\nsample_cache = torch.load(\n sample_file,\n map_location=\"cpu\"\n)\n\nprint(\"\\nSample cache contents:\")\n\nfor key, value in sample_cache.items():\n\n if torch.is_tensor(value):\n print(\n key,\n \"shape =\", tuple(value.shape),\n \"dtype =\", value.dtype\n )\n else:\n print(\n key,\n \"=\",\n value\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:29:23.876717Z","iopub.execute_input":"2026-08-17T17:29:23.879309Z","iopub.status.idle":"2026-08-17T17:29:24.015874Z","shell.execute_reply.started":"2026-08-17T17:29:23.879233Z","shell.execute_reply":"2026-08-17T17:29:24.014826Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import torch\nimport torchvision\n\nprint(\"PyTorch:\", torch.__version__)\nprint(\"Torchvision:\", torchvision.__version__)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:29:48.838664Z","iopub.execute_input":"2026-08-17T17:29:48.838976Z","iopub.status.idle":"2026-08-17T17:29:53.207404Z","shell.execute_reply.started":"2026-08-17T17:29:48.838952Z","shell.execute_reply":"2026-08-17T17:29:53.206382Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nimport torchvision.models as models\n\n\nclass BaselineV1(nn.Module):\n\n def __init__(\n self,\n num_classes=12,\n num_sequences=4,\n pretrained=True,\n ):\n super().__init__()\n\n weights = (\n models.ResNet18_Weights.DEFAULT\n if pretrained\n else None\n )\n\n backbone = models.resnet18(\n weights=weights\n )\n\n feature_dim = backbone.fc.in_features\n backbone.fc = nn.Identity()\n\n self.encoder = backbone\n self.feature_dim = feature_dim\n self.num_sequences = num_sequences\n\n fusion_dim = (\n num_sequences * feature_dim\n + num_sequences\n )\n\n self.fusion = nn.Sequential(\n nn.Linear(fusion_dim, 512),\n nn.ReLU(),\n nn.Dropout(0.30),\n\n nn.Linear(512, 256),\n nn.ReLU(),\n nn.Dropout(0.20),\n )\n\n self.classifier = nn.Linear(\n 256,\n num_classes\n )\n\n def forward(self, x, availability):\n\n B, S, N, C, H, W = x.shape\n\n x = x.reshape(\n B * S * N,\n C,\n H,\n W\n )\n\n # Grayscale → RGB\n x = x.repeat(1, 3, 1, 1)\n\n features = self.encoder(x)\n\n features = features.reshape(\n B,\n S,\n N,\n self.feature_dim\n )\n\n # Slice pooling\n sequence_features = features.mean(dim=2)\n\n # Explicitly remove missing-sequence contribution\n sequence_features = (\n sequence_features\n * availability.unsqueeze(-1)\n )\n\n sequence_features = sequence_features.reshape(\n B,\n S * self.feature_dim\n )\n\n fused = torch.cat(\n [\n sequence_features,\n availability\n ],\n dim=1\n )\n\n fused = self.fusion(fused)\n\n logits = self.classifier(fused)\n\n return logits","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:31:49.897709Z","iopub.execute_input":"2026-08-17T17:31:49.898118Z","iopub.status.idle":"2026-08-17T17:31:49.909292Z","shell.execute_reply.started":"2026-08-17T17:31:49.898069Z","shell.execute_reply":"2026-08-17T17:31:49.908196Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"device = torch.device(\n \"cuda\" if torch.cuda.is_available()\n else \"cpu\"\n)\n\nprint(\"Device:\", device)\n\nmodel = BaselineV1(\n num_classes=12,\n num_sequences=4,\n pretrained=True\n).to(device)\n\nprint(\"\\nModel created successfully.\")\n\n# Load one cached study\ncache_file = sorted(cache_root.glob(\"*.pt\"))[0]\n\nsample_cache = torch.load(\n cache_file,\n map_location=\"cpu\",\n weights_only=False\n)\n\nimages = sample_cache[\"images\"].unsqueeze(0).to(device)\navailability = sample_cache[\"availability\"].unsqueeze(0).to(device)\n\nprint(\"\\nInput shape:\", images.shape)\nprint(\"Availability shape:\", availability.shape)\n\nwith torch.no_grad():\n logits = model(\n images,\n availability\n )\n\nprint(\"\\nOutput shape:\", logits.shape)\nprint(\"Output:\", logits)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:30:23.934535Z","iopub.execute_input":"2026-08-17T17:30:23.934893Z","iopub.status.idle":"2026-08-17T17:30:29.277421Z","shell.execute_reply.started":"2026-08-17T17:30:23.934854Z","shell.execute_reply":"2026-08-17T17:30:29.276577Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model = BaselineV1(\n num_classes=12,\n num_sequences=4,\n pretrained=True\n).to(device)\n\nwith torch.no_grad():\n logits = model(\n images,\n availability\n )\n\nprint(\"Input:\", images.shape)\nprint(\"Output:\", logits.shape)\nprint(logits)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:32:10.088927Z","iopub.execute_input":"2026-08-17T17:32:10.089395Z","iopub.status.idle":"2026-08-17T17:32:14.350908Z","shell.execute_reply.started":"2026-08-17T17:32:10.089358Z","shell.execute_reply":"2026-08-17T17:32:14.349832Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nfrom torch.utils.data import Dataset, DataLoader\nfrom sklearn.metrics import roc_auc_score\nfrom pathlib import Path\nimport numpy as np\n\n\n# =========================================================\n# 1. Cached dataset\n# =========================================================\n\nclass CachedKneeDataset(Dataset):\n\n def __init__(self, cache_files):\n self.files = list(cache_files)\n\n def __len__(self):\n return len(self.files)\n\n def __getitem__(self, idx):\n\n data = torch.load(\n self.files[idx],\n map_location=\"cpu\",\n weights_only=False\n )\n\n return {\n \"images\": data[\"images\"].float(),\n \"availability\": data[\"availability\"].float(),\n \"labels\": data[\"labels\"].float(),\n \"study_id\": data[\"study_id\"],\n \"fold\": int(data[\"fold\"])\n }\n\n\ncache_files = sorted(\n Path(\"/kaggle/working/knee_cache\").glob(\"*.pt\")\n)\n\nall_cached = CachedKneeDataset(cache_files)\n\nprint(\"Cached studies:\", len(all_cached))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:32:45.813751Z","iopub.execute_input":"2026-08-17T17:32:45.814302Z","iopub.status.idle":"2026-08-17T17:32:47.107494Z","shell.execute_reply.started":"2026-08-17T17:32:45.814265Z","shell.execute_reply":"2026-08-17T17:32:47.10655Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"fold = 0\n\ntrain_files = []\nvalid_files = []\n\nfor file in cache_files:\n\n data = torch.load(\n file,\n map_location=\"cpu\",\n weights_only=False\n )\n\n if int(data[\"fold\"]) == fold:\n valid_files.append(file)\n else:\n train_files.append(file)\n\nprint(\"Fold:\", fold)\nprint(\"Training studies:\", len(train_files))\nprint(\"Validation studies:\", len(valid_files))\n\ntrain_dataset = CachedKneeDataset(train_files)\nvalid_dataset = CachedKneeDataset(valid_files)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:32:58.031267Z","iopub.execute_input":"2026-08-17T17:32:58.032383Z","iopub.status.idle":"2026-08-17T17:33:00.465396Z","shell.execute_reply.started":"2026-08-17T17:32:58.032349Z","shell.execute_reply":"2026-08-17T17:33:00.464153Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_loader = DataLoader(\n train_dataset,\n batch_size=2,\n shuffle=True,\n num_workers=0,\n pin_memory=torch.cuda.is_available()\n)\n\nvalid_loader = DataLoader(\n valid_dataset,\n batch_size=2,\n shuffle=False,\n num_workers=0,\n pin_memory=torch.cuda.is_available()\n)\n\nprint(\"Train