{"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":"markdown","source":"
\n
\n

Published on September 08, 2026

\n

Author: Jocelyn C. Dumlao

\n

\n LinkedIn\n GitHub\n YouTube\n Kaggle\n
\n
","metadata":{}},{"cell_type":"markdown","source":"#

Setup

\n","metadata":{}},{"cell_type":"code","source":"# ============================================================\n# 0. STRICT OFFLINE MODE\n# MUST RUN BEFORE importing timm / huggingface_hub\n# ============================================================\n\nimport os\n\nos.environ[\"HF_HUB_OFFLINE\"] = \"1\"\nos.environ[\"HF_DATASETS_OFFLINE\"] = \"1\"\nos.environ[\"TRANSFORMERS_OFFLINE\"] = \"1\"\n\nos.environ[\"HF_HUB_DISABLE_TELEMETRY\"] = \"1\"\nos.environ[\"HF_HUB_DISABLE_XET\"] = \"1\"\nos.environ[\"HF_HUB_ENABLE_HF_TRANSFER\"] = \"0\"\n\nprint(\"=\" * 70)\nprint(\"OFFLINE MODE ENABLED\")\nprint(\"=\" * 70)\nprint(\"HF_HUB_OFFLINE =\", os.environ[\"HF_HUB_OFFLINE\"])\nprint(\"HF_DATASETS_OFFLINE =\", os.environ[\"HF_DATASETS_OFFLINE\"])\nprint(\"TRANSFORMERS_OFFLINE =\", os.environ[\"TRANSFORMERS_OFFLINE\"])\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-14T08:01:45.939961Z","iopub.execute_input":"2026-09-14T08:01:45.940858Z","iopub.status.idle":"2026-09-14T08:01:45.958812Z","shell.execute_reply.started":"2026-09-14T08:01:45.940822Z","shell.execute_reply":"2026-09-14T08:01:45.958014Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ============================================================\n# 1. OFFLINE ENVIRONMENT — MUST COME FIRST\n# ============================================================\n\nimport os\n\n# ------------------------------------------------------------\n# Hugging Face / Transformers offline mode\n# MUST be configured before importing timm.\n# ------------------------------------------------------------\n\nos.environ[\"HF_HUB_OFFLINE\"] = \"1\"\nos.environ[\"TRANSFORMERS_OFFLINE\"] = \"1\"\nos.environ[\"HF_HUB_DISABLE_TELEMETRY\"] = \"1\"\n\n# Disable Hugging Face Xet/network-related behavior.\nos.environ[\"HF_HUB_DISABLE_XET\"] = \"1\"\n\n# Avoid token/network lookup.\nos.environ[\"HF_HUB_DISABLE_IMPLICIT_TOKEN\"] = \"1\"\n\n# Force local cache locations where possible.\nos.environ.setdefault(\n \"HF_HOME\",\n \"/kaggle/working/huggingface\"\n)\n\nos.environ.setdefault(\n \"TRANSFORMERS_CACHE\",\n \"/kaggle/working/huggingface\"\n)\n\nos.environ.setdefault(\n \"HF_DATASETS_OFFLINE\",\n \"1\"\n)\n\nprint(\"=\" * 70)\nprint(\"STRICT OFFLINE MODE\")\nprint(\"=\" * 70)\n\nprint(\n \"HF_HUB_OFFLINE:\",\n os.environ.get(\"HF_HUB_OFFLINE\")\n)\n\nprint(\n \"TRANSFORMERS_OFFLINE:\",\n os.environ.get(\"TRANSFORMERS_OFFLINE\")\n)\n\nprint(\n \"HF_HUB_DISABLE_XET:\",\n os.environ.get(\"HF_HUB_DISABLE_XET\")\n)\n\nprint()\n\n\n# ============================================================\n# 2. IMPORTS\n# ============================================================\n\nimport gc\nimport math\nimport pickle\nimport time\nimport warnings\n\nfrom pathlib import Path\n\nimport cv2\nimport numpy as np\nimport pandas as pd\nimport pydicom\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n# ------------------------------------------------------------\n# timm is imported ONLY AFTER offline environment variables\n# have been configured.