{"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":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n for filename in filenames:\n print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session\n\n# Use the kagglehub client library to attach Kaggle resources like competitions, datasets, and models to your session\n# Learn more about kagglehub: https://github.com/Kaggle/kagglehub/blob/main/README.md\n\nimport kagglehub\n# kagglehub.dataset_download('/')","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:04.08729Z","iopub.execute_input":"2026-09-26T01:55:04.087881Z","iopub.status.idle":"2026-09-26T01:55:10.131846Z","shell.execute_reply.started":"2026-09-26T01:55:04.087851Z","shell.execute_reply":"2026-09-26T01:55:10.131181Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import torch\n\nprint(\"PyTorch version:\", torch.__version__)\nprint(\"CUDA available:\", torch.cuda.is_available())\n\nif torch.cuda.is_available():\n print(\"GPU:\", torch.cuda.get_device_name(0))\n print(\"GPU count:\", torch.cuda.device_count())\nelse:\n print(\"❌ GPU NOT AVAILABLE\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.133255Z","iopub.execute_input":"2026-09-26T01:55:10.133598Z","iopub.status.idle":"2026-09-26T01:55:10.13856Z","shell.execute_reply.started":"2026-09-26T01:55:10.133578Z","shell.execute_reply":"2026-09-26T01:55:10.13774Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nprint(os.listdir(\"/kaggle/input\"))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.139458Z","iopub.execute_input":"2026-09-26T01:55:10.13997Z","iopub.status.idle":"2026-09-26T01:55:10.161004Z","shell.execute_reply.started":"2026-09-26T01:55:10.139946Z","shell.execute_reply":"2026-09-26T01:55:10.160281Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!ls -lah /kaggle/input/competitions","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.161823Z","iopub.execute_input":"2026-09-26T01:55:10.162002Z","iopub.status.idle":"2026-09-26T01:55:10.306625Z","shell.execute_reply.started":"2026-09-26T01:55:10.161984Z","shell.execute_reply":"2026-09-26T01:55:10.305941Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!find /kaggle/input/competitions -maxdepth 2 -type d | head -30","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.309003Z","iopub.execute_input":"2026-09-26T01:55:10.309298Z","iopub.status.idle":"2026-09-26T01:55:10.547627Z","shell.execute_reply.started":"2026-09-26T01:55:10.309268Z","shell.execute_reply":"2026-09-26T01:55:10.546991Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"dataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.548828Z","iopub.execute_input":"2026-09-26T01:55:10.549151Z","iopub.status.idle":"2026-09-26T01:55:10.5534Z","shell.execute_reply.started":"2026-09-26T01:55:10.549092Z","shell.execute_reply":"2026-09-26T01:55:10.552535Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nprint(\"Dataset exists:\", os.path.exists(dataset_root))\nprint(\"\\nContents:\")\nprint(os.listdir(dataset_root))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.554384Z","iopub.execute_input":"2026-09-26T01:55:10.554672Z","iopub.status.idle":"2026-09-26T01:55:10.569482Z","shell.execute_reply.started":"2026-09-26T01:55:10.55465Z","shell.execute_reply":"2026-09-26T01:55:10.568636Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\n\ntrain_csv = os.path.join(dataset_root, \"train.csv\")\n\ntrain_df = pd.read_csv(train_csv)\n\nprint(\"Train shape:\", train_df.shape)\nprint(\"\\nColumns:\")\nprint(train_df.columns.tolist())\n\nprint(\"\\nFirst 5 rows:\")\ndisplay(train_df.head())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.570416Z","iopub.execute_input":"2026-09-26T01:55:10.570685Z","iopub.status.idle":"2026-09-26T01:55:10.619189Z","shell.execute_reply.started":"2026-09-26T01:55:10.570661Z","shell.execute_reply":"2026-09-26T01:55:10.618541Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport glob\nimport pandas as pd\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\n# Load training metadata\ntrain_df = pd.read_csv(os.path.join(dataset_root, \"train.csv\"))\n\n# Select the first actual WSI (not TMA)\nwsi_row = train_df[train_df[\"is_tma\"] == False].iloc[0]\n\nimage_id = wsi_row[\"image_id\"]\n\nprint(\"Selected image ID:\", image_id)\nprint(\"Label:\", wsi_row[\"label\"])\nprint(\"Is TMA:\", wsi_row[\"is_tma\"])\nprint(\"Dimensions:\", wsi_row[\"image_width\"], \"x\", wsi_row[\"image_height\"])\n\n# Find the actual image file\npattern = os.path.join(\n dataset_root,\n \"train_images\",\n f\"{image_id}.*\"\n)\n\nmatches = glob.glob(pattern)\n\nprint(\"\\nMatching files:\")\nprint(matches)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.620162Z","iopub.execute_input":"2026-09-26T01:55:10.620458Z","iopub.status.idle":"2026-09-26T01:55:10.631534Z","shell.execute_reply.started":"2026-09-26T01:55:10.620437Z","shell.execute_reply":"2026-09-26T01:55:10.630576Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nthumbnail_dir = \"/kaggle/input/competitions/UBC-OCEAN/train_thumbnails\"\n\nfiles = os.listdir(thumbnail_dir)\n\nprint(\"Number of thumbnails:\", len(files))\nprint(\"First 20 files:\")\nprint(files[:20])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.63255Z","iopub.execute_input":"2026-09-26T01:55:10.632924Z","iopub.status.idle":"2026-09-26T01:55:10.648989Z","shell.execute_reply.started":"2026-09-26T01:55:10.632885Z","shell.execute_reply":"2026-09-26T01:55:10.64829Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"matches = [f for f in files if f.startswith(\"4\")]\n\nprint(\"Files starting with 4:\")\nprint(matches)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.649868Z","iopub.execute_input":"2026-09-26T01:55:10.650284Z","iopub.status.idle":"2026-09-26T01:55:10.663316Z","shell.execute_reply.started":"2026-09-26T01:55:10.650261Z","shell.execute_reply":"2026-09-26T01:55:10.662555Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nthumbnail_path = os.path.join(\n dataset_root,\n \"train_thumbnails\",\n f\"{image_id}_thumbnail.png\"\n)\n\nprint(\"Thumbnail path:\", thumbnail_path)\nprint(\"File exists:\", os.path.exists(thumbnail_path))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.66425Z","iopub.execute_input":"2026-09-26T01:55:10.664536Z","iopub.status.idle":"2026-09-26T01:55:10.680925Z","shell.execute_reply.started":"2026-09-26T01:55:10.664504Z","shell.execute_reply":"2026-09-26T01:55:10.680278Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from PIL import Image\nimport matplotlib.pyplot as plt\n\nthumbnail = Image.open(thumbnail_path)\n\nprint(\"Thumbnail size:\", thumbnail.size)\n\nplt.figure(figsize=(12, 8))\nplt.imshow(thumbnail)\nplt.axis(\"off\")\nplt.title(f\"WSI Thumbnail - Image ID {image_id} - {wsi_row['label']}\")\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:10.68178Z","iopub.execute_input":"2026-09-26T01:55:10.682045Z","iopub.status.idle":"2026-09-26T01:55:11.690339Z","shell.execute_reply.started":"2026-09-26T01:55:10.682014Z","shell.execute_reply":"2026-09-26T01:55:11.689457Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from PIL import Image\n\n# Temporarily disable Pillow's pixel-count safety check\n# ONLY for opening/inspecting this trusted competition image.\nImage.MAX_IMAGE_PIXELS = None\n\nwith Image.open(wsi_path) as wsi:\n print(\"WSI opened successfully!\")\n print(\"Format:\", wsi.format)\n print(\"Size:\", wsi.size)\n print(\"Mode:\", wsi.mode)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.69306Z","iopub.execute_input":"2026-09-26T01:55:11.69383Z","iopub.status.idle":"2026-09-26T01:55:11.704028Z","shell.execute_reply.started":"2026-09-26T01:55:11.693803Z","shell.execute_reply":"2026-09-26T01:55:11.699936Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\nimport pandas as pd\n\n\nclass UBCOceanMetadata:\n \"\"\"\n Handles metadata and file paths for the UBC-OCEAN dataset.\n \"\"\"\n\n def __init__(self, dataset_root: str):\n self.dataset_root = Path(dataset_root)\n\n self.train_csv = self.dataset_root / \"train.csv\"\n self.train_images_dir = self.dataset_root / \"train_images\"\n self.train_thumbnails_dir = self.dataset_root / \"train_thumbnails\"\n\n if not self.dataset_root.exists():\n raise FileNotFoundError(\n f\"Dataset root not found: {self.dataset_root}\"\n )\n\n if not self.train_csv.exists():\n raise FileNotFoundError(\n f\"train.csv not found: {self.train_csv}\"\n )\n\n self.df = pd.read_csv(self.train_csv)\n\n self._validate_columns()\n\n def _validate_columns(self):\n required_columns = {\n \"image_id\",\n \"label\",\n \"image_width\",\n \"image_height\",\n \"is_tma\",\n }\n\n missing = required_columns - set(self.df.columns)\n\n if missing:\n raise ValueError(\n f\"Missing required columns: {sorted(missing)}\"\n )\n\n def get_record(self, image_id):\n \"\"\"Return metadata for one image.\"\"\"\n\n rows = self.df[self.df[\"image_id\"] == image_id]\n\n if rows.empty:\n raise ValueError(\n f\"Image ID {image_id} not found in train.csv\"\n )\n\n row = rows.iloc[0]\n\n return {\n \"image_id\": int(row[\"image_id\"]),\n \"label\": str(row[\"label\"]),\n \"width\": int(row[\"image_width\"]),\n \"height\": int(row[\"image_height\"]),\n \"is_tma\": bool(row[\"is_tma\"]),\n }\n\n def get_image_path(self, image_id):\n \"\"\"Return path to the original training image.\"\"\"\n\n return self.train_images_dir / f\"{image_id}.png\"\n\n def get_thumbnail_path(self, image_id):\n \"\"\"\n Return thumbnail path for a WSI.\n\n TMAs do not have thumbnails.\n \"\"\"\n\n record = self.get_record(image_id)\n\n if record[\"is_tma\"]:\n return None\n\n return (\n self.train_thumbnails_dir\n / f\"{image_id}_thumbnail.png\"\n )\n\n def get_all_records(self):\n \"\"\"Return metadata for all training images.\"\"\"\n\n records = []\n\n for _, row in self.df.iterrows():\n image_id = int(row[\"image_id\"])\n\n records.append(\n {\n \"image_id\": image_id,\n \"label\": str(row[\"label\"]),\n \"width\": int(row[\"image_width\"]),\n \"height\": int(row[\"image_height\"]),\n \"is_tma\": bool(row[\"is_tma\"]),\n }\n )\n\n return records","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.704728Z","iopub.status.idle":"2026-09-26T01:55:11.705035Z","shell.execute_reply.started":"2026-09-26T01:55:11.704911Z","shell.execute_reply":"2026-09-26T01:55:11.704926Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nprint(\"Current directory:\")\nprint(os.getcwd())\n\nprint(\"\\nContents of /kaggle/working:\")\nprint(os.listdir(\"/kaggle/working\"))\n\nprint(\"\\nSearching for metadata.py:\")\nfor root, dirs, files in os.walk(\"/kaggle/working\"):\n if \"metadata.py\" in files:\n print(os.path.join(root, \"metadata.py\"))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.706194Z","iopub.status.idle":"2026-09-26T01:55:11.706608Z","shell.execute_reply.started":"2026-09-26T01:55:11.706408Z","shell.execute_reply":"2026-09-26T01:55:11.706434Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\n\nfolders = [\n \"config\",\n \"src\",\n \"src/preprocessing\",\n \"src/quality\",\n \"src/features\",\n \"src/classification\",\n \"src/calibration\",\n \"src/explainability\",\n \"src/routing\",\n \"src/evaluation\",\n \"cache/tiles\",\n \"cache/embeddings\",\n \"outputs/logs\",\n \"outputs/metrics\",\n \"outputs/visualizations\",\n \"tests\",\n \"notebooks\",\n]\n\nfor folder in folders:\n (project_root / folder).mkdir(parents=True, exist_ok=True)\n\n# Create Python package marker files\npackage_files = [\n \"src/__init__.py\",\n \"src/preprocessing/__init__.py\",\n \"src/quality/__init__.py\",\n \"src/features/__init__.py\",\n \"src/classification/__init__.py\",\n \"src/calibration/__init__.py\",\n \"src/explainability/__init__.py\",\n \"src/routing/__init__.py\",\n \"src/evaluation/__init__.py\",\n]\n\nfor file in package_files:\n (project_root / file).touch()\n\nprint(\"Project structure created!\")\nprint(project_root)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.707658Z","iopub.status.idle":"2026-09-26T01:55:11.707892Z","shell.execute_reply.started":"2026-09-26T01:55:11.707778Z","shell.execute_reply":"2026-09-26T01:55:11.707792Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nfor root, dirs, files in os.walk(\"/kaggle/working/ovarian-trustworthy-ai\"):\n level = root.replace(\"/kaggle/working/ovarian-trustworthy-ai\", \"\").count(os.sep)\n indent = \" \" * level\n print(f\"{indent}{os.path.basename(root)}/\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.709606Z","iopub.status.idle":"2026-09-26T01:55:11.710047Z","shell.execute_reply.started":"2026-09-26T01:55:11.7098Z","shell.execute_reply":"2026-09-26T01:55:11.709821Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nmetadata_code = r'''\nfrom pathlib import Path\nimport pandas as pd\n\n\nclass UBCOceanMetadata:\n \"\"\"\n Handles metadata and file paths for the UBC-OCEAN dataset.