batches:\", len(train_loader))\nprint(\"Validation batches:\", len(valid_loader))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:33:15.149047Z","iopub.execute_input":"2026-08-17T17:33:15.149735Z","iopub.status.idle":"2026-08-17T17:33:15.156322Z","shell.execute_reply.started":"2026-08-17T17:33:15.149705Z","shell.execute_reply":"2026-08-17T17:33:15.155263Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"device = torch.device(\n \"cuda\" if torch.cuda.is_available()\n else \"cpu\"\n)\n\nmodel = BaselineV1(\n num_classes=12,\n num_sequences=4,\n pretrained=True\n).to(device)\n\ncriterion = nn.BCEWithLogitsLoss()\n\noptimizer = torch.optim.AdamW(\n model.parameters(),\n lr=1e-4,\n weight_decay=1e-4\n)\n\nprint(\"Device:\", device)\nprint(\"Model parameters:\",\n sum(p.numel() for p in model.parameters()))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T17:33:26.845384Z","iopub.execute_input":"2026-08-17T17:33:26.846244Z","iopub.status.idle":"2026-08-17T17:33:27.049073Z","shell.execute_reply.started":"2026-08-17T17:33:26.846207Z","shell.execute_reply":"2026-08-17T17:33:27.04802Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model.train()\n\nrunning_loss = 0.0\n\nfor batch_idx, batch in enumerate(train_loader):\n\n images = batch[\"images\"].to(\n device,\n non_blocking=True\n )\n\n availability = batch[\"availability\"].to(\n device,\n non_blocking=True\n )\n\n labels = batch[\"labels\"].to(\n device,\n non_blocking=True\n )\n\n optimizer.zero_grad()\n\n logits = model(\n images,\n availability\n )\n\n loss = criterion(\n logits,\n labels\n )\n\n loss.backward()\n\n optimizer.step()\n\n running_loss += loss.item()\n\n print(\n f\"Batch {batch_idx + 1}/{len(train_loader)} \"\n f\"| loss = {loss.item():.4f}\"\n )\n\nprint(\n \"\\nAverage training loss:\",\n running_loss / len(train_loader)\n)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model.eval()\n\nall_predictions = []\nall_labels = []\n\nwith torch.no_grad():\n\n for batch in valid_loader:\n\n images = batch[\"images\"].to(device)\n availability = batch[\"availability\"].to(device)\n\n logits = model(\n images,\n availability\n )\n\n probabilities = torch.sigmoid(\n logits\n )\n\n all_predictions.append(\n probabilities.cpu().numpy()\n )\n\n all_labels.append(\n batch[\"labels\"].numpy()\n )\n\npredictions = np.concatenate(\n all_predictions,\n axis=0\n)\n\nlabels = np.concatenate(\n all_labels,\n axis=0\n)\n\nprint(\"Prediction shape:\", predictions.shape)\nprint(\"Label shape:\", labels.shape)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"auc_results = {}\n\nfor i, target in enumerate(targets):\n\n y_true = labels[:, i]\n y_pred = predictions[:, i]\n\n # AUC is undefined if validation contains only one class\n if len(np.unique(y_true)) < 2:\n auc_results[target] = np.nan\n continue\n\n auc_results[target] = roc_auc_score(\n y_true,\n y_pred\n )\n\nauc_table = pd.DataFrame(\n {\n \"target\": list(auc_results.keys()),\n \"auc\": list(auc_results.values())\n }\n)\n\ndisplay(auc_table)\n\nmacro_auc = np.nanmean(\n list(auc_results.values())\n)\n\nprint(\"\\nFold 0 Macro ROC-AUC:\", macro_auc)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# RSNA Knee Abnormality Detection\n# BASELINE V1 — Frozen ResNet18 + Study-Level Classifier\n#\n# Pipeline:\n# cached MRI tensors\n# ↓\n# ImageNet-pretrained ResNet18 (FROZEN)\n# ↓\n# 16 slice embeddings → mean pooling\n# ↓\n# 4 sequence embeddings\n# ↓\n# availability mask\n# ↓\n# study-level logistic regression\n# ↓\n# 5-fold OOF predictions\n# ↓\n# macro ROC-AUC\n# ============================================================\n\nimport os\nimport time\nimport random\nfrom pathlib import Path\n\nimport numpy as np\nimport pandas as pd\nimport torch\nimport torch.nn as nn\nimport torchvision.models as models\n\nfrom sklearn.pipeline import Pipeline\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.linear_model import LogisticRegression\nfrom sklearn.metrics import roc_auc_score\n\n\n# ------------------------------------------------------------\n# 0. 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\ndevice = torch.device(\n \"cuda\" if torch.cuda.is_available() else \"cpu\"\n)\n\nprint(\"Device:\", device)\n\n\n# ------------------------------------------------------------\n# 1. Paths and configuration\n# ------------------------------------------------------------\n\nCACHE_ROOT = Path(\n \"/kaggle/working/knee_cache\"\n)\n\nMETADATA_PATH = Path(\n \"/kaggle/working/study_metadata_v2_folds.csv\"\n)\n\nOUTPUT_ROOT = Path(\n \"/kaggle/working/baseline_v1\"\n)\n\nOUTPUT_ROOT.mkdir(\n parents=True,\n exist_ok=True\n)\n\nNUM_SEQUENCES = 4\nNUM_SLICES = 16\nFEATURE_DIM = 512\n\nSEQUENCES = [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\",\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# ------------------------------------------------------------\n# 2. Load metadata\n# ------------------------------------------------------------\n\nmetadata = pd.read_csv(\n METADATA_PATH\n).reset_index(drop=True)\n\nprint(\n \"\\nNumber of labelled studies:\",\n len(metadata)\n)\n\nprint(\n \"Fold counts:\"\n)\n\nprint(\n metadata[\"fold\"]\n .value_counts()\n .sort_index()\n)\n\n\n# ------------------------------------------------------------\n# 3. Load pretrained ResNet18\n# ------------------------------------------------------------\n\nprint(\"\\nLoading pretrained ResNet18...\")\n\nweights = models.ResNet18_Weights.DEFAULT\n\nbackbone = models.resnet18(\n weights=weights\n)\n\nbackbone.fc = nn.Identity()\n\nbackbone = backbone.to(device)\n\nbackbone.eval()\n\n# IMPORTANT:\n# We are freezing the image encoder.\nfor parameter in backbone.parameters():\n parameter.requires_grad = False\n\nfeature_extractor_preprocess = weights.transforms()\n\nprint(\n \"Feature dimension:\",\n FEATURE_DIM\n)\n\n\n# ------------------------------------------------------------\n# 4. Extract frozen image features\n# ------------------------------------------------------------\n#\n# Each study becomes:\n#\n# sequence 1 → 512\n# sequence 2 → 512\n# sequence 3 → 512\n# sequence 4 → 512\n# availability → 4\n#\n# Final vector:\n#\n# 512*4 + 4 = 2052 dimensions\n#\n# ------------------------------------------------------------\n\nembedding_path = (\n OUTPUT_ROOT / \"study_embeddings.pt\"\n)\n\nif embedding_path.exists():\n\n print(\n \"\\nExisting embedding file found.\"\n )\n\n embedding_data = torch.load(\n embedding_path,\n map_location=\"cpu\",\n weights_only=False\n )\n\n embeddings = embedding_data[\"embeddings\"]\n availability_matrix = embedding_data[\n \"availability\"\n ]\n study_ids = embedding_data[\n \"study_ids\"\n ]\n\n print(\n \"Loaded cached embeddings:\",\n embeddings.shape\n )\n\nelse:\n\n print(\n \"\\nExtracting frozen image embeddings...