\n# ------------------------------------------------------------\n\nimport timm\n\nfrom pydicom.pixel_data_handlers.util import (\n apply_modality_lut,\n)\n\nwarnings.filterwarnings(\"ignore\")\n\n\n# ============================================================\n# 3. ENVIRONMENT\n# ============================================================\n\nDEVICE = torch.device(\n \"cuda\"\n if torch.cuda.is_available()\n else \"cpu\"\n)\n\nif DEVICE.type == \"cuda\":\n\n torch.backends.cudnn.benchmark = True\n\n try:\n\n torch.backends.cuda.matmul.allow_tf32 = True\n torch.backends.cudnn.allow_tf32 = True\n\n except Exception:\n\n pass\n\n\nprint(\"=\" * 70)\nprint(\"RSNA KNEE ABNORMALITY — OFFLINE V6\")\nprint(\"=\" * 70)\n\nprint(\n f\"Device: {DEVICE}\"\n)\n\nif DEVICE.type == \"cuda\":\n\n print(\n f\"GPU: \"\n f\"{torch.cuda.get_device_name(0)}\"\n )\n\nprint()\n\n\n# ============================================================\n# 4. CONFIGURATION\n# ============================================================\n\nIMAGE_SIZE = 384\n\nCROP_MM = 140.0\n\nTOTAL_SLICES = 64\n\nDEFAULT_ARCH = (\n \"coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k\"\n)\n\nDEFAULT_RESOLUTION = 384\n\nOUTPUT_FILE = (\n \"/kaggle/working/submission.csv\"\n)\n\nCACHE_FILE = (\n \"/kaggle/working/\"\n \"rsna_knee_study_lookup_cache_v6.pkl\"\n)\n\nCACHE_VERSION = (\n \"rsna_knee_lookup_v6_strict_offline\"\n)\n\nPROGRESS_EVERY_STUDIES = 25\n\nPROGRESS_EVERY_FILES = 1000\n\n# ------------------------------------------------------------\n# Optional local checkpoint.\n#\n# If you later attach a trained RSNA checkpoint as a Kaggle\n# Dataset, put its path here.\n#\n# Example:\n#\n# MODEL_CHECKPOINT = (\n# \"/kaggle/input/my-rsna-model/\"\n# \"best_model.pth\"\n# )\n#\n# Leave as None for checkpoint-free mode.\n# ------------------------------------------------------------\n\nMODEL_CHECKPOINT = None\n\n\n# ============================================================\n# 5. LABELS\n# ============================================================\n\nLABELS = [\n\n \"ACL\",\n\n \"MCL\",\n\n \"Medial Meniscus\",\n\n \"Lateral Meniscus\",\n\n \"Medial OA\",\n\n \"Lateral OA\",\n\n \"PF OA\",\n\n \"Effusion\",\n\n \"Synovitis\",\n\n \"Baker's\",\n\n \"Contusion\",\n\n \"Fracture\",\n]\n\nNUM_CLASSES = len(\n LABELS\n)\n\n\n# ============================================================\n# 6. IMAGENET NORMALIZATION\n# ============================================================\n\nIMAGENET_MEAN = np.array(\n [\n 0.485,\n 0.456,\n 0.406,\n ],\n dtype=np.float32,\n)\n\nIMAGENET_STD = np.array(\n [\n 0.229,\n 0.224,\n 0.225,\n ],\n dtype=np.float32,\n)\n\n\n# ============================================================\n# 7. ANATOMICAL SERIES SLOTS\n# ============================================================\n\nSERIES_SLOTS = [\n\n (\n \"Sagittal\",\n 1,\n 18,\n ),\n\n (\n \"Sagittal\",\n 0,\n 14,\n ),\n\n (\n \"Coronal\",\n 1,\n 12,\n ),\n\n (\n \"Coronal\",\n 0,\n 8,\n ),\n\n (\n \"Axial\",\n -1,\n 12,\n ),\n]\n\n\n# ============================================================\n# 8. TTA CONFIGURATIONS\n# ============================================================\n\nTTA_CONFIGS = [\n\n {\n \"name\": \"wide_02_98\",\n \"lo\": 0.02,\n \"hi\": 0.98,\n \"k\": 64,\n \"weight\": 