\n \"\"\"\n\n def __init__(self, dataset_root: str):\n self.dataset_root = Path(dataset_root)\n\n self.train_csv = self.dataset_root / \"train.csv\"\n self.train_images_dir = self.dataset_root / \"train_images\"\n self.train_thumbnails_dir = self.dataset_root / \"train_thumbnails\"\n\n if not self.dataset_root.exists():\n raise FileNotFoundError(\n f\"Dataset root not found: {self.dataset_root}\"\n )\n\n if not self.train_csv.exists():\n raise FileNotFoundError(\n f\"train.csv not found: {self.train_csv}\"\n )\n\n self.df = pd.read_csv(self.train_csv)\n\n self._validate_columns()\n\n def _validate_columns(self):\n required_columns = {\n \"image_id\",\n \"label\",\n \"image_width\",\n \"image_height\",\n \"is_tma\",\n }\n\n missing = required_columns - set(self.df.columns)\n\n if missing:\n raise ValueError(\n f\"Missing required columns: {sorted(missing)}\"\n )\n\n def get_record(self, image_id):\n \"\"\"Return metadata for one image.\"\"\"\n\n rows = self.df[self.df[\"image_id\"] == image_id]\n\n if rows.empty:\n raise ValueError(\n f\"Image ID {image_id} not found in train.csv\"\n )\n\n row = rows.iloc[0]\n\n return {\n \"image_id\": int(row[\"image_id\"]),\n \"label\": str(row[\"label\"]),\n \"width\": int(row[\"image_width\"]),\n \"height\": int(row[\"image_height\"]),\n \"is_tma\": bool(row[\"is_tma\"]),\n }\n\n def get_image_path(self, image_id):\n \"\"\"Return path to the original training image.\"\"\"\n\n return self.train_images_dir / f\"{image_id}.png\"\n\n def get_thumbnail_path(self, image_id):\n \"\"\"\n Return thumbnail path for a WSI.\n\n TMAs do not have thumbnails.\n \"\"\"\n\n record = self.get_record(image_id)\n\n if record[\"is_tma\"]:\n return None\n\n return (\n self.train_thumbnails_dir\n / f\"{image_id}_thumbnail.png\"\n )\n\n def get_all_records(self):\n \"\"\"Return metadata for all training images.\"\"\"\n\n records = []\n\n for _, row in self.df.iterrows():\n image_id = int(row[\"image_id\"])\n\n records.append(\n {\n \"image_id\": image_id,\n \"label\": str(row[\"label\"]),\n \"width\": int(row[\"image_width\"]),\n \"height\": int(row[\"image_height\"]),\n \"is_tma\": bool(row[\"is_tma\"]),\n }\n )\n\n return records\n'''\n\nmetadata_path = (\n Path(\"/kaggle/working/ovarian-trustworthy-ai\")\n / \"src\"\n / \"preprocessing\"\n / \"metadata.py\"\n)\n\nmetadata_path.write_text(metadata_code)\n\nprint(\"Created:\")\nprint(metadata_path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.711555Z","iopub.status.idle":"2026-09-26T01:55:11.712004Z","shell.execute_reply.started":"2026-09-26T01:55:11.711817Z","shell.execute_reply":"2026-09-26T01:55:11.711841Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\nimport os\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\n\nprint(\"Project exists:\", project_root.exists())\nprint(\"Project path:\", project_root)\n\nprint(\"\\nProject contents:\")\nif project_root.exists():\n for path in project_root.rglob(\"*\"):\n print(path)\nelse:\n print(\"PROJECT DOES NOT EXIST\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.713576Z","iopub.status.idle":"2026-09-26T01:55:11.713879Z","shell.execute_reply.started":"2026-09-26T01:55:11.71374Z","shell.execute_reply":"2026-09-26T01:55:11.713761Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nfrom pathlib import Path\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\n\nprint(\"Project exists:\", project_root.exists())\nprint(\"src exists:\", (project_root / \"src\").exists())\nprint(\"metadata.py exists:\", (project_root / \"src/preprocessing/metadata.py\").exists())\n\n# Add the project root to Python's import path\nsys.path.insert(0, str(project_root))\n\nprint(\"\\nProject root added to sys.path:\")\nprint(str(project_root) in sys.path)\n\nprint(\"\\nFirst entries in sys.path:\")\nprint(sys.path[:5])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.714916Z","iopub.status.idle":"2026-09-26T01:55:11.715309Z","shell.execute_reply.started":"2026-09-26T01:55:11.715094Z","shell.execute_reply":"2026-09-26T01:55:11.715139Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import importlib.util\n\nspec = importlib.util.find_spec(\"src\")\n\nprint(\"\\nPython can find src:\", spec is not None)\n\nif spec is not None:\n print(\"src location:\", spec.origin)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.716592Z","iopub.status.idle":"2026-09-26T01:55:11.716952Z","shell.execute_reply.started":"2026-09-26T01:55:11.716788Z","shell.execute_reply":"2026-09-26T01:55:11.716817Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nfrom pathlib import Path\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nprint(\"1. Project root exists:\")\nprint(Path(project_root).exists())\n\nprint(\"\\n2. src exists:\")\nprint(Path(project_root, \"src\").exists())\n\nprint(\"\\n3. src __init__.py exists:\")\nprint(Path(project_root, \"src\", \"__init__.py\").exists())\n\nprint(\"\\n4. metadata.py exists:\")\nprint(Path(project_root, \"src\", \"preprocessing\", \"metadata.py\").exists())\n\nprint(\"\\n5. Project root in sys.path BEFORE:\")\nprint(project_root in sys.path)\n\nsys.path.insert(0, project_root)\n\nprint(\"\\n6. Project root in sys.path AFTER:\")\nprint(project_root in sys.path)\n\nprint(\"\\n7. sys.path first 10 entries:\")\nfor p in sys.path[:10]:\n print(repr(p))\n\nprint(\"\\n8. Importlib caches invalidated\")\nimportlib.invalidate_caches()\n\nprint(\"\\n9. Find src:\")\nprint(importlib.util.find_spec(\"src\"))\n\nprint(\"\\n10. Find metadata directly:\")\nmetadata_file = Path(project_root) / \"src\" / \"preprocessing\" / \"metadata.py\"\nprint(metadata_file.exists())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.718047Z","iopub.status.idle":"2026-09-26T01:55:11.718542Z","shell.execute_reply.started":"2026-09-26T01:55:11.718309Z","shell.execute_reply":"2026-09-26T01:55:11.718331Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from src.preprocessing.metadata import UBCOceanMetadata\n\nprint(\"Import successful! ✅\")\nprint(UBCOceanMetadata)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.71967Z","iopub.status.idle":"2026-09-26T01:55:11.720506Z","shell.execute_reply.started":"2026-09-26T01:55:11.720326Z","shell.execute_reply":"2026-09-26T01:55:11.720346Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"dataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nmetadata = UBCOceanMetadata(dataset_root)\n\nprint(\"Metadata loaded successfully!\")\nprint(\"Number of training records:\", len(metadata.df))\n\nrecord = metadata.get_record(4)\n\nprint(\"\\nImage 4 metadata:\")\nfor key, value in record.items():\n print(f\"{key}: {value}\")\n\nprint(\"\\nOriginal image:\")\nprint(metadata.get_image_path(4))\n\nprint(\"\\nThumbnail:\")\nprint(metadata.get_thumbnail_path(4))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.72236Z","iopub.status.idle":"2026-09-26T01:55:11.722613Z","shell.execute_reply.started":"2026-09-26T01:55:11.7225Z","shell.execute_reply":"2026-09-26T01:55:11.722514Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pyvips\n\nprint(\"pyvips imported successfully! ✅\")\nprint(\"libvips version:\", pyvips.version(0), \n pyvips.version(1), \n pyvips.version(2))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.724292Z","iopub.status.idle":"2026-09-26T01:55:11.724664Z","shell.execute_reply.started":"2026-09-26T01:55:11.724467Z","shell.execute_reply":"2026-09-26T01:55:11.724488Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!pip install -q \"pyvips[binary]\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.725544Z","iopub.status.idle":"2026-09-26T01:55:11.7259Z","shell.execute_reply.started":"2026-09-26T01:55:11.725722Z","shell.execute_reply":"2026-09-26T01:55:11.725743Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pyvips\n\nprint(\"pyvips imported successfully! ✅\")\nprint(\n \"libvips version:\",\n pyvips.version(0),\n pyvips.version(1),\n pyvips.version(2)\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.727413Z","iopub.status.idle":"2026-09-26T01:55:11.727759Z","shell.execute_reply.started":"2026-09-26T01:55:11.72759Z","shell.execute_reply":"2026-09-26T01:55:11.727612Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\nfile_path = project_root / \"src/preprocessing/image_reader.py\"\n\ncode = '''\nfrom pathlib import Path\nimport numpy as np\nimport pyvips\n\nfrom .metadata import UBCOceanMetadata\n\n\nclass UBCOceanImageReader:\n \"\"\"\n Safely opens UBC-OCEAN images and reads small image regions.\n\n Large WSI files are never loaded completely into NumPy.\n \"\"\"\n\n def __init__(self, dataset_root: str):\n self.metadata = UBCOceanMetadata(dataset_root)\n\n def open_image(self, image_id: int):\n \"\"\"\n Open an image using pyvips.\n\n Only image metadata/header information is accessed initially.\n \"\"\"\n image_path = self.metadata.get_image_path(image_id)\n\n if not image_path.exists():\n raise FileNotFoundError(\n f\"Image file not found: {image_path}\"\n )\n\n image = pyvips.Image.new_from_file(\n str(image_path),\n access=\"random\"\n )\n\n return image\n\n def get_region(\n self,\n image_id: int,\n x: int,\n y: int,\n width: int,\n height: int\n ) -> np.ndarray:\n \"\"\"\n Extract a small region from an image and return it as NumPy RGB.\n\n The full WSI is never converted to NumPy.\n \"\"\"\n\n image = self.open_image(image_id)\n\n if x < 0 or y < 0:\n raise ValueError(\"x and y must be non-negative.\")\n\n if width <= 0 or height <= 0:\n raise ValueError(\"width and height must be positive.\")\n\n if x >= image.width or y >= image.height:\n raise ValueError(\n f\"Region origin ({x}, {y}) is outside \"\n f\"image dimensions ({image.width}, {image.height}).\"\n )\n\n width = min(width, image.width - x)\n height = min(height, image.height - y)\n\n region = image.crop(x, y, width, height)\n\n array = np.ndarray(\n buffer=region.write_to_memory(),\n dtype=np.uint8,\n shape=(region.height, region.width, region.bands)\n )\n\n if array.shape[2] >= 3:\n array = array[:, :, :3]\n\n return array\n\n def get_image_dimensions(self, image_id: int):\n \"\"\"\n Return actual image dimensions from pyvips.\n \"\"\"\n image = self.open_image(image_id)\n return image.width, image.height\n'''\n\nfile_path.write_text(code)\n\nprint(f\"Created: {file_path}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.728852Z","iopub.status.idle":"2026-09-26T01:55:11.729231Z","shell.execute_reply.started":"2026-09-26T01:55:11.729022Z","shell.execute_reply":"2026-09-26T01:55:11.729053Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.image_reader import UBCOceanImageReader\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nreader = UBCOceanImageReader(dataset_root)\n\nprint(\"Image reader imported successfully! ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.730328Z","iopub.status.idle":"2026-09-26T01:55:11.730621Z","shell.execute_reply.started":"2026-09-26T01:55:11.730499Z","shell.execute_reply":"2026-09-26T01:55:11.730513Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image = reader.open_image(4)\n\nprint(\"Image opened successfully! ✅\")\nprint(\"Width :\", image.width)\nprint(\"Height:\", image.height)\nprint(\"Bands :\", image.bands)\nprint(\"Format:\", image.format)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.732238Z","iopub.status.idle":"2026-09-26T01:55:11.732589Z","shell.execute_reply.started":"2026-09-26T01:55:11.732409Z","shell.execute_reply":"2026-09-26T01:55:11.732432Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tile = reader.get_region(\n image_id=4,\n x=5000,\n y=5000,\n width=256,\n height=256\n)\n\nprint(\"Region extracted successfully! ✅\")\nprint(\"Shape:\", tile.shape)\nprint(\"Data type:\", tile.dtype)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.734185Z","iopub.status.idle":"2026-09-26T01:55:11.734496Z","shell.execute_reply.started":"2026-09-26T01:55:11.734336Z","shell.execute_reply":"2026-09-26T01:55:11.734356Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import matplotlib.pyplot as plt\n\nplt.figure(figsize=(5, 5))\nplt.imshow(tile)\nplt.axis(\"off\")\nplt.title(\"WSI 4 — 256 × 256 Region\")\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.735804Z","iopub.status.idle":"2026-09-26T01:55:11.736155Z","shell.execute_reply.started":"2026-09-26T01:55:11.735955Z","shell.execute_reply":"2026-09-26T01:55:11.735975Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\nfile_path = project_root / \"src/preprocessing/slide_type.py\"\n\ncode = '''\nfrom dataclasses import dataclass\n\n\n@dataclass\nclass SlideInfo:\n image_id: int\n slide_type: str\n width: int\n height: int\n magnification: str\n\n\nclass SlideTypeHandler:\n \"\"\"\n Determines how a UBC-OCEAN image should be handled\n based on the official train.csv metadata.