\"\n )\n\n study_embeddings = []\n study_availability = []\n study_ids = []\n\n cache_files = sorted(\n CACHE_ROOT.glob(\"*.pt\")\n )\n\n # Map study ID → cache file\n cache_map = {}\n\n for file_path in cache_files:\n\n cache_map[\n file_path.stem\n ] = file_path\n\n start_time = time.time()\n\n with torch.inference_mode():\n\n for index, row in metadata.iterrows():\n\n study_id = row[\n \"StudyInstanceUID\"\n ]\n\n cache_file = cache_map.get(\n study_id\n )\n\n if cache_file is None:\n raise FileNotFoundError(\n f\"No cache found for {study_id}\"\n )\n\n sample = torch.load(\n cache_file,\n map_location=\"cpu\",\n weights_only=False\n )\n\n images = sample[\n \"images\"\n ].float()\n\n availability = sample[\n \"availability\"\n ].float()\n\n # ------------------------------------------------\n # images shape:\n # [4, 16, 1, 224, 224]\n # ------------------------------------------------\n\n sequence_features = []\n\n for sequence_idx in range(\n NUM_SEQUENCES\n ):\n\n if (\n availability[\n sequence_idx\n ].item() == 0\n ):\n # Missing sequence:\n # use an all-zero feature vector.\n sequence_feature = torch.zeros(\n FEATURE_DIM,\n dtype=torch.float32\n )\n\n else:\n\n sequence_images = images[\n sequence_idx\n ]\n\n # [16, 1, 224, 224]\n #\n # Repeat grayscale → RGB\n sequence_images = (\n sequence_images\n .repeat(1, 3, 1, 1)\n )\n\n # ImageNet normalization\n processed = []\n\n for image in sequence_images:\n\n image = (\n image * 255.0\n ).clamp(\n 0,\n 255\n )\n\n image = (\n image / 255.0\n )\n\n image = (\n image\n )\n\n image = (\n image\n )\n\n image = (\n image\n )\n\n # Normalize with\n # ImageNet statistics\n image = (\n image\n - torch.tensor(\n [0.485, 0.456, 0.406]\n ).view(3, 1, 1)\n ) / torch.tensor(\n [0.229, 0.224, 0.225]\n ).view(3, 1, 1)\n\n processed.append(\n image\n )\n\n sequence_images = torch.stack(\n processed,\n dim=0\n )\n\n # Process all 16 slices\n # together through ResNet.\n sequence_images = (\n sequence_images\n .to(device)\n )\n\n features = backbone(\n sequence_images\n )\n\n # [16, 512]\n #\n # Study-level representation\n # for this sequence.\n sequence_feature = (\n features.mean(\n dim=0\n )\n .cpu()\n )\n\n sequence_features.append(\n sequence_feature\n )\n\n # [4, 512]\n sequence_features = torch.stack(\n sequence_features\n )\n\n # Flatten:\n # [2048]\n sequence_features = (\n sequence_features\n .flatten()\n )\n\n # Add availability:\n # [2052]\n final_embedding = torch.cat(\n [\n sequence_features,\n availability\n ]\n )\n\n study_embeddings.append(\n final_embedding\n )\n\n study_availability.append(\n availability\n )\n\n study_ids.append(\n study_id\n )\n\n if (\n (index + 1) % 10 == 0\n or index == len(metadata) - 1\n ):\n\n elapsed = (\n time.time()\n - start_time\n )\n\n print(\n f\"Processed \"\n f\"{index + 1}/{len(metadata)} \"\n f\"| elapsed \"\n f\"{elapsed:.1f}s\"\n )\n\n embeddings = torch.stack(\n study_embeddings\n )\n\n availability_matrix = torch.stack(\n study_availability\n )\n\n torch.save(\n {\n \"embeddings\": embeddings,\n \"availability\": availability_matrix,\n \"study_ids\": study_ids,\n },\n embedding_path\n )\n\n print(\n \"\\nSaved embeddings:\",\n embedding_path\n )\n\n\nprint(\n \"\\nEmbedding shape:\",\n tuple(embeddings.shape)\n)\n\n\n# ------------------------------------------------------------\n# 5. Prepare labels and folds\n# ------------------------------------------------------------\n\nX = embeddings.numpy().astype(\n np.float32\n)\n\nY = metadata[\n TARGETS\n].values.astype(\n np.float32\n)\n\nfolds = metadata[\n \"fold\"\n].values.astype(\n int\n)\n\n\n# ------------------------------------------------------------\n# 6. Five-fold OOF training\n# ------------------------------------------------------------\n#\n# We use a regularized logistic regression because:\n#\n# - only 58 labelled studies exist\n# - feature dimension is high\n# - a large neural classifier would overfit easily\n# - the resulting model is fast and reproducible\n#\n# ------------------------------------------------------------\n\noof_predictions = np.zeros_like(\n Y,\n dtype=np.float32\n)\n\nfold_results = []\n\nprint(\n \"\\n\"\n + \"=\" * 70\n)\n\nprint(\n \"STARTING 5-FOLD BASELINE V1\"\n)\n\nprint(\n \"=\" * 70\n)\n\n\nfor fold in range(5):\n\n print(\n f\"\\n{'=' * 70}\"\n )\n\n print(\n f\"FOLD {fold}\"\n )\n\n print(\n f\"{'=' * 70}\"\n )\n\n train_idx = np.where(\n folds != fold\n )[0]\n\n valid_idx = np.where(\n folds == fold\n )[0]\n\n X_train = X[\n train_idx\n ]\n\n X_valid = X[\n valid_idx\n ]\n\n print(\n \"Train studies:\",\n len(train_idx)\n )\n\n print(\n \"Validation studies:\",\n len(valid_idx)\n )\n\n fold_auc = []\n\n for target_idx, target in enumerate(\n TARGETS\n ):\n\n y_train = Y[\n train_idx,\n target_idx\n ]\n\n y_valid = Y[\n valid_idx,\n target_idx\n ]\n\n # ----------------------------------------------------\n # Regularized classifier\n # ----------------------------------------------------\n\n model = Pipeline(\n [\n (\n \"scaler\",\n StandardScaler()\n ),\n (\n \"classifier\",\n LogisticRegression(\n C=0.1,\n class_weight=\"balanced\",\n max_iter=3000,\n solver=\"liblinear\",\n random_state=SEED\n )\n )\n ]\n )\n\n model.fit(\n X_train,\n y_train\n )\n\n probabilities = model.predict_proba(\n X_valid\n )[:, 1]\n\n oof_predictions[\n valid_idx,\n target_idx\n ] = probabilities\n\n # AUC should be valid because\n # we explicitly checked that each\n # target has positives in every fold.\n auc = roc_auc_score(\n y_valid,\n probabilities\n )\n\n fold_auc.append(\n auc\n )\n\n print(\n f\"{target:20s} \"\n f\"AUC = {auc:.4f}\"\n )\n\n macro_auc = float(\n np.mean(fold_auc)\n )\n\n fold_results.append(\n {\n \"fold\": fold,\n \"macro_auc\": macro_auc,\n \"n_train\": len(train_idx),\n \"n_valid\": len(valid_idx)\n }\n )\n\n print(\n f\"\\nFold {fold} Macro AUC: \"\n f\"{macro_auc:.4f}\"\n )\n\n\n# ------------------------------------------------------------\n# 7. Overall OOF score\n# ------------------------------------------------------------\n\nprint(\n \"\\n\"\n + \"=\" * 70\n)\n\nprint(\n \"OVERALL OOF RESULTS\"\n)\n\nprint(\n \"=\" * 70\n)\n\n\noof_auc = {}\n\nfor target_idx, target in enumerate(\n TARGETS\n):\n\n auc = roc_auc_score(\n Y[:, target_idx],\n oof_predictions[:, target_idx]\n )\n\n oof_auc[target] = auc\n\n\nresults_table = pd.DataFrame(\n {\n \"target\": TARGETS,\n \"OOF_AUC\": [\n oof_auc[t]\n for t in TARGETS\n ]\n }\n)\n\nresults_table = results_table.sort_values(\n \"OOF_AUC\",\n ascending=False\n).reset_index(\n drop=True\n)\n\ndisplay(\n results_table\n)\n\noverall_macro_auc = float(\n results_table[\"OOF_AUC\"].mean()\n)\n\nprint(\n \"\\nOverall OOF Macro ROC-AUC:\",\n f\"{overall_macro_auc:.6f}\"\n)\n\n\n# ------------------------------------------------------------\n# 8. Fold summary\n# ------------------------------------------------------------\n\nfold_results_df = pd.DataFrame(\n fold_results\n)\n\nprint(\n \"\\nFold-level results:\"\n)\n\ndisplay(\n fold_results_df\n)\n\n\nprint(\n \"\\nMean fold Macro AUC:\",\n f\"{fold_results_df['macro_auc'].mean():.6f}\"\n)\n\nprint(\n \"Std fold Macro AUC:\",\n f\"{fold_results_df['macro_auc'].std(ddof=1):.6f}\"\n)\n\n\n# ------------------------------------------------------------\n# 