0.28,\n },\n\n {\n \"name\": \"wide_04_96\",\n \"lo\": 0.04,\n \"hi\": 0.96,\n \"k\": 64,\n \"weight\": 0.24,\n },\n\n {\n \"name\": \"balanced_06_94\",\n \"lo\": 0.06,\n \"hi\": 0.94,\n \"k\": 60,\n \"weight\": 0.20,\n },\n\n {\n \"name\": \"central_10_90\",\n \"lo\": 0.10,\n \"hi\": 0.90,\n \"k\": 56,\n \"weight\": 0.16,\n },\n\n {\n \"name\": \"focused_15_85\",\n \"lo\": 0.15,\n \"hi\": 0.85,\n \"k\": 48,\n \"weight\": 0.12,\n },\n]\n\n\n# ============================================================\n# 9. TARGET RANK WEIGHTS\n# ============================================================\n\nTARGET_RANK_WEIGHTS = {\n\n \"ACL\": 0.70,\n\n \"MCL\": 0.55,\n\n \"Medial Meniscus\": 0.65,\n\n \"Lateral Meniscus\": 0.72,\n\n \"Medial OA\": 0.60,\n\n \"Lateral OA\": 0.68,\n\n \"PF OA\": 0.58,\n\n \"Effusion\": 0.52,\n\n \"Synovitis\": 0.58,\n\n \"Baker's\": 0.55,\n\n \"Contusion\": 0.60,\n\n \"Fracture\": 0.70,\n}\n\n\n# ============================================================\n# 10. LOCATE COMPETITION ROOT\n# ============================================================\n\ndef locate_competition_root():\n\n candidates = [\n\n Path(\n \"/kaggle/input/competitions/\"\n \"rsna-knee-abnormality-detection\"\n ),\n\n Path(\n \"/kaggle/input/\"\n \"rsna-knee-abnormality-detection\"\n ),\n\n Path(\"data\"),\n\n Path(\".\"),\n ]\n\n for root in candidates:\n\n if not root.exists():\n\n continue\n\n if (\n\n (root / \"test.csv\").is_file()\n\n and\n\n (\n root /\n \"test_series.csv\"\n ).is_file()\n\n and\n\n (\n root /\n \"test_series\"\n ).is_dir()\n\n ):\n\n print(\n \"Using competition root:\"\n )\n\n print(\n f\" {root}\"\n )\n\n return root\n\n # --------------------------------------------------------\n # Fallback search.\n # --------------------------------------------------------\n\n input_root = Path(\n \"/kaggle/input\"\n )\n\n if input_root.exists():\n\n print(\n \"Searching /kaggle/input \"\n \"for test.csv + test_series...\"\n )\n\n for csv_path in input_root.glob(\n \"**/test.csv\"\n ):\n\n root = (\n csv_path.parent\n )\n\n if (\n\n (\n root /\n \"test_series\"\n ).is_dir()\n\n and\n\n (\n root /\n \"test_series.csv\"\n ).is_file()\n\n ):\n\n print(\n \"Discovered competition root:\"\n )\n\n print(\n f\" {root}\"\n )\n\n return root\n\n raise FileNotFoundError(\n\n \"\\nCould not locate the RSNA competition root.\\n\"\n\n \"\\nExpected:\\n\"\n\n \" test.csv\\n\"\n\n \" test_series.csv\\n\"\n\n \" test_series/\\n\"\n )\n\n\n# ============================================================\n# 11. LOCATE TEST SERIES\n# ============================================================\n\ndef locate_test_series(root):\n\n candidates = [\n\n root / \"test_series\",\n\n root / \"test_images\",\n\n root / \"test\",\n\n root / \"images\" / \"test\",\n ]\n\n official = (\n root /\n \"test_series\"\n )\n\n if official.is_dir():\n\n print(\n \"Using official test DICOM directory:\"\n )\n\n print(\n f\" {official}\"\n )\n\n return official\n\n for path in candidates:\n\n if path.is_dir():\n\n print(\n \"Using discovered test DICOM directory:\"\n )\n\n print(\n f\" {path}\"\n )\n\n return path\n\n raise FileNotFoundError(\n\n \"\\nCould not locate test DICOM directory.