\n \"\"\"\n\n def __init__(self, metadata):\n self.metadata = metadata\n\n def get_slide_info(self, image_id: int) -> SlideInfo:\n\n record = self.metadata.get_record(image_id)\n\n if record[\"is_tma\"]:\n slide_type = \"TMA\"\n magnification = \"40x\"\n else:\n slide_type = \"WSI\"\n magnification = \"20x\"\n\n return SlideInfo(\n image_id=record[\"image_id\"],\n slide_type=slide_type,\n width=record[\"width\"],\n height=record[\"height\"],\n magnification=magnification,\n )\n\n def is_wsi(self, image_id: int) -> bool:\n return not self.metadata.get_record(image_id)[\"is_tma\"]\n\n def is_tma(self, image_id: int) -> bool:\n return self.metadata.get_record(image_id)[\"is_tma\"]\n'''\n\nfile_path.write_text(code)\n\nprint(f\"Created: {file_path}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.739013Z","iopub.status.idle":"2026-09-26T01:55:11.739472Z","shell.execute_reply.started":"2026-09-26T01:55:11.739331Z","shell.execute_reply":"2026-09-26T01:55:11.739353Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.metadata import UBCOceanMetadata\nfrom src.preprocessing.slide_type import SlideTypeHandler\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nmetadata = UBCOceanMetadata(dataset_root)\nslide_handler = SlideTypeHandler(metadata)\n\ninfo = slide_handler.get_slide_info(4)\n\nprint(info)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.740803Z","iopub.status.idle":"2026-09-26T01:55:11.741218Z","shell.execute_reply.started":"2026-09-26T01:55:11.741032Z","shell.execute_reply":"2026-09-26T01:55:11.741053Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"info_tma = slide_handler.get_slide_info(91)\n\nprint(info_tma)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.742449Z","iopub.status.idle":"2026-09-26T01:55:11.742823Z","shell.execute_reply.started":"2026-09-26T01:55:11.742675Z","shell.execute_reply":"2026-09-26T01:55:11.7427Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\nfile_path = project_root / \"src/preprocessing/tile_extractor.py\"\n\ncode = '''\nfrom dataclasses import dataclass\nfrom typing import List\n\nimport cv2\nimport numpy as np\n\nfrom .image_reader import UBCOceanImageReader\nfrom .slide_type import SlideTypeHandler\n\n\n@dataclass\nclass TileCoordinate:\n image_id: int\n x: int\n y: int\n source_size: int\n standardized_size: int\n slide_type: str\n\n\nclass TileExtractor:\n \"\"\"\n Extracts standardized tiles from UBC-OCEAN WSI and TMA images.\n\n WSI:\n 256 x 256 native tiles at approximately 20x.\n\n TMA:\n 512 x 512 native tiles at approximately 40x,\n resized to 256 x 256 for standardization.\n \"\"\"\n\n def __init__(self, image_reader, slide_handler):\n self.image_reader = image_reader\n self.slide_handler = slide_handler\n\n def get_tile_size(self, image_id: int):\n \"\"\"\n Return the native tile size according to slide type.\n \"\"\"\n\n if self.slide_handler.is_tma(image_id):\n return 512\n\n return 256\n\n def generate_grid(\n self,\n image_id: int,\n stride: int = None\n ) -> List[TileCoordinate]:\n \"\"\"\n Generate candidate tile coordinates.\n\n This function only generates coordinates.\n It does NOT read the image.\n \"\"\"\n\n info = self.slide_handler.get_slide_info(image_id)\n\n tile_size = self.get_tile_size(image_id)\n\n if stride is None:\n stride = tile_size\n\n coordinates = []\n\n for y in range(0, info.height - tile_size + 1, stride):\n for x in range(0, info.width - tile_size + 1, stride):\n\n coordinates.append(\n TileCoordinate(\n image_id=image_id,\n x=x,\n y=y,\n source_size=tile_size,\n standardized_size=256,\n slide_type=info.slide_type,\n )\n )\n\n return coordinates\n\n def extract_tile(\n self,\n coordinate: TileCoordinate\n ) -> np.ndarray:\n \"\"\"\n Extract one tile and convert it to the standardized\n 256 x 256 RGB representation.\n \"\"\"\n\n tile = self.image_reader.get_region(\n image_id=coordinate.image_id,\n x=coordinate.x,\n y=coordinate.y,\n width=coordinate.source_size,\n height=coordinate.source_size,\n )\n\n if coordinate.source_size != coordinate.standardized_size:\n\n tile = cv2.resize(\n tile,\n (\n coordinate.standardized_size,\n coordinate.standardized_size\n ),\n interpolation=cv2.INTER_AREA,\n )\n\n return tile\n\n def extract_tiles(\n self,\n image_id: int,\n max_tiles: int = None\n ) -> List[np.ndarray]:\n \"\"\"\n Extract tiles from a regular grid.\n\n max_tiles can be used to prevent excessive memory usage.\n \"\"\"\n\n coordinates = self.generate_grid(image_id)\n\n if max_tiles is not None:\n coordinates = coordinates[:max_tiles]\n\n tiles = []\n\n for coordinate in coordinates:\n\n tile = self.extract_tile(coordinate)\n\n tiles.append(tile)\n\n return tiles\n'''\n\nfile_path.write_text(code)\n\nprint(f\"Created: {file_path}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.744286Z","iopub.status.idle":"2026-09-26T01:55:11.744622Z","shell.execute_reply.started":"2026-09-26T01:55:11.74451Z","shell.execute_reply":"2026-09-26T01:55:11.744524Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.tile_extractor import TileExtractor\n\nprint(\"Tile extractor imported successfully! ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.746174Z","iopub.status.idle":"2026-09-26T01:55:11.746524Z","shell.execute_reply.started":"2026-09-26T01:55:11.746336Z","shell.execute_reply":"2026-09-26T01:55:11.746358Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from src.preprocessing.metadata import UBCOceanMetadata\nfrom src.preprocessing.image_reader import UBCOceanImageReader\nfrom src.preprocessing.slide_type import SlideTypeHandler\nfrom src.preprocessing.tile_extractor import TileExtractor\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nmetadata = UBCOceanMetadata(dataset_root)\n\nreader = UBCOceanImageReader(dataset_root)\n\nslide_handler = SlideTypeHandler(metadata)\n\ntile_extractor = TileExtractor(\n image_reader=reader,\n slide_handler=slide_handler\n)\n\nprint(\"Tile extraction pipeline initialized! ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.747372Z","iopub.status.idle":"2026-09-26T01:55:11.747617Z","shell.execute_reply.started":"2026-09-26T01:55:11.74751Z","shell.execute_reply":"2026-09-26T01:55:11.747524Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wsi_coords = tile_extractor.generate_grid(\n image_id=4,\n stride=256\n)\n\nprint(\"Number of candidate WSI tiles:\", len(wsi_coords))\nprint(\"First 5 coordinates:\")\n\nfor coord in wsi_coords[:5]:\n print(coord)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.748539Z","iopub.status.idle":"2026-09-26T01:55:11.748756Z","shell.execute_reply.started":"2026-09-26T01:55:11.748654Z","shell.execute_reply":"2026-09-26T01:55:11.748667Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wsi_tiles = []\n\nfor coord in wsi_coords[:3]:\n tile = tile_extractor.extract_tile(coord)\n wsi_tiles.append(tile)\n\nprint(\"Extracted:\", len(wsi_tiles), \"tiles\")\n\nfor i, tile in enumerate(wsi_tiles):\n print(f\"Tile {i}: shape={tile.shape}, dtype={tile.dtype}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.749944Z","iopub.status.idle":"2026-09-26T01:55:11.750226Z","shell.execute_reply.started":"2026-09-26T01:55:11.750057Z","shell.execute_reply":"2026-09-26T01:55:11.75007Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wsi_tiles = []\n\nfor coord in wsi_coords[:3]:\n tile = tile_extractor.extract_tile(coord)\n wsi_tiles.append(tile)\n\nprint(\"Extracted:\", len(wsi_tiles), \"tiles\")\n\nfor i, tile in enumerate(wsi_tiles):\n print(f\"Tile {i}: shape={tile.shape}, dtype={tile.dtype}\")\n ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.751777Z","iopub.status.idle":"2026-09-26T01:55:11.75205Z","shell.execute_reply.started":"2026-09-26T01:55:11.751945Z","shell.execute_reply":"2026-09-26T01:55:11.751959Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tma_coords = tile_extractor.generate_grid(\n image_id=91,\n stride=512\n)\n\nprint(\"Number of candidate TMA tiles:\", len(tma_coords))\n\nprint(\"First 3 coordinates:\")\n\nfor coord in tma_coords[:3]:\n print(coord)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.753506Z","iopub.status.idle":"2026-09-26T01:55:11.753846Z","shell.execute_reply.started":"2026-09-26T01:55:11.7537Z","shell.execute_reply":"2026-09-26T01:55:11.753714Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tma_tile = tile_extractor.extract_tile(tma_coords[0])\n\nprint(\"TMA standardized tile:\")\nprint(\"Shape:\", tma_tile.shape)\nprint(\"Dtype:\", tma_tile.dtype)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.75478Z","iopub.status.idle":"2026-09-26T01:55:11.755076Z","shell.execute_reply.started":"2026-09-26T01:55:11.754964Z","shell.execute_reply":"2026-09-26T01:55:11.754978Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\nfile_path = project_root / \"src/preprocessing/thumbnail_selector.py\"\n\ncode = '''\nfrom dataclasses import replace\n\nimport cv2\nimport numpy as np\nimport pyvips\n\n\nclass ThumbnailCandidateSelector:\n \"\"\"\n Uses a WSI thumbnail to identify candidate tissue regions.\n\n This is a coarse candidate-selection step.\n Detailed adaptive quality assessment is performed later in M2.\n \"\"\"\n\n def __init__(self, metadata):\n self.metadata = metadata\n\n def load_thumbnail(self, image_id: int) -> np.ndarray:\n \"\"\"\n Load the WSI thumbnail as an RGB NumPy array.\n\n TMAs do not have official thumbnails in UBC-OCEAN,\n so this method is intended for WSIs only.\n \"\"\"\n\n thumbnail_path = self.metadata.get_thumbnail_path(image_id)\n\n if thumbnail_path is None:\n raise ValueError(\n f\"Image {image_id} is a TMA and has no official thumbnail.\"\n )\n\n if not thumbnail_path.exists():\n raise FileNotFoundError(\n f\"Thumbnail not found: {thumbnail_path}\"\n )\n\n thumbnail = pyvips.Image.new_from_file(\n str(thumbnail_path),\n access=\"sequential\"\n )\n\n array = np.ndarray(\n buffer=thumbnail.write_to_memory(),\n dtype=np.uint8,\n shape=(thumbnail.height, thumbnail.width, thumbnail.bands)\n )\n\n if array.shape[2] >= 3:\n array = array[:, :, :3]\n\n return array\n\n def create_tissue_mask(self, image_id: int) -> np.ndarray:\n \"\"\"\n Create a coarse tissue mask using HSV saturation\n and Otsu thresholding.\n\n Higher-saturation regions are treated as likely tissue.\n \"\"\"\n\n thumbnail = self.load_thumbnail(image_id)\n\n hsv = cv2.cvtColor(thumbnail, cv2.COLOR_RGB2HSV)\n\n saturation = hsv[:, :, 1]\n\n _, mask = cv2.threshold(\n saturation,\n 0,\n 255,\n cv2.THRESH_BINARY + cv2.THRESH_OTSU\n )\n\n # Remove tiny isolated regions.\n kernel = np.ones((3, 3), np.uint8)\n\n mask = cv2.morphologyEx(\n mask,\n cv2.MORPH_OPEN,\n kernel\n )\n\n mask = cv2.morphologyEx(\n mask,\n cv2.MORPH_CLOSE,\n kernel\n )\n\n return mask > 0\n\n def get_thumbnail_tissue_ratio(\n self,\n image_id: int,\n x: int,\n y: int,\n width: int,\n height: int\n ) -> float:\n \"\"\"\n Estimate the tissue fraction corresponding to one\n full-resolution WSI tile.\n \"\"\"\n\n record = self.metadata.get_record(image_id)\n\n if record[\"is_tma\"]:\n raise ValueError(\n \"Thumbnail-guided selection is only for WSIs.\"\n )\n\n mask = self.create_tissue_mask(image_id)\n\n thumb_height, thumb_width = mask.shape\n\n scale_x = thumb_width / record[\"width\"]\n scale_y = thumb_height / record[\"height\"]\n\n tx0 = max(0, int(np.floor(x * scale_x)))\n ty0 = max(0, int(np.floor(y * scale_y)))\n\n tx1 = min(\n thumb_width,\n int(np.ceil((x + width) * scale_x))\n )\n\n ty1 = min(\n thumb_height,\n int(np.ceil((y + height) * scale_y))\n )\n\n if tx1 <= tx0 or ty1 <= ty0:\n return 0.0\n\n region = mask[ty0:ty1, tx0:tx1]\n\n return float(region.mean())\n\n def select_candidates(\n self,\n coordinates,\n tissue_ratio_threshold: float = 0.20\n ):\n \"\"\"\n Keep candidate WSI coordinates whose corresponding\n thumbnail region contains enough likely tissue.\n\n TMA coordinates are retained because TMAs do not have\n official thumbnails.