9. Save predictions/results\n# ------------------------------------------------------------\n\noof_df = pd.DataFrame(\n oof_predictions,\n columns=[\n f\"{target}_prediction\"\n for target in TARGETS\n ]\n)\n\noof_df.insert(\n 0,\n \"StudyInstanceUID\",\n metadata[\n \"StudyInstanceUID\"\n ].values\n)\n\noof_df[\"fold\"] = folds\n\noof_path = (\n OUTPUT_ROOT\n / \"oof_predictions.csv\"\n)\n\noof_df.to_csv(\n oof_path,\n index=False\n)\n\nresults_path = (\n OUTPUT_ROOT\n / \"baseline_v1_results.csv\"\n)\n\nresults_table.to_csv(\n results_path,\n index=False\n)\n\nfold_results_path = (\n OUTPUT_ROOT\n / \"fold_results.csv\"\n)\n\nfold_results_df.to_csv(\n fold_results_path,\n index=False\n)\n\n\n# ------------------------------------------------------------\n# 10. Final report\n# ------------------------------------------------------------\n\nprint(\n \"\\n\"\n + \"=\" * 70\n)\n\nprint(\n \"BASELINE V1 COMPLETE\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Overall OOF Macro ROC-AUC:\",\n f\"{overall_macro_auc:.6f}\"\n)\n\nprint(\n \"\\nSaved:\"\n)\n\nprint(\n \"Embeddings:\",\n embedding_path\n)\n\nprint(\n \"OOF predictions:\",\n oof_path\n)\n\nprint(\n \"Target results:\",\n results_path\n)\n\nprint(\n \"Fold results:\",\n fold_results_path\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T18:10:43.667708Z","iopub.execute_input":"2026-08-17T18:10:43.668222Z","iopub.status.idle":"2026-08-17T18:13:07.016147Z","shell.execute_reply.started":"2026-08-17T18:10:43.668176Z","shell.execute_reply":"2026-08-17T18:13:07.014951Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"results = pd.read_csv(\n \"/kaggle/working/baseline_v1/baseline_v1_results.csv\"\n)\n\ndisplay(\n results.sort_values(\n \"OOF_AUC\",\n ascending=False\n )\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T18:15:32.746598Z","iopub.execute_input":"2026-08-17T18:15:32.747267Z","iopub.status.idle":"2026-08-17T18:15:32.766132Z","shell.execute_reply.started":"2026-08-17T18:15:32.747215Z","shell.execute_reply":"2026-08-17T18:15:32.764954Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# RSNA Knee Abnormality Detection\n# BASELINE V2 — Target-Aware Slice Attention\n#\n# V1:\n# 16 slice embeddings -> MEAN -> sequence feature\n#\n# V2:\n# 16 slice embeddings -> target-aware ATTENTION -> sequence feature\n#\n# The ResNet-18 image encoder remains frozen.\n# We keep the exact same 5 folds as V1.\n# ============================================================\n\nimport random\nimport time\nfrom pathlib import Path\n\nimport numpy as np\nimport pandas as pd\nimport torch\nimport torch.nn as nn\nimport torchvision.models as models\n\nfrom sklearn.metrics import roc_auc_score\n\n\n# ============================================================\n# 0. Configuration\n# ============================================================\n\nSEED = 42\nNUM_FOLDS = 5\n\nNUM_SEQUENCES = 4\nNUM_SLICES = 16\nFEATURE_DIM = 512\nNUM_TARGETS = 12\n\nDEVICE = torch.device(\n \"cuda\" if torch.cuda.is_available() else \"cpu\"\n)\n\nCACHE_ROOT = Path(\n \"/kaggle/working/knee_cache\"\n)\n\nMETADATA_PATH = Path(\n \"/kaggle/working/study_metadata_v2_folds.csv\"\n)\n\nOUTPUT_ROOT = Path(\n \"/kaggle/working/baseline_v2\"\n)\n\nOUTPUT_ROOT.mkdir(\n parents=True,\n exist_ok=True\n)\n\nSEQUENCES = [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\",\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# ============================================================\n# 1. 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\nprint(\"Device:\", DEVICE)\n\n\n# ============================================================\n# 2. Load fixed metadata / folds\n# ============================================================\n\nmetadata = pd.read_csv(\n METADATA_PATH\n).reset_index(drop=True)\n\nY = metadata[TARGETS].values.astype(\n np.float32\n)\n\nFOLDS = metadata[\"fold\"].values.astype(\n int\n)\n\nprint(\"\\nStudies:\", len(metadata))\nprint(\"\\nFrozen fold sizes:\")\nprint(\n metadata[\"fold\"]\n .value_counts()\n .sort_index()\n)\n\n\n# ============================================================\n# 3. Frozen ResNet-18\n# ============================================================\n\nprint(\"\\nLoading pretrained ResNet-18...\")\n\nweights = models.ResNet18_Weights.DEFAULT\n\nbackbone = models.resnet18(\n weights=weights\n)\n\nbackbone.fc = nn.Identity()\n\nbackbone = backbone.to(DEVICE)\nbackbone.eval()\n\nfor p in backbone.parameters():\n p.requires_grad = False\n\nprint(\"Backbone feature dimension:\", FEATURE_DIM)\n\n\n# ============================================================\n# 4. Extract PER-SLICE embeddings\n#\n# Saved structure:\n#\n# [58, 4, 16, 512]\n#\n# 58 studies\n# 4 sequences\n# 16 slices\n# 512 features per slice\n# ============================================================\n\nslice_embedding_path = (\n OUTPUT_ROOT / \"slice_embeddings.pt\"\n)\n\nif slice_embedding_path.exists():\n\n print(\n \"\\nLoading existing slice embeddings...\"\n )\n\n saved = torch.load(\n slice_embedding_path,\n map_location=\"cpu\",\n weights_only=False\n )\n\n slice_embeddings = saved[\n \"slice_embeddings\"\n ]\n\n availability = saved[\n \"availability\"\n ]\n\n print(\n \"Loaded:\",\n tuple(slice_embeddings.shape)\n )\n\nelse:\n\n print(\n \"\\nExtracting frozen per-slice embeddings...\"\n )\n\n cache_map = {\n p.stem: p\n for p in CACHE_ROOT.glob(\"*.pt\")\n }\n\n all_embeddings = []\n all_availability = []\n\n extraction_start = time.time()\n\n with torch.inference_mode():\n\n for study_idx, row in metadata.iterrows():\n\n study_id = row[\n \"StudyInstanceUID\"\n ]\n\n cache_file = cache_map.get(\n study_id\n )\n\n if cache_file is None:\n raise FileNotFoundError(\n f\"Cache missing: {study_id}\"\n )\n\n sample = torch.load(\n cache_file,\n map_location=\"cpu\",\n weights_only=False\n )\n\n images = sample[\n \"images\"\n ].float()\n\n study_availability = sample[\n \"availability\"\n ].float()\n\n sequence_features = []\n\n for seq_idx in range(\n NUM_SEQUENCES\n ):\n\n if study_availability[\n seq_idx\n ].item() == 0:\n\n sequence_features.append(\n torch.zeros(\n NUM_SLICES,\n FEATURE_DIM,\n dtype=torch.float32\n )\n )\n\n continue\n\n sequence_images = images[\n seq_idx\n ]\n\n # [16,1,224,224] -> [16,3,224,224]\n sequence_images = (\n sequence_images\n .repeat(1, 3, 1, 1)\n )\n\n # ImageNet normalization\n mean = torch.tensor(\n [0.485, 0.456, 0.406],\n dtype=torch.float32\n ).view(1, 3, 1, 1)\n\n std = torch.tensor(\n [0.229, 0.224, 0.225],\n dtype=torch.float32\n ).view(1, 3, 1, 1)\n\n sequence_images = (\n sequence_images - mean\n ) / std\n\n sequence_images = (\n sequence_images\n .to(DEVICE)\n )\n\n features = backbone(\n sequence_images\n )\n\n # [16,512]\n sequence_features.append(\n features.cpu()\n )\n\n # [4,16,512]\n all_embeddings.append(\n torch.stack(\n sequence_features\n )\n )\n\n all_availability.append(\n study_availability\n )\n\n if (\n (study_idx + 1) % 10 == 0\n or study_idx == len(metadata) - 1\n ):\n\n elapsed = (\n time.time()\n - extraction_start\n )\n\n print(\n f\"Processed \"\n f\"{study_idx + 1}/\"\n f\"{len(metadata)} \"\n f\"| {elapsed:.1f}s\"\n )\n\n slice_embeddings = torch.stack(\n all_embeddings\n )\n\n availability = torch.stack(\n all_availability\n )\n\n torch.save(\n {\n \"slice_embeddings\":\n slice_embeddings,\n \"availability\":\n availability,\n },\n slice_embedding_path\n )\n\n print(\n \"\\nSaved:\",\n slice_embedding_path\n )\n\nprint(\n \"Slice embedding shape:\",\n tuple(slice_embeddings.shape)\n)\n\n\n# ============================================================\n# 5. Target-Aware Attention Model\n# ============================================================\n\nclass TargetAwareAttention(nn.Module):\n \"\"\"\n For each target, learn a query vector.