\\n\"\n\n \"\\nExpected:\\n\"\n\n f\" {root / 'test_series'}\\n\"\n )\n\n\n# ============================================================\n# 12. LOAD TEST CSV\n# ============================================================\n\ndef load_test_csv(root):\n\n test_csv = (\n root /\n \"test.csv\"\n )\n\n if not test_csv.is_file():\n\n matches = list(\n root.glob(\n \"**/test.csv\"\n )\n )\n\n if not matches:\n\n raise FileNotFoundError(\n \"test.csv not found.\"\n )\n\n test_csv = matches[0]\n\n df = pd.read_csv(\n test_csv\n )\n\n if df.empty:\n\n raise ValueError(\n \"test.csv is empty.\"\n )\n\n if (\n \"StudyInstanceUID\"\n not in df.columns\n ):\n\n raise ValueError(\n \"test.csv does not contain \"\n \"'StudyInstanceUID'.\"\n )\n\n print(\n \"Test CSV:\"\n )\n\n print(\n f\" {test_csv}\"\n )\n\n print(\n f\"Test studies: \"\n f\"{len(df):,}\"\n )\n\n return (\n df,\n \"StudyInstanceUID\",\n test_csv,\n )\n\n\n# ============================================================\n# 13. LOAD TEST SERIES CSV\n# ============================================================\n\ndef load_test_series_csv(root):\n\n test_series_csv = (\n root /\n \"test_series.csv\"\n )\n\n if not test_series_csv.is_file():\n\n matches = list(\n root.glob(\n \"**/test_series.csv\"\n )\n )\n\n if not matches:\n\n raise FileNotFoundError(\n \"test_series.csv not found.\"\n )\n\n test_series_csv = matches[0]\n\n df = pd.read_csv(\n test_series_csv\n )\n\n required = [\n\n \"StudyInstanceUID\",\n\n \"SeriesInstanceUID\",\n\n \"Fluid_Sensitive\",\n\n \"Anatomical_Plane\",\n ]\n\n missing = [\n\n col\n\n for col in required\n\n if col not in df.columns\n ]\n\n if missing:\n\n raise ValueError(\n\n \"test_series.csv is missing \"\n\n f\"columns: {missing}\"\n )\n\n print(\n \"Test series CSV:\"\n )\n\n print(\n f\" {test_series_csv}\"\n )\n\n print(\n f\"Test series: \"\n f\"{len(df):,}\"\n )\n\n print()\n\n print(\n \"Series metadata columns:\"\n )\n\n print(\n f\" {list(df.columns)}\"\n )\n\n return (\n df,\n test_series_csv,\n )\n\n\n# ============================================================\n# 14. DICOM HEADER TAGS\n# ============================================================\n\nDICOM_TAGS = [\n\n \"StudyInstanceUID\",\n\n \"SeriesInstanceUID\",\n\n \"SOPInstanceUID\",\n\n \"SeriesDescription\",\n\n \"ProtocolName\",\n\n \"SequenceName\",\n\n \"ImageOrientationPatient\",\n\n \"ImagePositionPatient\",\n\n \"PixelSpacing\",\n\n \"InstanceNumber\",\n\n \"SliceThickness\",\n\n \"SpacingBetweenSlices\",\n\n \"PatientPosition\",\n\n \"Modality\",\n\n \"BodyPartExamined\",\n]\n\n\n# ============================================================\n# 15. SAFE HELPERS\n# ============================================================\n\ndef safe_float_array(value):\n\n if value is None:\n\n return None\n\n try:\n\n return np.asarray(\n\n [\n float(x)\n for x in value\n ],\n\n dtype=np.float32,\n )\n\n except Exception:\n\n return None\n\n\ndef safe_float(\n value,\n default=None,\n):\n\n try:\n\n return float(value)\n\n except Exception:\n\n return default\n\n\ndef safe_int(\n value,\n default=None,\n):\n\n try:\n\n return int(value)\n\n except Exception:\n\n return default\n\n\ndef get_string(\n ds,\n name,\n):\n\n try:\n\n value = getattr(\n ds,\n name,\n None,\n )\n\n if value is None:\n\n return \"\"\n\n return str(value)\n\n