\n \"\"\"\n\n if not coordinates:\n return []\n\n image_id = coordinates[0].image_id\n\n record = self.metadata.get_record(image_id)\n\n if record[\"is_tma\"]:\n return list(coordinates)\n\n mask = self.create_tissue_mask(image_id)\n\n thumb_height, thumb_width = mask.shape\n\n scale_x = thumb_width / record[\"width\"]\n scale_y = thumb_height / record[\"height\"]\n\n selected = []\n\n for coordinate in coordinates:\n\n tx0 = max(\n 0,\n int(np.floor(coordinate.x * scale_x))\n )\n\n ty0 = max(\n 0,\n int(np.floor(coordinate.y * scale_y))\n )\n\n tx1 = min(\n thumb_width,\n int(np.ceil(\n (coordinate.x + coordinate.source_size) * scale_x\n ))\n )\n\n ty1 = min(\n thumb_height,\n int(np.ceil(\n (coordinate.y + coordinate.source_size) * scale_y\n ))\n )\n\n if tx1 <= tx0 or ty1 <= ty0:\n continue\n\n region = mask[ty0:ty1, tx0:tx1]\n\n tissue_ratio = float(region.mean())\n\n if tissue_ratio >= tissue_ratio_threshold:\n selected.append(coordinate)\n\n return selected\n'''\n\nfile_path.write_text(code)\n\nprint(f\"Created: {file_path}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.756093Z","iopub.status.idle":"2026-09-26T01:55:11.756358Z","shell.execute_reply.started":"2026-09-26T01:55:11.756247Z","shell.execute_reply":"2026-09-26T01:55:11.756262Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.thumbnail_selector import ThumbnailCandidateSelector\n\nprint(\"Thumbnail selector imported successfully! ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.756829Z","iopub.status.idle":"2026-09-26T01:55:11.757062Z","shell.execute_reply.started":"2026-09-26T01:55:11.756956Z","shell.execute_reply":"2026-09-26T01:55:11.756969Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"selector = ThumbnailCandidateSelector(metadata)\n\nprint(\"Thumbnail candidate selector initialized! ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.757569Z","iopub.status.idle":"2026-09-26T01:55:11.757798Z","shell.execute_reply.started":"2026-09-26T01:55:11.757672Z","shell.execute_reply":"2026-09-26T01:55:11.757685Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wsi_mask = selector.create_tissue_mask(4)\n\nprint(\"WSI tissue mask created! ✅\")\nprint(\"Mask shape:\", wsi_mask.shape)\nprint(\"Tissue fraction:\", wsi_mask.mean())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.758342Z","iopub.status.idle":"2026-09-26T01:55:11.758582Z","shell.execute_reply.started":"2026-09-26T01:55:11.758474Z","shell.execute_reply":"2026-09-26T01:55:11.758486Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import matplotlib.pyplot as plt\n\nplt.figure(figsize=(8, 7))\nplt.imshow(wsi_mask, cmap=\"gray\")\nplt.title(\"WSI 4 — Thumbnail Tissue Mask\")\nplt.axis(\"off\")\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.759922Z","iopub.status.idle":"2026-09-26T01:55:11.760301Z","shell.execute_reply.started":"2026-09-26T01:55:11.760101Z","shell.execute_reply":"2026-09-26T01:55:11.760142Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"selected_wsi_coords = selector.select_candidates(\n wsi_coords,\n tissue_ratio_threshold=0.20\n)\n\nprint(\"Original WSI candidates:\", len(wsi_coords))\nprint(\"Thumbnail-selected candidates:\", len(selected_wsi_coords))\nprint(\n \"Candidates removed:\",\n len(wsi_coords) - len(selected_wsi_coords)\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.761507Z","iopub.status.idle":"2026-09-26T01:55:11.761744Z","shell.execute_reply.started":"2026-09-26T01:55:11.761627Z","shell.execute_reply":"2026-09-26T01:55:11.76164Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"First 5 selected coordinates:\")\n\nfor coord in selected_wsi_coords[:5]:\n print(coord)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.762958Z","iopub.status.idle":"2026-09-26T01:55:11.763292Z","shell.execute_reply.started":"2026-09-26T01:55:11.763143Z","shell.execute_reply":"2026-09-26T01:55:11.763171Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"selected_tile = tile_extractor.extract_tile(\n selected_wsi_coords[0]\n)\n\nprint(\"Selected tile extracted successfully! ✅\")\nprint(\"Shape:\", selected_tile.shape)\nprint(\"Dtype:\", selected_tile.dtype)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.764352Z","iopub.status.idle":"2026-09-26T01:55:11.764686Z","shell.execute_reply.started":"2026-09-26T01:55:11.764504Z","shell.execute_reply":"2026-09-26T01:55:11.764525Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"plt.figure(figsize=(5, 5))\nplt.imshow(selected_tile)\nplt.title(\"Thumbnail-Selected WSI Tile\")\nplt.axis(\"off\")\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.766154Z","iopub.status.idle":"2026-09-26T01:55:11.766467Z","shell.execute_reply.started":"2026-09-26T01:55:11.766282Z","shell.execute_reply":"2026-09-26T01:55:11.766295Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"selected_tma_coords = selector.select_candidates(\n tma_coords,\n tissue_ratio_threshold=0.20\n)\n\nprint(\"Original TMA candidates:\", len(tma_coords))\nprint(\"Selected TMA candidates:\", len(selected_tma_coords))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.767753Z","iopub.status.idle":"2026-09-26T01:55:11.768167Z","shell.execute_reply.started":"2026-09-26T01:55:11.767938Z","shell.execute_reply":"2026-09-26T01:55:11.767962Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%writefile /kaggle/working/ovarian-trustworthy-ai/src/preprocessing/stain_normalization.py\n\nfrom dataclasses import dataclass\n\nimport numpy as np\n\n\n@dataclass\nclass StainReference:\n stain_matrix: np.ndarray\n max_concentration: np.ndarray\n io: float = 255.0\n\n\nclass MacenkoNormalizer:\n \"\"\"\n Macenko stain normalization for H&E histopathology images.\n\n Pipeline:\n RGB\n -> Optical Density\n -> foreground filtering\n -> covariance/eigenvector analysis\n -> angular percentile stain vectors\n -> stain concentration estimation\n -> concentration normalization\n -> RGB reconstruction\n \"\"\"\n\n def __init__(\n self,\n beta: float = 0.15,\n alpha: float = 1.0,\n max_samples: int = 10000,\n io: float = 255.0,\n ):\n self.beta = beta\n self.alpha = alpha\n self.max_samples = max_samples\n self.io = io\n\n self.reference = None\n\n # =========================================================\n # RGB -> Optical Density\n # =========================================================\n def rgb_to_od(self, image: np.ndarray) -> np.ndarray:\n\n image = image.astype(np.float32)\n\n image = np.clip(image, 0, 255)\n\n od = -np.log(\n (image + 1.0) / self.io\n )\n\n return od\n\n # =========================================================\n # Optical Density -> RGB\n # =========================================================\n def od_to_rgb(self, od: np.ndarray) -> np.ndarray:\n\n rgb = self.io * np.exp(-od) - 1.0\n\n rgb = np.clip(rgb, 0, 255)\n\n return rgb.astype(np.uint8)\n\n # =========================================================\n # Pixel sampling\n # =========================================================\n def _sample_pixels(\n self,\n pixels: np.ndarray\n ) -> np.ndarray:\n\n if len(pixels) <= self.max_samples:\n return pixels\n\n rng = np.random.default_rng(42)\n\n indices = rng.choice(\n len(pixels),\n size=self.max_samples,\n replace=False,\n )\n\n return pixels[indices]\n\n # =========================================================\n # Estimate Macenko stain matrix\n # =========================================================\n def _estimate_stain_matrix(\n self,\n image: np.ndarray,\n ) -> np.ndarray:\n\n if image.ndim != 3 or image.shape[2] != 3:\n raise ValueError(\n \"Image must have shape H x W x 3.\"\n )\n\n # -----------------------------------------------------\n # Step 1: RGB -> OD\n # -----------------------------------------------------\n od = self.rgb_to_od(image)\n\n pixels = od.reshape(-1, 3)\n\n # -----------------------------------------------------\n # Step 2: Remove low-OD/background pixels\n # -----------------------------------------------------\n mask = np.min(pixels, axis=1) >= self.beta\n\n odhat = pixels[mask]\n\n if len(odhat) < 100:\n raise ValueError(\n \"Not enough foreground pixels for \"\n \"Macenko stain estimation.\"\n )\n\n # -----------------------------------------------------\n # Step 3: Sample pixels for efficiency\n # -----------------------------------------------------\n odhat = self._sample_pixels(odhat)\n\n # -----------------------------------------------------\n # Step 4: Covariance matrix\n # -----------------------------------------------------\n covariance = np.cov(\n odhat.T\n )\n\n # -----------------------------------------------------\n # Step 5: Eigen decomposition\n # -----------------------------------------------------\n eigvals, eigvecs = np.linalg.eigh(\n covariance\n )\n\n # Eigenvalues are returned in ascending order.\n # Select the two largest eigenvectors.\n V = eigvecs[:, [2, 1]]\n\n # -----------------------------------------------------\n # Step 6: Consistent eigenvector orientation\n # -----------------------------------------------------\n if V[0, 0] < 0:\n V[:, 0] *= -1\n\n if V[0, 1] < 0:\n V[:, 1] *= -1\n\n # -----------------------------------------------------\n # Step 7: Project OD data onto stain plane\n # -----------------------------------------------------\n projected = odhat @ V\n\n # -----------------------------------------------------\n # Step 8: Calculate angular coordinates\n # -----------------------------------------------------\n phi = np.arctan2(\n projected[:, 1],\n projected[:, 0],\n )\n\n # -----------------------------------------------------\n # Step 9: Robust angular extremes\n # -----------------------------------------------------\n min_phi = np.percentile(\n phi,\n self.alpha,\n )\n\n max_phi = np.percentile(\n phi,\n 100.0 - self.alpha,\n )\n\n # -----------------------------------------------------\n # Step 10: Convert extremes back to OD space\n # -----------------------------------------------------\n v_min = V @ np.array([\n np.cos(min_phi),\n np.sin(min_phi),\n ])\n\n v_max = V @ np.array([\n np.cos(max_phi),\n np.sin(max_phi),\n ])\n\n # Normalize vectors\n v_min = v_min / (\n np.linalg.norm(v_min) + 1e-8\n )\n\n v_max = v_max / (\n np.linalg.norm(v_max) + 1e-8\n )\n\n # -----------------------------------------------------\n # Step 11: Consistent H&E ordering\n # -----------------------------------------------------\n if v_min[0] > v_max[0]:\n stain_matrix = np.array([\n v_min,\n v_max,\n ])\n else:\n stain_matrix = np.array([\n v_max,\n v_min,\n ])\n\n # Normalize rows\n stain_matrix = (\n stain_matrix\n / (\n np.linalg.norm(\n stain_matrix,\n axis=1,\n keepdims=True,\n )\n + 1e-8\n )\n )\n\n return stain_matrix\n\n # =========================================================\n # Estimate stain concentrations\n # =========================================================\n def _get_concentrations(\n self,\n image: np.ndarray,\n stain_matrix: np.ndarray,\n ):\n\n od = self.rgb_to_od(image)\n\n height, width, _ = od.shape\n\n pixels = od.reshape(-1, 3)\n\n # Foreground mask\n mask = np.min(\n pixels,\n axis=1\n ) >= self.beta\n\n odhat = pixels[mask]\n\n if len(odhat) == 0:\n raise ValueError(\n \"No foreground pixels found.\"\n )\n\n # Solve:\n #\n # OD = C @ stain_matrix\n #\n concentrations = np.linalg.lstsq(\n stain_matrix.T,\n odhat.T,\n rcond=None,\n )[0].T\n\n return (\n concentrations,\n mask,\n (height, width),\n )\n\n # =========================================================\n # Fit reference\n # =========================================================\n def fit(\n self,\n reference_images,\n ):\n\n if not reference_images:\n raise ValueError(\n \"At least one reference image is required.\"\n )\n\n stain_matrices = []\n concentration_samples = []\n\n for image in reference_images:\n\n stain_matrix = (\n self._estimate_stain_matrix(image)\n )\n\n concentrations, _, _ = (\n self._get_concentrations(\n image,\n stain_matrix,\n )\n )\n\n stain_matrices.append(\n stain_matrix\n )\n\n concentration_samples.append(\n concentrations\n )\n\n # First reference tile determines\n # the target stain appearance.\n target_stain_matrix = (\n stain_matrices[0]\n )\n\n all_concentrations = np.concatenate(\n concentration_samples,\n axis=0,\n )\n\n target_max_concentration = (\n np.percentile(\n all_concentrations,\n 99,\n axis=0,\n )\n )\n\n self.reference = StainReference(\n stain_matrix=target_stain_matrix,\n max_concentration=target_max_concentration,\n io=self.io,\n )\n\n return self\n\n # =========================================================\n # Transform / normalize image\n # =========================================================\n def transform(\n self,\n image: np.ndarray,\n ) -> np.ndarray:\n\n if self.reference is None:\n raise RuntimeError(\n \"Normalizer has not been fitted. \"\n \"Call fit() first.