\n\n For each sequence:\n\n 16 slice features\n ↓\n similarity with target query\n ↓\n softmax attention\n ↓\n weighted slice representation\n\n This lets ACL, MCL, OA, etc. focus on\n different slices.\n \"\"\"\n\n def __init__(\n self,\n num_targets=12,\n num_sequences=4,\n feature_dim=512,\n ):\n super().__init__()\n\n self.num_targets = num_targets\n self.num_sequences = num_sequences\n self.feature_dim = feature_dim\n\n # Small target-specific queries.\n self.query = nn.Parameter(\n torch.randn(\n num_targets,\n feature_dim\n ) * 0.02\n )\n\n # Temperature prevents extremely sharp\n # attention at initialization.\n self.temperature = nn.Parameter(\n torch.ones(num_targets)\n )\n\n # One lightweight target-specific\n # classifier per target.\n self.heads = nn.ModuleList(\n [\n nn.Sequential(\n nn.Linear(\n num_sequences * feature_dim\n + num_sequences,\n 128\n ),\n nn.LayerNorm(128),\n nn.GELU(),\n nn.Dropout(0.25),\n nn.Linear(128, 1)\n )\n for _ in range(num_targets)\n ]\n )\n\n def forward(\n self,\n x,\n availability\n ):\n \"\"\"\n x:\n [B, S, N, D]\n\n availability:\n [B, S]\n\n returns:\n logits [B, T]\n \"\"\"\n\n B, S, N, D = x.shape\n\n target_outputs = []\n\n # ----------------------------------------------------\n # Each target gets its own attention mechanism.\n # ----------------------------------------------------\n\n for target_idx in range(\n self.num_targets\n ):\n\n query = self.query[\n target_idx\n ]\n\n temperature = (\n self.temperature[\n target_idx\n ].abs()\n + 1e-4\n )\n\n # ------------------------------------------------\n # x:\n # [B,S,N,D]\n #\n # query:\n # [D]\n #\n # scores:\n # [B,S,N]\n # ------------------------------------------------\n\n scores = torch.einsum(\n \"bsnd,d->bsn\",\n x,\n query\n )\n\n scores = (\n scores\n / temperature\n )\n\n # ------------------------------------------------\n # Mask completely missing sequences.\n # ------------------------------------------------\n\n missing_mask = (\n availability == 0\n )\n\n scores = scores.masked_fill(\n missing_mask.unsqueeze(-1),\n -1e4\n )\n\n # ------------------------------------------------\n # Attention over slices\n # ------------------------------------------------\n\n attention = torch.softmax(\n scores,\n dim=2\n )\n\n # ------------------------------------------------\n # Weighted slice pooling\n # ------------------------------------------------\n\n pooled = torch.sum(\n x * attention.unsqueeze(-1),\n dim=2\n )\n\n # Remove representations from\n # missing sequences explicitly.\n pooled = (\n pooled\n * availability.unsqueeze(-1)\n )\n\n # [B,S,D] -> [B,S*D]\n pooled = pooled.reshape(\n B,\n S * D\n )\n\n fused = torch.cat(\n [\n pooled,\n availability\n ],\n dim=1\n )\n\n logits = self.heads[\n target_idx\n ](\n fused\n )\n\n target_outputs.append(\n logits\n )\n\n return torch.cat(\n target_outputs,\n dim=1\n )\n\n\n# ============================================================\n# 6. Training helpers\n# ============================================================\n\ndef evaluate_model(\n model,\n X_valid,\n A_valid,\n y_valid\n):\n\n model.eval()\n\n with torch.no_grad():\n\n logits = model(\n X_valid,\n A_valid\n )\n\n probabilities = (\n torch.sigmoid(\n logits\n )\n .cpu()\n .numpy()\n )\n\n y_true = (\n y_valid\n .cpu()\n .numpy()\n )\n\n aucs = []\n\n for target_idx in range(\n NUM_TARGETS\n ):\n\n auc = roc_auc_score(\n y_true[:, target_idx],\n probabilities[:, target_idx]\n )\n\n aucs.append(auc)\n\n return (\n np.asarray(aucs),\n probabilities\n )\n\n\n# ============================================================\n# 7. Five-fold training\n# ============================================================\n\noof_predictions = np.zeros(\n (\n len(metadata),\n NUM_TARGETS\n ),\n dtype=np.float32\n)\n\nfold_summaries = []\n\nprint(\n \"\\n\"\n + \"=\" * 70\n)\n\nprint(\n \"BASELINE V2 — TARGET-AWARE SLICE ATTENTION\"\n)\n\nprint(\n \"=\" * 70\n)\n\n\nfor fold in range(\n NUM_FOLDS\n):\n\n print(\n f\"\\n{'=' * 70}\"\n )\n\n print(\n f\"FOLD {fold}\"\n )\n\n print(\n f\"{'=' * 70}\"\n )\n\n train_idx = np.where(\n FOLDS != fold\n )[0]\n\n valid_idx = np.where(\n FOLDS == fold\n )[0]\n\n X_train = (\n slice_embeddings[\n train_idx\n ]\n .to(DEVICE)\n )\n\n A_train = (\n availability[\n train_idx\n ]\n .to(DEVICE)\n )\n\n y_train = torch.tensor(\n Y[train_idx],\n dtype=torch.float32,\n device=DEVICE\n )\n\n X_valid = (\n slice_embeddings[\n valid_idx\n ]\n .to(DEVICE)\n )\n\n A_valid = (\n availability[\n valid_idx\n ]\n .to(DEVICE)\n )\n\n y_valid = torch.tensor(\n Y[valid_idx],\n dtype=torch.float32,\n device=DEVICE\n )\n\n model = TargetAwareAttention(\n num_targets=NUM_TARGETS,\n num_sequences=NUM_SEQUENCES,\n feature_dim=FEATURE_DIM\n ).to(DEVICE)\n\n # --------------------------------------------------------\n # Positive-class weighting computed ONLY from training fold.\n # --------------------------------------------------------\n\n pos = (\n Y[train_idx]\n .sum(axis=0)\n )\n\n neg = (\n len(train_idx)\n - pos\n )\n\n pos_weight = torch.tensor(\n neg / np.maximum(pos, 1),\n dtype=torch.float32,\n device=DEVICE\n )\n\n criterion = nn.BCEWithLogitsLoss(\n pos_weight=pos_weight\n )\n\n optimizer = torch.optim.AdamW(\n model.parameters(),\n lr=2e-3,\n weight_decay=1e-3\n )\n\n scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(\n optimizer,\n T_max=120\n )\n\n best_auc = -np.inf\n best_state = None\n patience = 25\n epochs_without_improvement = 0\n\n for epoch in range(1, 121):\n\n model.train()\n\n optimizer.zero_grad()\n\n logits = model(\n X_train,\n A_train\n )\n\n loss = criterion(\n logits,\n y_train\n )\n\n loss.backward()\n\n torch.nn.utils.clip_grad_norm_(\n model.parameters(),\n max_norm=1.0\n )\n\n optimizer.step()\n scheduler.step()\n\n # ----------------------------------------------------\n # Validation every epoch.\n # Dataset is tiny, so this is cheap.\n # ----------------------------------------------------\n\n valid_aucs, _ = evaluate_model(\n model,\n X_valid,\n A_valid,\n y_valid\n )\n\n macro_auc = float(\n np.mean(valid_aucs)\n )\n\n if macro_auc > best_auc:\n\n best_auc = macro_auc\n\n best_state = {\n k: v.detach()\n .cpu()\n .clone()\n for k, v in model.state_dict().items()\n }\n\n epochs_without_improvement = 0\n\n else:\n\n epochs_without_improvement += 1\n\n if (\n epoch == 1\n or epoch % 10 == 0\n or epoch == 120\n ):\n\n print(\n f\"Epoch {epoch:03d} \"\n f\"| loss {loss.item():.4f} \"\n f\"| val macro AUC \"\n f\"{macro_auc:.4f} \"\n f\"| best \"\n f\"{best_auc:.4f}\"\n )\n\n if (\n epochs_without_improvement\n >= patience\n ):\n print(\n f\"Early stopping at epoch \"\n f\"{epoch}\"\n )\n break\n\n # --------------------------------------------------------\n # Restore best epoch.\n # --------------------------------------------------------\n\n model.load_state_dict(\n best_state\n )\n\n final_aucs, final_predictions = (\n evaluate_model(\n model,\n X_valid,\n A_valid,\n y_valid\n )\n )\n\n oof_predictions[\n valid_idx\n ] = final_predictions\n\n fold_macro = float(\n np.mean(final_aucs)\n )\n\n fold_summaries.append(\n {\n \"fold\": fold,\n \"n_train\": len(train_idx),\n \"n_valid\": len(valid_idx),\n \"best_macro_auc\": fold_macro\n }\n )\n\n print(\n f\"\\nFold {fold} FINAL macro AUC:\"\n f\" {fold_macro:.6f}\"\n )\n\n # Save fold model\n torch.save(\n {\n \"model_state_dict\":\n model.state_dict(),\n \"fold\": fold,\n \"best_macro_auc\":\n fold_macro\n },\n OUTPUT_ROOT\n / f\"fold_{fold}_model.pt\"\n )\n\n\n# ============================================================\n# 8. Overall OOF evaluation\n# ============================================================\n\nprint(\n \"\\n\"\n + \"=\" * 70\n)\n\nprint(\n \"BASELINE V2 — OOF RESULTS\"\n)\n\nprint(\n \"=\" * 70\n)\n\noverall_aucs = []\n\nfor target_idx, target in enumerate(\n TARGETS\n):\n\n auc = roc_auc_score(\n Y[:, target_idx],\n oof_predictions[:, target_idx]\n )\n\n overall_aucs.append(\n auc\n )\n\n\nresults = pd.DataFrame(\n {\n \"target\": TARGETS,\n \"OOF_AUC\": overall_aucs\n }\n).sort_values(\n \"OOF_AUC\",\n ascending=False\n).reset_index(\n drop=True\n)\n\ndisplay(\n results\n)\n\noverall_macro_auc = float(\n np.mean(overall_aucs)\n)\n\nprint(\n \"\\nOverall OOF Macro ROC-AUC:\",\n f\"{overall_macro_auc:.6f}\"\n)\n\n\n# ============================================================\n# 9. Fold summary\n# ============================================================\n\nfold_results = pd.DataFrame(\n fold_summaries\n)\n\nprint(\n \"\\nFold summary:\"\n)\n\ndisplay(\n fold_results\n)\n\nprint(\n \"\\nMean fold Macro AUC:\",\n f\"{fold_results['best_macro_auc'].mean():.6f}\"\n)\n\nprint(\n \"Std fold Macro AUC:\",\n f\"{fold_results['best_macro_auc'].std(ddof=1):.6f}\"\n)\n\n\n# ============================================================\n# 10. Compare against V1\n# ============================================================\n\nV1_AUC = 0.626641\n\nprint(\n \"\\n\"\n + \"=\" * 70\n)\n\nprint(\n \"V1 vs V2\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n f\"V1 Macro AUC: \"\n f\"{V1_AUC:.6f}\"\n)\n\nprint(\n f\"V2 Macro AUC: \"\n f\"{overall_macro_auc:.6f}\"\n)\n\nprint(\n f\"Absolute improvement: \"\n f\"{overall_macro_auc - V1_AUC:+.6f}\"\n)\n\n\n# ============================================================\n# 11. Save everything\n# ============================================================\n\noof_output = pd.DataFrame(\n oof_predictions,\n columns=[\n f\"{target}_prediction\"\n for target in TARGETS\n ]\n)\n\noof_output.insert(\n 0,\n \"StudyInstanceUID\",\n metadata[\n \"StudyInstanceUID\"\n ].values\n)\n\noof_output[\"fold\"] = FOLDS\n\noof_output.to_csv(\n OUTPUT_ROOT\n / \"oof_predictions.csv\",\n index=False\n)\n\nresults.to_csv(\n OUTPUT_ROOT\n / \"baseline_v2_results.csv\",\n index=False\n)\n\nfold_results.to_csv(\n OUTPUT_ROOT\n / \"fold_results.csv\",\n index=False\n)\n\nprint(\n \"\\nSaved V2 results to:\"\n)\n\nprint(\n OUTPUT_ROOT\n)\n\nprint(\n \"\\nBASELINE V2 COMPLETE.\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T18:17:31.669858Z","iopub.execute_input":"2026-08-17T18:17:31.670822Z","iopub.status.idle":"2026-08-17T18:20:23.415656Z","shell.execute_reply.started":"2026-08-17T18:17:31.670784Z","shell.execute_reply":"2026-08-17T18:20:23.414674Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# RSNA KNEE ABNORMALITY DETECTION\n# FINAL V1 SUBMISSION PIPELINE\n#\n# Uses the BEST VALIDATED model so far:\n# Frozen ResNet18 embeddings\n# + mean slice pooling\n# + 4 sequence fusion\n# + availability mask\n# + one LogisticRegression classifier per target\n#\n# IMPORTANT:\n# - Uses the SAME preprocessing as V1\n# - Uses the SAME four sequence types\n# - Trains final classifiers on all 58 labelled studies\n# - Generates predictions for ALL test studies\n# - Writes submission.csv\n# ============================================================\n\nimport time\nimport random\nfrom pathlib import Path\n\nimport numpy as np\nimport pandas as pd\nimport torch\nimport torch.nn as nn\nimport pydicom\nimport torchvision.models as models\n\nfrom sklearn.pipeline import Pipeline\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.linear_model import LogisticRegression\n\n\n# ============================================================\n# 1. Configuration\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\nDEVICE = torch.device(\n \"cuda\" if torch.cuda.is_available() else \"cpu\"\n)\n\nROOT = Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n)\n\nTRAIN_PATH = ROOT / \"train.csv\"\nTEST_PATH = ROOT / \"test.csv\"\nTRAIN_SERIES_PATH = ROOT / \"train_series.csv\"\nTEST_SERIES_PATH = ROOT / \"test_series.csv\"\nSAMPLE_PATH = ROOT / \"sample_submission.csv\"\n\nOUTPUT_DIR = Path(\n \"/kaggle/working/final_submission_v1\"\n)\nOUTPUT_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\nSEQUENCE_TYPES = [\n (\"Axial\", 1, 1, \"axial_f1_fs1\"),\n (\"Sagittal\", 0, 0, \"sagittal_f0_fs0\"),\n (\"Coronal\", 1, 1, \"coronal_f1_fs1\"),\n (\"Sagittal\", 1, 1, \"sagittal_f1_fs1\"),\n]\n\nNUM_SLICES = 16\nIMAGE_SIZE = 224\nFEATURE_DIM = 512\n\nprint(\"Device:\", DEVICE)\n\n\n# ============================================================\n# 2. Load competition tables\n# ============================================================\n\ntrain = pd.read_csv(TRAIN_PATH)\ntest = pd.read_csv(TEST_PATH)\ntrain_series = pd.read_csv(TRAIN_SERIES_PATH)\ntest_series = pd.read_csv(TEST_SERIES_PATH)\nsample_submission = pd.read_csv(SAMPLE_PATH)\n\nprint(\"\\nTrain shape:\", train.shape)\nprint(\"Test shape:\", test.shape)\nprint(\"Train series shape:\", train_series.shape)\nprint(\"Test series shape:\", test_series.shape)\nprint(\"Sample submission shape:\", sample_submission.shape)\n\n\n# ============================================================\n# 3. Select the 58 genuinely labelled studies\n# ============================================================\n\nlabelled_train = train[\n train[TARGETS].notna().all(axis=1)\n].copy()\n\nprint(\n \"\\nFully labelled studies:\",\n len(labelled_train)\n)\n\nif len(labelled_train) != 58:\n raise RuntimeError(\n f\"Expected 58 labelled studies, \"\n f\"found {len(labelled_train)}.