\"\n )\n\n source_stain_matrix = (\n self._estimate_stain_matrix(image)\n )\n\n concentrations, mask, shape = (\n self._get_concentrations(\n image,\n source_stain_matrix,\n )\n )\n\n # -----------------------------------------------------\n # Source stain quantity\n # -----------------------------------------------------\n source_max = np.percentile(\n concentrations,\n 99,\n axis=0,\n )\n\n source_max = np.maximum(\n source_max,\n 1e-8,\n )\n\n # -----------------------------------------------------\n # Match source concentration to reference\n # -----------------------------------------------------\n normalized_concentrations = (\n concentrations\n / source_max\n )\n\n normalized_concentrations *= (\n self.reference.max_concentration\n )\n\n height, width = shape\n\n # -----------------------------------------------------\n # Reconstruct normalized OD\n # -----------------------------------------------------\n normalized_od = np.zeros(\n (height * width, 3),\n dtype=np.float32,\n )\n\n normalized_od[mask] = (\n normalized_concentrations\n @ self.reference.stain_matrix\n )\n\n normalized_od = normalized_od.reshape(\n height,\n width,\n 3,\n )\n\n # -----------------------------------------------------\n # OD -> RGB\n # -----------------------------------------------------\n normalized_image = (\n self.od_to_rgb(\n normalized_od\n )\n )\n\n return normalized_image\n\n # =========================================================\n # Fit + Transform\n # =========================================================\n def fit_transform(\n self,\n reference_images,\n image: np.ndarray,\n ) -> np.ndarray:\n\n self.fit(reference_images)\n\n return self.transform(image)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.769274Z","iopub.status.idle":"2026-09-26T01:55:11.769604Z","shell.execute_reply.started":"2026-09-26T01:55:11.769475Z","shell.execute_reply":"2026-09-26T01:55:11.769498Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimport src.preprocessing.stain_normalization as stain_module\n\nimportlib.reload(stain_module)\n\nMacenkoNormalizer = stain_module.MacenkoNormalizer\n\nprint(\"Corrected Macenko implementation loaded! ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.771408Z","iopub.status.idle":"2026-09-26T01:55:11.771758Z","shell.execute_reply.started":"2026-09-26T01:55:11.77157Z","shell.execute_reply":"2026-09-26T01:55:11.77159Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"normalizer = MacenkoNormalizer()\n\nnormalizer.fit([reference_tile])\n\nprint(\"Macenko fitted successfully! ✅\")\n\nprint(\"\\nCorrected stain matrix:\")\nprint(normalizer.reference.stain_matrix)\n\nprint(\"\\nMaximum concentrations:\")\nprint(normalizer.reference.max_concentration)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.772649Z","iopub.status.idle":"2026-09-26T01:55:11.772975Z","shell.execute_reply.started":"2026-09-26T01:55:11.772795Z","shell.execute_reply":"2026-09-26T01:55:11.77283Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\n\nprint(\"Project exists:\", project_root.exists())\n\npreprocessing_dir = project_root / \"src\" / \"preprocessing\"\n\nprint(\"\\nPreprocessing folder:\")\nfor item in preprocessing_dir.iterdir():\n print(\" -\", item.name)\n\nprint(\"\\nStain normalization file exists:\",\n (preprocessing_dir / \"stain_normalization.py\").exists())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.774542Z","iopub.status.idle":"2026-09-26T01:55:11.774871Z","shell.execute_reply.started":"2026-09-26T01:55:11.774703Z","shell.execute_reply":"2026-09-26T01:55:11.774724Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nfile_path = Path(\n \"/kaggle/working/ovarian-trustworthy-ai/\"\n \"src/preprocessing/stain_normalization.py\"\n)\n\nfile_path.parent.mkdir(parents=True, exist_ok=True)\n\nfile_path.touch()\n\nprint(\"File created:\", file_path)\nprint(\"Exists:\", file_path.exists())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.776008Z","iopub.status.idle":"2026-09-26T01:55:11.776357Z","shell.execute_reply.started":"2026-09-26T01:55:11.776185Z","shell.execute_reply":"2026-09-26T01:55:11.776207Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nos.makedirs(\n f\"{project_root}/src/preprocessing\",\n exist_ok=True\n)\n\nprint(\"Folder created/check passed ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.78392Z","iopub.status.idle":"2026-09-26T01:55:11.784261Z","shell.execute_reply.started":"2026-09-26T01:55:11.784099Z","shell.execute_reply":"2026-09-26T01:55:11.78414Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%writefile /kaggle/working/ovarian-trustworthy-ai/src/preprocessing/stain_normalization.py\n\nfrom dataclasses import dataclass\n\nimport numpy as np\n\n\n@dataclass\nclass StainReference:\n stain_matrix: np.ndarray\n max_concentration: np.ndarray\n io: float = 255.0\n\n\nclass MacenkoNormalizer:\n\n def __init__(\n self,\n beta: float = 0.15,\n alpha: float = 1.0,\n max_samples: int = 10000,\n io: float = 255.0,\n black_threshold: float = 30.0,\n ):\n self.beta = beta\n self.alpha = alpha\n self.max_samples = max_samples\n self.io = io\n self.black_threshold = black_threshold\n\n self.reference = None\n\n # ---------------------------------------------------------\n # RGB -> Optical Density\n # ---------------------------------------------------------\n def rgb_to_od(self, image: np.ndarray) -> np.ndarray:\n\n image = image.astype(np.float32)\n image = np.clip(image, 0, 255)\n\n od = -np.log10(\n (image + 1.0) / self.io\n )\n\n return od\n\n # ---------------------------------------------------------\n # Optical Density -> RGB\n # ---------------------------------------------------------\n def od_to_rgb(self, od: np.ndarray) -> np.ndarray:\n\n rgb = (\n self.io * np.power(10.0, -od)\n - 1.0\n )\n\n rgb = np.clip(rgb, 0, 255)\n\n return rgb.astype(np.uint8)\n\n # ---------------------------------------------------------\n # Randomly sample pixels for efficiency\n # ---------------------------------------------------------\n def _sample_pixels(\n self,\n pixels: np.ndarray,\n ) -> np.ndarray:\n\n if len(pixels) <= self.max_samples:\n return pixels\n\n rng = np.random.default_rng(42)\n\n indices = rng.choice(\n len(pixels),\n size=self.max_samples,\n replace=False,\n )\n\n return pixels[indices]\n\n # ---------------------------------------------------------\n # Create stain/foreground mask\n # ---------------------------------------------------------\n def _get_stain_mask(\n self,\n image: np.ndarray,\n od: np.ndarray,\n ) -> np.ndarray:\n\n rgb_pixels = image.reshape(-1, 3).astype(\n np.float32\n )\n\n od_pixels = od.reshape(-1, 3)\n\n # Remove near-black scanner / border pixels\n not_black = (\n np.max(rgb_pixels, axis=1)\n >= self.black_threshold\n )\n\n # Standard Macenko-style OD foreground filtering\n sufficient_od = (\n np.min(od_pixels, axis=1)\n >= self.beta\n )\n\n mask = not_black & sufficient_od\n\n return mask\n\n # ---------------------------------------------------------\n # Estimate H&E stain matrix\n # ---------------------------------------------------------\n def _estimate_stain_matrix(\n self,\n image: np.ndarray,\n ):\n\n if image.ndim != 3 or image.shape[2] != 3:\n raise ValueError(\n \"Image must have shape H x W x 3.\"\n )\n\n od = self.rgb_to_od(image)\n\n pixels = od.reshape(-1, 3)\n\n # Create foreground/stain mask\n mask = self._get_stain_mask(\n image,\n od,\n )\n\n od_pixels = pixels[mask]\n\n if len(od_pixels) < 100:\n raise ValueError(\n \"Not enough stained pixels \"\n \"to estimate stain matrix.\"\n )\n\n # Limit computation for large tiles\n od_pixels = self._sample_pixels(\n od_pixels\n )\n\n # -----------------------------------------------------\n # Singular Value Decomposition\n # -----------------------------------------------------\n _, _, vh = np.linalg.svd(\n od_pixels,\n full_matrices=False,\n )\n\n # First two principal directions\n v = vh[:2].T\n\n # Project OD values onto SVD plane\n projected = od_pixels @ v\n\n # Calculate angular coordinates\n phi = np.arctan2(\n projected[:, 1],\n projected[:, 0],\n )\n\n # Robust percentile limits\n min_phi = np.percentile(\n phi,\n self.alpha,\n )\n\n max_phi = np.percentile(\n phi,\n 100.0 - self.alpha,\n )\n\n vector_1 = (\n v[:, 0] * np.cos(min_phi)\n + v[:, 1] * np.sin(min_phi)\n )\n\n vector_2 = (\n v[:, 0] * np.cos(max_phi)\n + v[:, 1] * np.sin(max_phi)\n )\n\n # Normalize stain vectors\n vector_1 = vector_1 / (\n np.linalg.norm(vector_1)\n + 1e-8\n )\n\n vector_2 = vector_2 / (\n np.linalg.norm(vector_2)\n + 1e-8\n )\n\n # Consistent H&E ordering\n if vector_1[0] > vector_2[0]:\n stain_matrix = np.stack(\n [vector_1, vector_2],\n axis=0,\n )\n else:\n stain_matrix = np.stack(\n [vector_2, vector_1],\n axis=0,\n )\n\n return stain_matrix\n\n # ---------------------------------------------------------\n # Estimate stain concentrations\n # ---------------------------------------------------------\n def _get_concentrations(\n self,\n image: np.ndarray,\n stain_matrix: np.ndarray,\n ):\n\n od = self.rgb_to_od(image)\n\n height, width, _ = od.shape\n\n pixels = od.reshape(-1, 3)\n\n # Same foreground mask used during\n # stain matrix estimation\n mask = self._get_stain_mask(\n image,\n od,\n )\n\n od_stained = pixels[mask]\n\n if len(od_stained) == 0:\n raise ValueError(\n \"No stained pixels found in image.\"\n )\n\n # Solve:\n # OD = stain_matrix.T @ concentrations\n concentrations = np.linalg.lstsq(\n stain_matrix.T,\n od_stained.T,\n rcond=None,\n )[0]\n\n # Prevent small numerical negative values\n concentrations = np.maximum(\n concentrations,\n 0.0,\n )\n\n return (\n concentrations,\n mask,\n (height, width),\n )\n\n # ---------------------------------------------------------\n # Fit reference stain appearance\n # ---------------------------------------------------------\n def fit(\n self,\n reference_images,\n ):\n\n if not reference_images:\n raise ValueError(\n \"At least one reference image \"\n \"is required.\"\n )\n\n stain_matrices = []\n concentration_samples = []\n\n for image in reference_images:\n\n if image.ndim != 3 or image.shape[2] != 3:\n raise ValueError(\n \"Reference image must have \"\n \"shape H x W x 3.\"\n )\n\n stain_matrix = (\n self._estimate_stain_matrix(\n image\n )\n )\n\n concentrations, _, _ = (\n self._get_concentrations(\n image,\n stain_matrix,\n )\n )\n\n stain_matrices.append(\n stain_matrix\n )\n\n concentration_samples.append(\n concentrations\n )\n\n # First reference image defines\n # target stain color appearance\n target_stain_matrix = (\n stain_matrices[0]\n )\n\n # Combine concentration samples\n all_concentrations = np.concatenate(\n concentration_samples,\n axis=1,\n )\n\n # 99th percentile defines target\n # stain intensity\n target_max_concentration = (\n np.percentile(\n all_concentrations,\n 99,\n axis=1,\n )\n )\n\n self.reference = StainReference(\n stain_matrix=target_stain_matrix,\n max_concentration=(\n target_max_concentration\n ),\n io=self.io,\n )\n\n return self\n\n # ---------------------------------------------------------\n # Transform one image\n # ---------------------------------------------------------\n def transform(\n self,\n image: np.ndarray,\n ) -> np.ndarray:\n\n if self.reference is None:\n raise RuntimeError(\n \"Normalizer has not been fitted. \"\n \"Call fit() first.\"\n )\n\n if image.ndim != 3 or image.shape[2] != 3:\n raise ValueError(\n \"Image must have shape H x W x 3.\"\n )\n\n # Estimate source stain matrix\n source_stain_matrix = (\n self._estimate_stain_matrix(\n image\n )\n )\n\n # Get source concentrations\n concentrations, mask, shape = (\n self._get_concentrations(\n image,\n source_stain_matrix,\n )\n )\n\n # Source 99th percentile\n source_max_concentration = (\n np.percentile(\n concentrations,\n 99,\n axis=1,\n )\n )\n\n source_max_concentration = (\n np.maximum(\n source_max_concentration,\n 1e-8,\n )\n )\n\n # Normalize stain intensity\n normalized_concentrations = (\n concentrations\n / source_max_concentration[:, None]\n )\n\n # Match target stain intensity\n normalized_concentrations *= (\n self.reference.max_concentration[\n :, None\n ]\n )\n\n height, width = shape\n\n # -----------------------------------------------------\n # IMPORTANT:\n # Start with ORIGINAL image.