\"\n )\n\n\n# ============================================================\n# 4. DICOM utilities\n# ============================================================\n\ndef load_dicom_volume(series_path):\n \"\"\"\n Load one DICOM series and order slices using\n ImageOrientationPatient + ImagePositionPatient.\n \"\"\"\n\n series_path = Path(series_path)\n\n files = list(series_path.glob(\"*.dcm\"))\n\n if not files:\n raise FileNotFoundError(\n f\"No DICOM files found in {series_path}\"\n )\n\n records = []\n\n for file_path in files:\n\n ds = pydicom.dcmread(\n file_path,\n stop_before_pixels=False\n )\n\n position = np.asarray(\n ds.ImagePositionPatient,\n dtype=np.float64\n )\n\n orientation = np.asarray(\n ds.ImageOrientationPatient,\n dtype=np.float64\n )\n\n row_direction = orientation[:3]\n column_direction = orientation[3:]\n\n normal = np.cross(\n row_direction,\n column_direction\n )\n\n physical_position = np.dot(\n position,\n normal\n )\n\n records.append(\n (\n physical_position,\n file_path\n )\n )\n\n records.sort(\n key=lambda x: x[0]\n )\n\n images = []\n\n for _, file_path in records:\n\n ds = pydicom.dcmread(\n file_path,\n stop_before_pixels=False\n )\n\n image = ds.pixel_array.astype(\n np.float32\n )\n\n images.append(image)\n\n return np.stack(\n images,\n axis=0\n )\n\n\ndef sample_slices(\n volume,\n num_slices=NUM_SLICES\n):\n\n n = volume.shape[0]\n\n indices = np.linspace(\n 0,\n n - 1,\n num_slices\n ).round().astype(int)\n\n return volume[indices]\n\n\ndef normalize_slice(image):\n\n image = image.astype(\n np.float32\n )\n\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 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\ndef preprocess_volume(volume):\n\n volume = sample_slices(\n volume,\n NUM_SLICES\n )\n\n processed = []\n\n for image in volume:\n\n image = normalize_slice(\n image\n )\n\n tensor = torch.from_numpy(\n image\n ).unsqueeze(0).unsqueeze(0)\n\n tensor = torch.nn.functional.interpolate(\n tensor,\n size=(\n IMAGE_SIZE,\n IMAGE_SIZE\n ),\n mode=\"bilinear\",\n align_corners=False\n )\n\n processed.append(\n tensor.squeeze(0)\n )\n\n return torch.stack(\n processed\n )\n\n\n# ============================================================\n# 5. Build deterministic series selection\n# ============================================================\n\ndef build_selected_series_table(\n series_df,\n study_ids\n):\n\n rows = []\n\n study_id_set = set(\n study_ids\n )\n\n for plane, fluid, fat, name in SEQUENCE_TYPES:\n\n subset = series_df[\n (series_df[\"Anatomical_Plane\"] == plane) &\n (series_df[\"Fluid_Sensitive\"] == fluid) &\n (series_df[\"Fat_Suppression\"] == fat) &\n (series_df[\"StudyInstanceUID\"].isin(study_id_set))\n ].copy()\n\n for study_id, group in subset.groupby(\n \"StudyInstanceUID\"\n ):\n\n best_series = None\n best_count = -1\n\n for _, series_row in group.iterrows():\n\n series_id = series_row[\n \"SeriesInstanceUID\"\n ]\n\n series_path = (\n ROOT\n / (\n \"test_series\"\n if series_df is test_series\n else \"train_series\"\n )\n / str(study_id)\n / str(series_id)\n )\n\n n_slices = len(\n list(\n series_path.glob(\n \"*.dcm\"\n )\n )\n )\n\n if n_slices > best_count:\n\n best_count = n_slices\n best_series = series_id\n\n if best_series is not None:\n\n rows.append(\n {\n \"StudyInstanceUID\":\n study_id,\n \"sequence_type\":\n name,\n \"SeriesInstanceUID\":\n best_series,\n \"num_slices\":\n best_count\n }\n )\n\n selected = pd.DataFrame(rows)\n\n pivot = selected.pivot(\n index=\"StudyInstanceUID\",\n columns=\"sequence_type\",\n values=\"SeriesInstanceUID\"\n ).reset_index()\n\n for _, _, _, name in SEQUENCE_TYPES:\n\n if name not in pivot.columns:\n pivot[name] = np.nan\n\n return pivot[\n [\n \"StudyInstanceUID\",\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\"\n ]\n ]\n\n\nprint(\n \"\\nBuilding test study → series mapping...\"\n)\n\ntest_series_map = build_selected_series_table(\n test_series,\n test[\"StudyInstanceUID\"].tolist()\n)\n\nprint(\n \"Test studies with mappings:\",\n len(test_series_map)\n)\n\ndisplay(\n test_series_map.head()\n)\n\n\n# ============================================================\n# 6. Load frozen ResNet-18 feature extractor\n# ============================================================\n\nprint(\n \"\\nLoading pretrained ResNet-18...\"\n)\n\nweights = models.ResNet18_Weights.DEFAULT\n\nbackbone = models.resnet18(\n weights=weights\n)\n\nbackbone.fc = nn.Identity()\n\nbackbone = backbone.to(\n DEVICE\n)\n\nbackbone.eval()\n\nfor parameter in backbone.parameters():\n parameter.requires_grad = False\n\n\n# ImageNet normalization\nimagenet_mean = torch.tensor(\n [0.485, 0.456, 0.406],\n dtype=torch.float32,\n device=DEVICE\n).view(\n 1, 3, 1, 1\n)\n\nimagenet_std = torch.tensor(\n [0.229, 0.224, 0.225],\n dtype=torch.float32,\n device=DEVICE\n).view(\n 1, 3, 1, 1\n)\n\n\n# ============================================================\n# 7. Extract final TRAIN embeddings\n#\n# We use the SAME representation as V1:\n#\n# 4 sequences × mean of 16 slice embeddings\n# + 4 availability indicators\n#\n# Final dimension = 2052\n# ============================================================\n\nprint(\n \"\\nExtracting final training embeddings...\"\n)\n\ntrain_series_map = build_selected_series_table(\n train_series,\n labelled_train[\"StudyInstanceUID\"].tolist()\n)\n\ntrain_embedding_rows = []\n\nstart = time.time()\n\nwith torch.inference_mode():\n\n for count, study_id in enumerate(\n labelled_train[\"StudyInstanceUID\"]\n ):\n\n mapping = train_series_map[\n train_series_map[\"StudyInstanceUID\"]\n == study_id\n ]\n\n if len(mapping) != 1:\n raise RuntimeError(\n f\"Missing/duplicate mapping for \"\n f\"train study {study_id}\"\n )\n\n mapping = mapping.iloc[0]\n\n sequence_features = []\n availability = []\n\n for sequence in [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\"\n ]:\n\n series_id = mapping[\n sequence\n ]\n\n if pd.isna(series_id):\n\n sequence_features.append(\n torch.zeros(\n FEATURE_DIM,\n dtype=torch.float32\n )\n )\n\n availability.append(0.0)\n continue\n\n series_path = (\n ROOT\n / \"train_series\"\n / str(study_id)\n / str(series_id)\n )\n\n volume = load_dicom_volume(\n series_path\n )\n\n images = preprocess_volume(\n volume\n )\n\n # [16,1,224,224] → RGB\n images = images.repeat(\n 1, 3, 1, 1\n )\n\n images = (\n images\n .to(DEVICE)\n )\n\n images = (\n images - imagenet_mean\n ) / imagenet_std\n\n slice_features = backbone(\n images\n )\n\n study_feature = (\n slice_features\n .mean(dim=0)\n .cpu()\n )\n\n sequence_features.append(\n study_feature\n )\n\n availability.append(1.0)\n\n feature_vector = torch.cat(\n sequence_features\n + [\n torch.tensor(\n availability,\n dtype=torch.float32\n )\n ]\n )\n\n train_embedding_rows.append(\n feature_vector.numpy()\n )\n\n if (\n (count + 1) % 10 == 0\n or count == len(labelled_train) - 1\n ):\n\n elapsed = (\n time.time() - start\n )\n\n print(\n f\"Train embeddings: \"\n f\"{count + 1}/\"\n f\"{len(labelled_train)} \"\n f\"| {elapsed:.1f}s\"\n )\n\nX_train = np.stack(\n train_embedding_rows\n).astype(\n np.float32\n)\n\nY_train = labelled_train[\n TARGETS\n].values.astype(\n np.float32\n)\n\nprint(\n \"\\nFinal train embedding shape:\",\n X_train.shape\n)\n\n\n# ============================================================\n# 8. Train FINAL classifiers on ALL 58 labelled studies\n# ============================================================\n\nprint(\n \"\\nTraining final 12 target classifiers...\"\n)\n\nfinal_models = []\n\nfor target_idx, target in enumerate(\n TARGETS\n):\n\n y = Y_train[\n :,\n target_idx\n ]\n\n model = Pipeline(\n [\n (\n \"scaler\",\n StandardScaler()\n ),\n (\n \"classifier\",\n LogisticRegression(\n C=0.1,\n class_weight=\"balanced\",\n max_iter=3000,\n solver=\"liblinear\",\n random_state=SEED\n )\n )\n ]\n )\n\n model.fit(\n X_train,\n y\n )\n\n final_models.append(\n model\n )\n\n print(\n f\"Trained: {target}\"\n )\n\n\n# ============================================================\n# 9. Extract TEST embeddings\n# ============================================================\n\nprint(\n \"\\nExtracting test embeddings...\"\n)\n\ntest_embedding_rows = []\ntest_ids = test[\n \"StudyInstanceUID\"\n].tolist()\n\nstart = time.time()\n\nwith torch.inference_mode():\n\n for count, study_id in enumerate(\n test_ids\n ):\n\n mapping = test_series_map[\n test_series_map[\"StudyInstanceUID\"]\n == study_id\n ]\n\n if len(mapping) != 1:\n raise RuntimeError(\n f\"Missing/duplicate mapping \"\n f\"for test study {study_id}\"\n )\n\n mapping = mapping.iloc[0]\n\n sequence_features = []\n availability = []\n\n for sequence in [\n \"axial_f1_fs1\",\n \"sagittal_f0_fs0\",\n \"coronal_f1_fs1\",\n \"sagittal_f1_fs1\"\n ]:\n\n series_id = mapping[\n sequence\n ]\n\n if pd.isna(series_id):\n\n sequence_features.append(\n torch.zeros(\n FEATURE_DIM,\n dtype=torch.float32\n )\n )\n\n availability.append(0.0)\n\n continue\n\n series_path = (\n ROOT\n / \"test_series\"\n / str(study_id)\n / str(series_id)\n )\n\n if not series_path.exists():\n\n raise FileNotFoundError(\n f\"Missing test series:\\n\"\n f\"{series_path}\"\n )\n\n volume = load_dicom_volume(\n series_path\n )\n\n images = preprocess_volume(\n volume\n )\n\n images = images.repeat(\n 1, 3, 1, 1\n )\n\n images = images.to(\n DEVICE\n )\n\n images = (\n images - imagenet_mean\n ) / imagenet_std\n\n slice_features = backbone(\n images\n )\n\n study_feature = (\n slice_features\n .mean(dim=0)\n .cpu()\n )\n\n sequence_features.append(\n study_feature\n )\n\n availability.append(\n 1.0\n )\n\n feature_vector = torch.cat(\n sequence_features\n + [\n torch.tensor(\n availability,\n dtype=torch.float32\n )\n ]\n )\n\n test_embedding_rows.append(\n feature_vector.numpy()\n )\n\n if (\n (count + 1) % 25 == 0\n or count == len(test_ids) - 1\n ):\n\n elapsed = (\n time.time() - start\n )\n\n print(\n f\"Test embeddings: \"\n f\"{count + 1}/\"\n f\"{len(test_ids)} \"\n f\"| {elapsed:.1f}s\"\n )\n\nX_test = np.stack(\n test_embedding_rows\n).astype(\n np.float32\n)\n\nprint(\n \"\\nFinal test embedding shape:\",\n X_test.shape\n)\n\n\n# ============================================================\n# 10. Generate FINAL predictions\n# ============================================================\n\npredictions = {}\n\nfor target_idx, target in enumerate(\n TARGETS\n):\n\n probabilities = (\n final_models[target_idx]\n .predict_proba(\n X_test\n )[:, 1]\n )\n\n probabilities = np.clip(\n probabilities,\n 1e-6,\n 1 - 1e-6\n )\n\n predictions[target] = probabilities\n\n\nprediction_df = pd.DataFrame(\n predictions\n)\n\nprediction_df.insert(\n 0,\n \"StudyInstanceUID\",\n test_ids\n)\n\n\n# ============================================================\n# 11. Force exact sample-submission column order\n# ============================================================\n\nexpected_columns = list(\n sample_submission.columns\n)\n\nmissing_columns = [\n col\n for col in expected_columns\n if col not in prediction_df.columns\n]\n\nextra_columns = [\n col\n for col in prediction_df.columns\n if col not in expected_columns\n]\n\nif missing_columns:\n raise RuntimeError(\n f\"Missing submission columns: \"\n f\"{missing_columns}\"\n )\n\nif extra_columns:\n print(\n \"Extra columns removed:\",\n extra_columns\n )\n\nsubmission = prediction_df[\n expected_columns\n].copy()\n\n\n# ============================================================\n# 12. Final validation\n# ============================================================\n\nprint(\n \"\\n\"\n + \"=\" * 70\n)\n\nprint(\n \"SUBMISSION VALIDATION\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"Rows:\",\n len(submission)\n)\n\nprint(\n \"Expected rows:\",\n len(sample_submission)\n)\n\nprint(\n \"Columns:\"\n)\n\nprint(\n submission.columns.tolist()\n)\n\nprint(\n \"\\nMissing values:\",\n int(\n submission.isna().sum().sum()\n )\n)\n\nprediction_columns = [\n col\n for col in expected_columns\n if col != \"StudyInstanceUID\"\n]\n\nprint(\n \"Minimum prediction:\",\n float(\n submission[\n prediction_columns\n ].min().min()\n )\n)\n\nprint(\n \"Maximum prediction:\",\n float(\n submission[\n prediction_columns\n ].max().max()\n )\n)\n\nassert len(submission) == len(\n sample_submission\n)\n\nassert submission.columns.tolist() == (\n expected_columns\n)\n\nassert (\n submission.isna().sum().sum()\n == 0\n)\n\nassert (\n submission[prediction_columns]\n .applymap(\n lambda x: 0 <= x <= 1\n )\n .all()\n .all()\n)\n\nassert (\n submission[\"StudyInstanceUID\"]\n .tolist()\n ==\n sample_submission[\n \"StudyInstanceUID\"\n ].tolist()\n)\n\n# Save\nsubmission_path = (\n OUTPUT_DIR / \"submission.csv\"\n)\n\nsubmission.to_csv(\n submission_path,\n index=False\n)\n\nprint(\n \"\\n\"\n + \"=\" * 70\n)\n\nprint(\n \"SUBMISSION READY\"\n)\n\nprint(\n \"=\" * 70\n)\n\nprint(\n \"File:\",\n submission_path\n)\n\nprint(\n \"\\nPreview:\"\n)\n\ndisplay(\n submission.head()\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T18:34:30.241674Z","iopub.execute_input":"2026-08-17T18:34:30.242743Z","iopub.status.idle":"2026-08-17T18:38:02.833733Z","shell.execute_reply.started":"2026-08-17T18:34:30.242708Z","shell.execute_reply":"2026-08-17T18:38:02.83225Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# FINAL KAGGLE SUBMISSION FILE\n# Save directly in /kaggle/working/\n\nsubmission_path = \"/kaggle/working/submission.csv\"\n\nsubmission.to_csv(\n submission_path,\n index=False\n)\n\nprint(\"FINAL FILE:\", submission_path)\nprint(\"EXISTS:\", os.path.exists(submission_path))\nprint(\"SIZE:\", os.path.getsize(submission_path), \"bytes\")\n\nprint(\"\\nSubmission preview:\")\ndisplay(submission.head())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-08-17T19:03:52.92279Z","iopub.execute_input":"2026-08-17T19:03:52.923823Z","iopub.status.idle":"2026-08-17T19:03:52.959549Z","shell.execute_reply.started":"2026-08-17T19:03:52.923788Z","shell.execute_reply":"2026-08-17T19:03:52.958473Z"}},"outputs":[],"execution_count":null}]}