\n #\n # This preserves:\n # - black borders\n # - white background\n # - non-stain pixels\n # -----------------------------------------------------\n normalized_image = image.copy()\n\n # Reconstruct only the pixels that\n # were identified as stained tissue\n normalized_od_stained = (\n self.reference.stain_matrix.T\n @ normalized_concentrations\n ).T\n\n normalized_rgb_stained = (\n self.od_to_rgb(\n normalized_od_stained\n )\n )\n\n # Put normalized pixels back into\n # their original locations\n normalized_image[\n mask.reshape(height, width)\n ] = normalized_rgb_stained\n\n return normalized_image\n\n # ---------------------------------------------------------\n # Fit + Transform\n # ---------------------------------------------------------\n def fit_transform(\n self,\n reference_images,\n image: np.ndarray,\n ):\n\n self.fit(reference_images)\n\n return self.transform(image)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.785824Z","iopub.status.idle":"2026-09-26T01:55:11.786185Z","shell.execute_reply.started":"2026-09-26T01:55:11.785977Z","shell.execute_reply":"2026-09-26T01:55:11.785991Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\npath = \"/kaggle/working/ovarian-trustworthy-ai/src/preprocessing/stain_normalization.py\"\n\nprint(\"Exists:\", os.path.exists(path))\nprint(\"Size:\", os.path.getsize(path), \"bytes\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.787294Z","iopub.status.idle":"2026-09-26T01:55:11.787665Z","shell.execute_reply.started":"2026-09-26T01:55:11.787505Z","shell.execute_reply":"2026-09-26T01:55:11.787528Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nfile_path = Path(\n \"/kaggle/working/ovarian-trustworthy-ai/\"\n \"src/preprocessing/stain_normalization.py\"\n)\n\nprint(\"Exists:\", file_path.exists())\nprint(\"Size:\", file_path.stat().st_size, \"bytes\")\n\nprint(\"\\nFirst 5 lines:\")\nprint(\"\\n\".join(file_path.read_text().splitlines()[:5]))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.789202Z","iopub.status.idle":"2026-09-26T01:55:11.789588Z","shell.execute_reply.started":"2026-09-26T01:55:11.789343Z","shell.execute_reply":"2026-09-26T01:55:11.789383Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.stain_normalization import MacenkoNormalizer\n\nprint(\"Macenko normalizer imported successfully! ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.79084Z","iopub.status.idle":"2026-09-26T01:55:11.791137Z","shell.execute_reply.started":"2026-09-26T01:55:11.790992Z","shell.execute_reply":"2026-09-26T01:55:11.791006Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# M1.5.2 — Get representative candidate tiles\n\nimport sys\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nfrom src.preprocessing.metadata import UBCOceanMetadata\nfrom src.preprocessing.image_reader import UBCOceanImageReader\nfrom src.preprocessing.slide_type import SlideTypeHandler\nfrom src.preprocessing.tile_extractor import TileExtractor\nfrom src.preprocessing.thumbnail_selector import ThumbnailCandidateSelector\n\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nmetadata = UBCOceanMetadata(dataset_root)\nimage_reader = UBCOceanImageReader(dataset_root)\nslide_handler = SlideTypeHandler(metadata)\n\ntile_extractor = TileExtractor(\n image_reader,\n slide_handler\n)\n\nselector = ThumbnailCandidateSelector(metadata)\n\n\n# Generate WSI candidate grid\ncoordinates = tile_extractor.generate_grid(4)\n\n# Apply thumbnail tissue filtering\nselected_coordinates = selector.select_candidates(\n coordinates,\n tissue_ratio_threshold=0.20\n)\n\nprint(\"Total candidates:\", len(coordinates))\nprint(\"Selected candidates:\", len(selected_coordinates))\n\n\n# Select 6 spread-out candidates\nindices = np.linspace(\n 0,\n len(selected_coordinates) - 1,\n 6,\n dtype=int\n)\n\nreference_tiles = []\n\nfor idx in indices:\n coordinate = selected_coordinates[idx]\n\n tile = tile_extractor.extract_tile(coordinate)\n\n reference_tiles.append(tile)\n\n print(\n f\"Tile {len(reference_tiles)}: \"\n f\"x={coordinate.x}, \"\n f\"y={coordinate.y}, \"\n f\"shape={tile.shape}\"\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.791991Z","iopub.status.idle":"2026-09-26T01:55:11.792335Z","shell.execute_reply.started":"2026-09-26T01:55:11.792147Z","shell.execute_reply":"2026-09-26T01:55:11.792171Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# M1.5.2 — Visualize candidate reference tiles\n\nimport matplotlib.pyplot as plt\n\nfig, axes = plt.subplots(2, 3, figsize=(12, 8))\n\nfor i, (ax, tile) in enumerate(zip(axes.ravel(), reference_tiles)):\n ax.imshow(tile)\n ax.set_title(\n f\"Reference Candidate {i + 1}\\n\"\n f\"x={selected_coordinates[indices[i]].x}, \"\n f\"y={selected_coordinates[indices[i]].y}\"\n )\n ax.axis(\"off\")\n\nplt.tight_layout()\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.793727Z","iopub.status.idle":"2026-09-26T01:55:11.79396Z","shell.execute_reply.started":"2026-09-26T01:55:11.793849Z","shell.execute_reply":"2026-09-26T01:55:11.793862Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# M1.5.3 — Fit Macenko using Candidate 2\n\nfrom src.preprocessing.stain_normalization import MacenkoNormalizer\n\n# Candidate 2 is index 1 because Python uses zero-based indexing\nreference_tile = reference_tiles[1]\n\nprint(\"Reference tile shape:\", reference_tile.shape)\nprint(\"Reference tile dtype:\", reference_tile.dtype)\n\nnormalizer = MacenkoNormalizer()\n\nnormalizer.fit([reference_tile])\n\nprint(\"Macenko fitted successfully! ✅\")\nprint(\"\\nReference stain matrix:\")\nprint(normalizer.reference.stain_matrix)\n\nprint(\"\\nReference maximum concentrations:\")\nprint(normalizer.reference.max_concentration)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.795641Z","iopub.status.idle":"2026-09-26T01:55:11.795917Z","shell.execute_reply.started":"2026-09-26T01:55:11.795806Z","shell.execute_reply":"2026-09-26T01:55:11.795819Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# M1.5.4 — Normalize a visually different tissue tile\n\ntest_tile = reference_tiles[4] # Candidate 5\n\nnormalized_tile = normalizer.transform(test_tile)\n\nprint(\"Original tile:\")\nprint(\" Shape:\", test_tile.shape)\nprint(\" Dtype:\", test_tile.dtype)\n\nprint(\"\\nNormalized tile:\")\nprint(\" Shape:\", normalized_tile.shape)\nprint(\" Dtype:\", normalized_tile.dtype)\n\nprint(\"\\nNormalization completed successfully! ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.796758Z","iopub.status.idle":"2026-09-26T01:55:11.797073Z","shell.execute_reply.started":"2026-09-26T01:55:11.796953Z","shell.execute_reply":"2026-09-26T01:55:11.796969Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import matplotlib.pyplot as plt\n\nfig, axes = plt.subplots(1, 2, figsize=(10, 5))\n\naxes[0].imshow(test_tile)\naxes[0].set_title(\"Original — Candidate 5\")\naxes[0].axis(\"off\")\n\naxes[1].imshow(normalized_tile)\naxes[1].set_title(\"Macenko Normalized\")\naxes[1].axis(\"off\")\n\nplt.tight_layout()\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.798283Z","iopub.status.idle":"2026-09-26T01:55:11.798627Z","shell.execute_reply.started":"2026-09-26T01:55:11.798456Z","shell.execute_reply":"2026-09-26T01:55:11.798472Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# M1.5.5 — Normalize multiple candidate tiles\n\ncomparison_results = []\n\nfor i, tile in enumerate(reference_tiles):\n\n if i == 1:\n # Candidate 2 is our reference.\n # Don't normalize it against itself.\n continue\n\n normalized = normalizer.transform(tile)\n\n comparison_results.append(\n {\n \"candidate\": i + 1,\n \"original\": tile,\n \"normalized\": normalized,\n }\n )\n\nprint(\n \"Successfully normalized\",\n len(comparison_results),\n \"test tiles. ✅\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.799712Z","iopub.status.idle":"2026-09-26T01:55:11.799983Z","shell.execute_reply.started":"2026-09-26T01:55:11.799826Z","shell.execute_reply":"2026-09-26T01:55:11.799839Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Visual comparison\n\nfig, axes = plt.subplots(\n 5,\n 2,\n figsize=(10, 20)\n)\n\nfor row, result in enumerate(comparison_results):\n\n axes[row, 0].imshow(result[\"original\"])\n axes[row, 0].set_title(\n f\"Candidate {result['candidate']} — Original\"\n )\n axes[row, 0].axis(\"off\")\n\n axes[row, 1].imshow(result[\"normalized\"])\n axes[row, 1].set_title(\n f\"Candidate {result['candidate']} — Normalized\"\n )\n axes[row, 1].axis(\"off\")\n\nplt.tight_layout()\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.801858Z","iopub.status.idle":"2026-09-26T01:55:11.802208Z","shell.execute_reply.started":"2026-09-26T01:55:11.802025Z","shell.execute_reply":"2026-09-26T01:55:11.802044Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nprint(\"Project exists:\", os.path.exists(project_root))\nprint(\"\\nProject contents:\")\nprint(os.listdir(project_root))\n\nprint(\"\\nSRC contents:\")\nprint(os.listdir(os.path.join(project_root, \"src\")))\n\nprint(\"\\nPREPROCESSING contents:\")\nprint(os.listdir(os.path.join(project_root, \"src\", \"preprocessing\")))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.803807Z","iopub.status.idle":"2026-09-26T01:55:11.804225Z","shell.execute_reply.started":"2026-09-26T01:55:11.803999Z","shell.execute_reply":"2026-09-26T01:55:11.804023Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nfolders = [\n \"src\",\n \"src/preprocessing\",\n \"src/quality\",\n \"src/features\",\n \"src/classification\",\n \"src/calibration\",\n \"src/explainability\",\n \"src/routing\",\n \"src/evaluation\",\n \"config\",\n \"data\",\n \"cache/tiles\",\n \"cache/embeddings\",\n \"outputs/logs\",\n \"outputs/metrics\",\n \"outputs/visualizations\",\n \"tests\",\n \"notebooks\",\n]\n\nfor folder in folders:\n os.makedirs(os.path.join(project_root, folder), exist_ok=True)\n\nprint(\"Project structure recreated ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.806222Z","iopub.status.idle":"2026-09-26T01:55:11.806583Z","shell.execute_reply.started":"2026-09-26T01:55:11.806399Z","shell.execute_reply":"2026-09-26T01:55:11.806422Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from pathlib import Path\n\nproject_root = Path(\"/kaggle/working/ovarian-trustworthy-ai\")\n\ninit_files = [\n project_root / \"src\" / \"__init__.py\",\n project_root / \"src\" / \"preprocessing\" / \"__init__.py\",\n project_root / \"src\" / \"quality\" / \"__init__.py\",\n project_root / \"src\" / \"features\" / \"__init__.py\",\n project_root / \"src\" / \"classification\" / \"__init__.py\",\n project_root / \"src\" / \"calibration\" / \"__init__.py\",\n project_root / \"src\" / \"explainability\" / \"__init__.py\",\n project_root / \"src\" / \"routing\" / \"__init__.py\",\n project_root / \"src\" / \"evaluation\" / \"__init__.py\",\n project_root / \"config\" / \"__init__.py\",\n]\n\nfor path in init_files:\n path.touch(exist_ok=True)\n\nprint(\"Package files recreated ✅\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.807471Z","iopub.status.idle":"2026-09-26T01:55:11.80783Z","shell.execute_reply.started":"2026-09-26T01:55:11.807647Z","shell.execute_reply":"2026-09-26T01:55:11.807669Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%writefile /kaggle/working/ovarian-trustworthy-ai/src/preprocessing/metadata.py\n\nfrom pathlib import Path\nimport pandas as pd\n\n\nclass UBCOceanMetadata:\n\n def __init__(self, dataset_root: str):\n\n self.dataset_root = Path(dataset_root)\n\n self.train_csv = self.dataset_root / \"train.csv\"\n self.train_images_dir = self.dataset_root / \"train_images\"\n self.train_thumbnails_dir = self.dataset_root / \"train_thumbnails\"\n\n if not self.dataset_root.exists():\n raise FileNotFoundError(\n f\"Dataset root not found: {self.dataset_root}\"\n )\n\n if not self.train_csv.exists():\n raise FileNotFoundError(\n f\"train.csv not found: {self.train_csv}\"\n )\n\n self.df = pd.read_csv(self.train_csv)\n\n self._validate_columns()\n\n\n def _validate_columns(self):\n\n required_columns = {\n \"image_id\",\n \"label\",\n \"image_width\",\n \"image_height\",\n \"is_tma\"\n }\n\n missing = required_columns - set(self.df.columns)\n\n if missing:\n raise ValueError(\n f\"Missing required columns: {sorted(missing)}\"\n )\n\n\n def get_record(self, image_id):\n\n rows = self.df[self.df[\"image_id\"] == image_id]\n\n if rows.empty:\n raise ValueError(\n f\"Image ID {image_id} not found in train.csv\"\n )\n\n row = rows.iloc[0]\n\n return {\n \"image_id\": int(row[\"image_id\"]),\n \"label\": str(row[\"label\"]),\n \"width\": int(row[\"image_width\"]),\n \"height\": int(row[\"image_height\"]),\n \"is_tma\": bool(row[\"is_tma\"])\n }\n\n\n def get_image_path(self, image_id):\n\n return self.train_images_dir / f\"{image_id}.png\"\n\n\n def get_thumbnail_path(self, image_id):\n\n record = self.get_record(image_id)\n\n # TMA slides do not have official thumbnails\n if record[\"is_tma\"]:\n return None\n\n return self.train_thumbnails_dir / f\"{image_id}_thumbnail.png\"\n\n\n def get_all_records(self):\n\n records = []\n\n for _, row in self.df.iterrows():\n\n records.append({\n \"image_id\": int(row[\"image_id\"]),\n \"label\": str(row[\"label\"]),\n \"width\": int(row[\"image_width\"]),\n \"height\": int(row[\"image_height\"]),\n \"is_tma\": bool(row[\"is_tma\"])\n })\n\n return records","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.809544Z","iopub.status.idle":"2026-09-26T01:55:11.809902Z","shell.execute_reply.started":"2026-09-26T01:55:11.809722Z","shell.execute_reply":"2026-09-26T01:55:11.809744Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.metadata import UBCOceanMetadata\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nmetadata = UBCOceanMetadata(dataset_root)\n\nprint(\"Number of records:\", len(metadata.df))\n\nrecord = metadata.get_record(4)\n\nprint(\"\\nImage 4:\")\nprint(record)\n\nprint(\"\\nImage path:\")\nprint(metadata.get_image_path(4))\n\nprint(\"\\nThumbnail path:\")\nprint(metadata.get_thumbnail_path(4))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.811005Z","iopub.status.idle":"2026-09-26T01:55:11.811394Z","shell.execute_reply.started":"2026-09-26T01:55:11.8112Z","shell.execute_reply":"2026-09-26T01:55:11.811226Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%writefile /kaggle/working/ovarian-trustworthy-ai/src/preprocessing/image_reader.py\n\nfrom pathlib import Path\nimport numpy as np\nimport pyvips\n\nfrom .metadata import UBCOceanMetadata\n\n\nclass UBCOceanImageReader:\n\n def __init__(self, dataset_root: str):\n\n self.metadata = UBCOceanMetadata(dataset_root)\n\n\n def open_image(self, image_id):\n\n image_path = self.metadata.get_image_path(image_id)\n\n if not image_path.exists():\n raise FileNotFoundError(\n f\"Image not found: {image_path}\"\n )\n\n return pyvips.Image.new_from_file(\n str(image_path),\n access=\"random\"\n )\n\n\n def get_region(\n self,\n image_id,\n x,\n y,\n width,\n height\n ):\n\n image = self.open_image(image_id)\n\n image_width = image.width\n image_height = image.height\n\n if x < 0 or y < 0:\n raise ValueError(\"x and y must be >= 0\")\n\n if width <= 0 or height <= 0:\n raise ValueError(\"width and height must be > 0\")\n\n if x + width > image_width:\n raise ValueError(\n f\"Requested region exceeds image width. \"\n f\"x={x}, width={width}, image_width={image_width}\"\n )\n\n if y + height > image_height:\n raise ValueError(\n f\"Requested region exceeds image height. \"\n f\"y={y}, height={height}, image_height={image_height}\"\n )\n\n region = image.crop(\n x,\n y,\n width,\n height\n )\n\n array = np.ndarray(\n buffer=region.write_to_memory(),\n dtype=np.uint8,\n shape=(region.height, region.width, region.bands)\n )\n\n # Keep RGB channels only\n if array.shape[2] > 3:\n array = array[:, :, :3]\n\n return array\n\n\n def get_image_dimensions(self, image_id):\n\n image = self.open_image(image_id)\n\n return image.width, image.height","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.812574Z","iopub.status.idle":"2026-09-26T01:55:11.812913Z","shell.execute_reply.started":"2026-09-26T01:55:11.812776Z","shell.execute_reply":"2026-09-26T01:55:11.812802Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!pip install -q \"pyvips[binary]\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.814795Z","iopub.status.idle":"2026-09-26T01:55:11.815149Z","shell.execute_reply.started":"2026-09-26T01:55:11.814976Z","shell.execute_reply":"2026-09-26T01:55:11.815Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pyvips\n\nprint(\"pyvips version:\", pyvips.__version__)\nprint(\"libvips version:\", pyvips.version(0))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.8162Z","iopub.status.idle":"2026-09-26T01:55:11.816487Z","shell.execute_reply.started":"2026-09-26T01:55:11.81635Z","shell.execute_reply":"2026-09-26T01:55:11.81637Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.image_reader import UBCOceanImageReader\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nreader = UBCOceanImageReader(dataset_root)\n\nwidth, height = reader.get_image_dimensions(4)\n\nprint(\"Image 4 dimensions:\")\nprint(\"Width :\", width)\nprint(\"Height:\", height)\n\ntile = reader.get_region(\n image_id=4,\n x=8960,\n y=4608,\n width=256,\n height=256\n)\n\nprint(\"\\nRegion:\")\nprint(\"Shape :\", tile.shape)\nprint(\"Dtype :\", tile.dtype)\nprint(\"Range :\", tile.min(), \"to\", tile.max())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.817741Z","iopub.status.idle":"2026-09-26T01:55:11.818099Z","shell.execute_reply.started":"2026-09-26T01:55:11.817956Z","shell.execute_reply":"2026-09-26T01:55:11.817978Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%writefile /kaggle/working/ovarian-trustworthy-ai/src/preprocessing/slide_type.py\n\nfrom dataclasses import dataclass\n\n\n@dataclass\nclass SlideInfo:\n image_id: int\n slide_type: str\n width: int\n height: int\n magnification: str\n\n\nclass SlideTypeHandler:\n\n def __init__(self, metadata):\n self.metadata = metadata\n\n\n def get_slide_info(self, image_id):\n\n record = self.metadata.get_record(image_id)\n\n if record[\"is_tma\"]:\n slide_type = \"TMA\"\n magnification = \"40x\"\n else:\n slide_type = \"WSI\"\n magnification = \"20x\"\n\n return SlideInfo(\n image_id=record[\"image_id\"],\n slide_type=slide_type,\n width=record[\"width\"],\n height=record[\"height\"],\n magnification=magnification\n )\n\n\n def is_wsi(self, image_id):\n\n return not self.metadata.get_record(image_id)[\"is_tma\"]\n\n\n def is_tma(self, image_id):\n\n return self.metadata.get_record(image_id)[\"is_tma\"]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.81969Z","iopub.status.idle":"2026-09-26T01:55:11.820039Z","shell.execute_reply.started":"2026-09-26T01:55:11.819891Z","shell.execute_reply":"2026-09-26T01:55:11.819921Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.metadata import UBCOceanMetadata\nfrom src.preprocessing.slide_type import SlideTypeHandler\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nmetadata = UBCOceanMetadata(dataset_root)\nslide_handler = SlideTypeHandler(metadata)\n\n# WSI test\nwsi_info = slide_handler.get_slide_info(4)\n\nprint(\"Image 4:\")\nprint(wsi_info)\n\n# TMA test\ntma_info = slide_handler.get_slide_info(91)\n\nprint(\"\\nImage 91:\")\nprint(tma_info)\n\nprint(\"\\nType checks:\")\nprint(\"Image 4 is WSI:\", slide_handler.is_wsi(4))\nprint(\"Image 91 is TMA:\", slide_handler.is_tma(91))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.821457Z","iopub.status.idle":"2026-09-26T01:55:11.821821Z","shell.execute_reply.started":"2026-09-26T01:55:11.821624Z","shell.execute_reply":"2026-09-26T01:55:11.821645Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%writefile /kaggle/working/ovarian-trustworthy-ai/src/preprocessing/tile_extractor.py\n\nfrom dataclasses import dataclass\nfrom typing import List\n\nimport cv2\nimport numpy as np\n\nfrom .image_reader import UBCOceanImageReader\nfrom .slide_type import SlideTypeHandler\n\n\n@dataclass\nclass TileCoordinate:\n image_id: int\n x: int\n y: int\n source_size: int\n standardized_size: int\n slide_type: str\n\n\nclass TileExtractor:\n\n def __init__(self, reader, slide_handler):\n\n self.reader = reader\n self.slide_handler = slide_handler\n\n self.standardized_size = 256\n\n\n def get_tile_size(self, image_id):\n\n if self.slide_handler.is_tma(image_id):\n return 512\n\n return 256\n\n\n def generate_grid(self, image_id, stride=None):\n\n slide_info = self.slide_handler.get_slide_info(image_id)\n\n tile_size = self.get_tile_size(image_id)\n\n if stride is None:\n stride = tile_size\n\n coordinates = []\n\n for y in range(\n 0,\n slide_info.height - tile_size + 1,\n stride\n ):\n\n for x in range(\n 0,\n slide_info.width - tile_size + 1,\n stride\n ):\n\n coordinates.append(\n TileCoordinate(\n image_id=image_id,\n x=x,\n y=y,\n source_size=tile_size,\n standardized_size=self.standardized_size,\n slide_type=slide_info.slide_type\n )\n )\n\n return coordinates\n\n\n def extract_tile(self, coordinate):\n\n tile = self.reader.get_region(\n image_id=coordinate.image_id,\n x=coordinate.x,\n y=coordinate.y,\n width=coordinate.source_size,\n height=coordinate.source_size\n )\n\n if coordinate.source_size != coordinate.standardized_size:\n\n tile = cv2.resize(\n tile,\n (\n coordinate.standardized_size,\n coordinate.standardized_size\n ),\n interpolation=cv2.INTER_AREA\n )\n\n return tile\n\n\n def extract_tiles(self, image_id, max_tiles=None):\n\n coordinates = self.generate_grid(image_id)\n\n if max_tiles is not None:\n coordinates = coordinates[:max_tiles]\n\n tiles = []\n\n for coordinate in coordinates:\n\n tile = self.extract_tile(coordinate)\n\n tiles.append(\n (coordinate, tile)\n )\n\n return tiles","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.822897Z","iopub.status.idle":"2026-09-26T01:55:11.823241Z","shell.execute_reply.started":"2026-09-26T01:55:11.823044Z","shell.execute_reply":"2026-09-26T01:55:11.823058Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.metadata import UBCOceanMetadata\nfrom src.preprocessing.image_reader import UBCOceanImageReader\nfrom src.preprocessing.slide_type import SlideTypeHandler\nfrom src.preprocessing.tile_extractor import TileExtractor\n\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nmetadata = UBCOceanMetadata(dataset_root)\n\nreader = UBCOceanImageReader(dataset_root)\n\nslide_handler = SlideTypeHandler(metadata)\n\ntile_extractor = TileExtractor(\n reader,\n slide_handler\n)\n\n\n# Generate WSI tile coordinates\ncoordinates = tile_extractor.generate_grid(4)\n\nprint(\"Number of candidate tiles:\", len(coordinates))\n\nprint(\"\\nFirst 5 coordinates:\")\n\nfor coordinate in coordinates[:5]:\n print(coordinate)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.824417Z","iopub.status.idle":"2026-09-26T01:55:11.824666Z","shell.execute_reply.started":"2026-09-26T01:55:11.82455Z","shell.execute_reply":"2026-09-26T01:55:11.824563Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"test_tiles = tile_extractor.extract_tiles(\n image_id=4,\n max_tiles=3\n)\n\nprint(\"Extracted tiles:\", len(test_tiles))\n\nfor i, (coordinate, tile) in enumerate(test_tiles):\n\n print(\n f\"Tile {i+1}:\",\n \"coordinate=\", (coordinate.x, coordinate.y),\n \"shape=\", tile.shape,\n \"dtype=\", tile.dtype\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.82572Z","iopub.status.idle":"2026-09-26T01:55:11.826078Z","shell.execute_reply.started":"2026-09-26T01:55:11.825942Z","shell.execute_reply":"2026-09-26T01:55:11.825959Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tma_coordinates = tile_extractor.generate_grid(91)\n\nprint(\"Image 91 candidate tiles:\", len(tma_coordinates))\n\nprint(\"\\nFirst coordinate:\")\nprint(tma_coordinates[0])\n\ntma_tile = tile_extractor.extract_tile(\n tma_coordinates[0]\n)\n\nprint(\"\\nExtracted TMA tile:\")\nprint(\"Shape:\", tma_tile.shape)\nprint(\"Dtype:\", tma_tile.dtype)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.827309Z","iopub.status.idle":"2026-09-26T01:55:11.827542Z","shell.execute_reply.started":"2026-09-26T01:55:11.827433Z","shell.execute_reply":"2026-09-26T01:55:11.827447Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%writefile /kaggle/working/ovarian-trustworthy-ai/src/preprocessing/thumbnail_selector.py\n\nfrom pathlib import Path\n\nimport cv2\nimport numpy as np\nimport pyvips\n\nfrom .metadata import UBCOceanMetadata\nfrom .tile_extractor import TileCoordinate\n\n\nclass ThumbnailSelector:\n\n def __init__(\n self,\n metadata,\n tissue_threshold=0.20\n ):\n\n self.metadata = metadata\n self.tissue_threshold = tissue_threshold\n\n\n def _load_thumbnail(self, image_id):\n\n thumbnail_path = self.metadata.get_thumbnail_path(image_id)\n\n if thumbnail_path is None:\n return None\n\n if not thumbnail_path.exists():\n raise FileNotFoundError(\n f\"Thumbnail not found: {thumbnail_path}\"\n )\n\n image = pyvips.Image.new_from_file(\n str(thumbnail_path),\n access=\"random\"\n )\n\n array = np.ndarray(\n buffer=image.write_to_memory(),\n dtype=np.uint8,\n shape=(image.height, image.width, image.bands)\n )\n\n if array.shape[2] > 3:\n array = array[:, :, :3]\n\n return array\n\n\n def _create_tissue_mask(self, thumbnail):\n\n hsv = cv2.cvtColor(\n thumbnail,\n cv2.COLOR_RGB2HSV\n )\n\n saturation = hsv[:, :, 1]\n\n # Otsu automatically chooses a saturation threshold\n _, mask = cv2.threshold(\n saturation,\n 0,\n 255,\n cv2.THRESH_BINARY + cv2.THRESH_OTSU\n )\n\n # Remove small isolated regions\n kernel = np.ones(\n (5, 5),\n dtype=np.uint8\n )\n\n mask = cv2.morphologyEx(\n mask,\n cv2.MORPH_OPEN,\n kernel\n )\n\n mask = cv2.morphologyEx(\n mask,\n cv2.MORPH_CLOSE,\n kernel\n )\n\n return mask > 0\n\n\n def get_thumbnail_tissue_ratio(\n self,\n image_id,\n x,\n y,\n tile_size\n ):\n\n thumbnail = self._load_thumbnail(image_id)\n\n if thumbnail is None:\n # TMA has no official thumbnail\n return 1.0\n\n tissue_mask = self._create_tissue_mask(\n thumbnail\n )\n\n record = self.metadata.get_record(image_id)\n\n scale_x = thumbnail.shape[1] / record[\"width\"]\n scale_y = thumbnail.shape[0] / record[\"height\"]\n\n x1 = int(x * scale_x)\n y1 = int(y * scale_y)\n\n x2 = int((x + tile_size) * scale_x)\n y2 = int((y + tile_size) * scale_y)\n\n x1 = max(0, min(x1, thumbnail.shape[1]))\n x2 = max(0, min(x2, thumbnail.shape[1]))\n\n y1 = max(0, min(y1, thumbnail.shape[0]))\n y2 = max(0, min(y2, thumbnail.shape[0]))\n\n if x2 <= x1 or y2 <= y1:\n return 0.0\n\n region = tissue_mask[y1:y2, x1:x2]\n\n if region.size == 0:\n return 0.0\n\n return float(region.mean())\n\n\n def select_candidates(\n self,\n image_id,\n coordinates\n ):\n\n record = self.metadata.get_record(image_id)\n\n # TMA images do not have official thumbnails.\n # Keep their generated coordinates for M2 QA.\n if record[\"is_tma\"]:\n return list(coordinates)\n\n selected = []\n\n for coordinate in coordinates:\n\n ratio = self.get_thumbnail_tissue_ratio(\n image_id=image_id,\n x=coordinate.x,\n y=coordinate.y,\n tile_size=coordinate.source_size\n )\n\n if ratio >= self.tissue_threshold:\n selected.append(coordinate)\n\n return selected","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.828874Z","iopub.status.idle":"2026-09-26T01:55:11.829267Z","shell.execute_reply.started":"2026-09-26T01:55:11.829039Z","shell.execute_reply":"2026-09-26T01:55:11.829055Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.metadata import UBCOceanMetadata\nfrom src.preprocessing.thumbnail_selector import ThumbnailSelector\n\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nmetadata = UBCOceanMetadata(dataset_root)\n\nselector = ThumbnailSelector(\n metadata=metadata,\n tissue_threshold=0.20\n)\n\nratio = selector.get_thumbnail_tissue_ratio(\n image_id=4,\n x=8960,\n y=4608,\n tile_size=256\n)\n\nprint(\"Tissue ratio:\", ratio)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.830743Z","iopub.status.idle":"2026-09-26T01:55:11.831141Z","shell.execute_reply.started":"2026-09-26T01:55:11.830919Z","shell.execute_reply":"2026-09-26T01:55:11.830934Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from src.preprocessing.image_reader import UBCOceanImageReader\nfrom src.preprocessing.slide_type import SlideTypeHandler\nfrom src.preprocessing.tile_extractor import TileExtractor\n\nreader = UBCOceanImageReader(dataset_root)\n\nslide_handler = SlideTypeHandler(metadata)\n\ntile_extractor = TileExtractor(\n reader,\n slide_handler\n)\n\ncoordinates = tile_extractor.generate_grid(4)\n\nprint(\"Total candidate tiles:\", len(coordinates))\n\nselected_coordinates = selector.select_candidates(\n image_id=4,\n coordinates=coordinates\n)\n\nprint(\"Selected tissue candidates:\", len(selected_coordinates))\n\nremoved = len(coordinates) - len(selected_coordinates)\n\nprint(\"Removed background candidates:\", removed)\n\nprint(\n \"Percentage removed:\",\n round(100 * removed / len(coordinates), 2),\n \"%\"\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.832261Z","iopub.status.idle":"2026-09-26T01:55:11.83257Z","shell.execute_reply.started":"2026-09-26T01:55:11.832419Z","shell.execute_reply":"2026-09-26T01:55:11.832464Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tma_coordinates = tile_extractor.generate_grid(91)\n\ntma_selected = selector.select_candidates(\n image_id=91,\n coordinates=tma_coordinates\n)\n\nprint(\"TMA total:\", len(tma_coordinates))\nprint(\"TMA selected:\", len(tma_selected))\n\nprint(\n \"All TMA candidates preserved:\",\n len(tma_coordinates) == len(tma_selected)\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.833772Z","iopub.status.idle":"2026-09-26T01:55:11.834101Z","shell.execute_reply.started":"2026-09-26T01:55:11.833945Z","shell.execute_reply":"2026-09-26T01:55:11.833968Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import sys\nimport importlib\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nproject_root = \"/kaggle/working/ovarian-trustworthy-ai\"\n\nif project_root not in sys.path:\n sys.path.insert(0, project_root)\n\nimportlib.invalidate_caches()\n\nfrom src.preprocessing.image_reader import UBCOceanImageReader\nfrom src.preprocessing.stain_normalization import MacenkoNormalizer\n\ndataset_root = \"/kaggle/input/competitions/UBC-OCEAN\"\n\nreader = UBCOceanImageReader(dataset_root)\n\nimage_id = 4\n\ntile = reader.get_region(\n image_id=image_id,\n x=8960,\n y=4608,\n width=256,\n height=256\n)\n\nprint(\"Tile shape:\", tile.shape)\nprint(\"Tile dtype:\", tile.dtype)\nprint(\"Pixel range:\", tile.min(), \"to\", tile.max())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.835478Z","iopub.status.idle":"2026-09-26T01:55:11.835758Z","shell.execute_reply.started":"2026-09-26T01:55:11.835639Z","shell.execute_reply":"2026-09-26T01:55:11.835655Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"normalizer = MacenkoNormalizer(\n beta=0.15,\n alpha=1.0,\n max_samples=10000,\n io=255.0,\n black_threshold=30.0\n)\n\nnormalizer.fit([tile])\n\nprint(\"Reference stain matrix:\")\nprint(normalizer.reference.stain_matrix)\n\nprint(\"\\nReference maximum concentration:\")\nprint(normalizer.reference.max_concentration)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.83686Z","iopub.status.idle":"2026-09-26T01:55:11.837225Z","shell.execute_reply.started":"2026-09-26T01:55:11.837043Z","shell.execute_reply":"2026-09-26T01:55:11.837068Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Get several tissue tiles from image 4\nreference_coordinates = [\n (8960, 4608),\n (9216, 4608),\n (9472, 4608),\n (9728, 4608),\n (8960, 4864),\n]\n\nreference_tiles = []\n\nfor x, y in reference_coordinates:\n tile = reader.get_region(\n image_id=4,\n x=x,\n y=y,\n width=256,\n height=256\n )\n reference_tiles.append(tile)\n\nprint(\"Number of reference tiles:\", len(reference_tiles))\nprint(\"Tile shapes:\", [t.shape for t in reference_tiles])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.838568Z","iopub.status.idle":"2026-09-26T01:55:11.838818Z","shell.execute_reply.started":"2026-09-26T01:55:11.838707Z","shell.execute_reply":"2026-09-26T01:55:11.838721Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(normalizer.__dict__.keys())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.840033Z","iopub.status.idle":"2026-09-26T01:55:11.840427Z","shell.execute_reply.started":"2026-09-26T01:55:11.840218Z","shell.execute_reply":"2026-09-26T01:55:11.840245Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"for key, value in normalizer.__dict__.items():\n print(key, type(value), getattr(value, \"shape\", None))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.841454Z","iopub.status.idle":"2026-09-26T01:55:11.841669Z","shell.execute_reply.started":"2026-09-26T01:55:11.841561Z","shell.execute_reply":"2026-09-26T01:55:11.841573Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(normalizer.reference)\nprint(normalizer.reference.__dict__)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.842542Z","iopub.status.idle":"2026-09-26T01:55:11.84277Z","shell.execute_reply.started":"2026-09-26T01:55:11.842663Z","shell.execute_reply":"2026-09-26T01:55:11.842677Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(normalizer.reference.__dict__.keys())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.843678Z","iopub.status.idle":"2026-09-26T01:55:11.843946Z","shell.execute_reply.started":"2026-09-26T01:55:11.843829Z","shell.execute_reply":"2026-09-26T01:55:11.843844Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"normalizer.fit(reference_tiles)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.845673Z","iopub.status.idle":"2026-09-26T01:55:11.846006Z","shell.execute_reply.started":"2026-09-26T01:55:11.84583Z","shell.execute_reply":"2026-09-26T01:55:11.845852Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import numpy as np\n\ntile = reference_tiles[0]\n\nod = normalizer.rgb_to_od(tile)\n\nmask = normalizer._get_stain_mask(tile, od)\n\nprint(\"Total pixels:\", tile.shape[0] * tile.shape[1])\nprint(\"Foreground pixels:\", mask.sum())\nprint(\"Foreground ratio:\", mask.mean())\n\nstain_matrix = normalizer._estimate_stain_matrix(tile)\n\nprint(\"\\nEstimated stain matrix:\")\nprint(stain_matrix)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.847213Z","iopub.status.idle":"2026-09-26T01:55:11.847568Z","shell.execute_reply.started":"2026-09-26T01:55:11.847395Z","shell.execute_reply":"2026-09-26T01:55:11.847417Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"\\nAngle between stain vectors:\")\n\nv1 = stain_matrix[0]\nv2 = stain_matrix[1]\n\ncosine = np.dot(v1, v2) / (\n np.linalg.norm(v1) * np.linalg.norm(v2)\n)\n\nangle = np.degrees(np.arccos(np.clip(cosine, -1, 1)))\n\nprint(angle, \"degrees\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.849367Z","iopub.status.idle":"2026-09-26T01:55:11.849663Z","shell.execute_reply.started":"2026-09-26T01:55:11.849496Z","shell.execute_reply":"2026-09-26T01:55:11.84951Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import inspect\n\nprint(inspect.getsource(normalizer._estimate_stain_matrix))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.850487Z","iopub.status.idle":"2026-09-26T01:55:11.850851Z","shell.execute_reply.started":"2026-09-26T01:55:11.850665Z","shell.execute_reply":"2026-09-26T01:55:11.85069Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(inspect.getsource(normalizer._get_stain_mask))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.852257Z","iopub.status.idle":"2026-09-26T01:55:11.852614Z","shell.execute_reply.started":"2026-09-26T01:55:11.852429Z","shell.execute_reply":"2026-09-26T01:55:11.852453Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"od = normalizer.rgb_to_od(tile)\n\nmask = normalizer._get_stain_mask(tile, od)\nod_pixels = od.reshape(-1, 3)[mask]\n\nod_pixels = normalizer._sample_pixels(od_pixels)\n\n# SVD\n_, singular_values, vh = np.linalg.svd(\n od_pixels,\n full_matrices=False\n)\n\nv = vh[:2].T\n\nprojected = od_pixels @ v\n\nphi = np.arctan2(\n projected[:, 1],\n projected[:, 0]\n)\n\nprint(\"Number of OD pixels:\", len(od_pixels))\n\nprint(\"\\nSingular values:\")\nprint(singular_values)\n\nprint(\"\\nOD min:\")\nprint(od_pixels.min(axis=0))\n\nprint(\"\\nOD max:\")\nprint(od_pixels.max(axis=0))\n\nprint(\"\\nOD mean:\")\nprint(od_pixels.mean(axis=0))\n\nprint(\"\\nPhi min/max:\")\nprint(phi.min(), phi.max())\n\nprint(\"\\nPhi percentiles:\")\nprint(np.percentile(phi, [0, 1, 5, 25, 50, 75, 95, 99, 100]))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.853725Z","iopub.status.idle":"2026-09-26T01:55:11.854156Z","shell.execute_reply.started":"2026-09-26T01:55:11.85395Z","shell.execute_reply":"2026-09-26T01:55:11.853978Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Check whether the angle distribution is wrapping around -pi / +pi\n\nphi_shifted = np.mod(phi, 2 * np.pi)\n\nprint(\"Original phi percentiles:\")\nprint(np.percentile(phi, [1, 5, 50, 95, 99]))\n\nprint(\"\\nShifted phi percentiles:\")\nprint(np.percentile(phi_shifted, [1, 5, 50, 95, 99]))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:11.855588Z","iopub.status.idle":"2026-09-26T01:55:11.856097Z","shell.execute_reply.started":"2026-09-26T01:55:11.85592Z","shell.execute_reply":"2026-09-26T01:55:11.85595Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"SVD basis V:\")\nprint(v)\n\nprint(\"\\nProjected coordinate ranges:\")\nprint(\"X:\", projected[:, 0].min(), \"to\", projected[:, 0].max())\nprint(\"Y:\", projected[:, 1].min(), \"to\", projected[:, 1].max())\n\nprint(\"\\nProjected coordinate means:\")\nprint(projected.mean(axis=0))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-26T01:55:32.177324Z","iopub.execute_input":"2026-09-26T01:55:32.177581Z","iopub.status.idle":"2026-09-26T01:55:32.184323Z","shell.execute_reply.started":"2026-09-26T01:55:32.177558Z","shell.execute_reply":"2026-09-26T01:55:32.183384Z"}},"outputs":[],"execution_count":null}]}