except Exception:\n\n return \"\"\n\n\n# ============================================================\n# 16. ESTIMATE PLANE\n# ============================================================\n\ndef estimate_plane(ds):\n\n try:\n\n iop = safe_float_array(\n\n getattr(\n\n ds,\n\n \"ImageOrientationPatient\",\n\n None,\n )\n )\n\n if (\n\n iop is None\n\n or\n\n len(iop) != 6\n\n ):\n\n return \"\"\n\n row = iop[:3]\n\n col = iop[3:]\n\n normal = np.cross(\n row,\n col,\n )\n\n axis = int(\n np.argmax(\n np.abs(normal)\n )\n )\n\n if axis == 0:\n\n return \"Sagittal\"\n\n if axis == 1:\n\n return \"Coronal\"\n\n if axis == 2:\n\n return \"Axial\"\n\n except Exception:\n\n pass\n\n return \"\"\n\n\n# ============================================================\n# 17. READ LIGHT DICOM METADATA\n# ============================================================\n\ndef read_dicom_metadata(path):\n\n try:\n\n ds = pydicom.dcmread(\n\n str(path),\n\n stop_before_pixels=True,\n\n specific_tags=DICOM_TAGS,\n\n force=True,\n )\n\n iop = safe_float_array(\n\n getattr(\n\n ds,\n\n \"ImageOrientationPatient\",\n\n None,\n )\n )\n\n ipp = safe_float_array(\n\n getattr(\n\n ds,\n\n \"ImagePositionPatient\",\n\n None,\n )\n )\n\n spacing = safe_float_array(\n\n getattr(\n\n ds,\n\n \"PixelSpacing\",\n\n None,\n )\n )\n\n if (\n\n spacing is None\n\n or\n\n len(spacing) < 2\n\n ):\n\n spacing = np.array(\n\n [\n 0.5,\n 0.5,\n ],\n\n dtype=np.float32,\n )\n\n return {\n\n \"path\":\n str(path),\n\n \"study_uid\":\n get_string(\n ds,\n \"StudyInstanceUID\",\n ),\n\n \"series_uid\":\n get_string(\n ds,\n \"SeriesInstanceUID\",\n ),\n\n \"sop_uid\":\n get_string(\n ds,\n \"SOPInstanceUID\",\n ),\n\n \"series_description\":\n get_string(\n ds,\n \"SeriesDescription\",\n ),\n\n \"protocol_name\":\n get_string(\n ds,\n \"ProtocolName\",\n ),\n\n \"sequence_name\":\n get_string(\n ds,\n \"SequenceName\",\n ),\n\n \"plane\":\n estimate_plane(ds),\n\n \"image_orientation\":\n (\n iop.tolist()\n if iop is not None\n else None\n ),\n\n \"image_position\":\n (\n ipp.tolist()\n if ipp is not None\n else None\n ),\n\n \"pixel_spacing\":\n spacing[:2].tolist(),\n\n \"instance_number\":\n safe_int(\n getattr(\n ds,\n \"InstanceNumber\",\n None,\n )\n ),\n\n \"slice_thickness\":\n safe_float(\n getattr(\n ds,\n \"SliceThickness\",\n None,\n )\n ),\n\n \"spacing_between_slices\":\n safe_float(\n getattr(\n ds,\n \"SpacingBetweenSlices\",\n None,\n )\n ),\n\n \"patient_position\":\n get_string(\n ds,\n \"PatientPosition\",\n ),\n\n \"modality\":\n get_string(\n ds,\n \"Modality\",\n ),\n\n \"body_part\":\n get_string(\n ds,\n \"BodyPartExamined\",\n ),\n }\n\n except Exception as exc:\n\n return {\n\n \"error\":\n str(exc),\n\n \"path\":\n str(path),\n }\n\n\n# ============================================================\n# 18. SORT SERIES\n# ============================================================\n\ndef sort_series_records(records):\n\n if not records:\n\n return records\n\n orientation = None\n\n for record in records:\n\n value = record.get(\n \"image_orientation\"\n )\n\n if value is None:\n\n continue\n\n try:\n\n arr = np.asarray(\n\n value,\n\n dtype=np.float32,\n )\n\n if len(arr) == 6:\n\n orientation = arr\n\n break\n\n except Exception:\n\n pass\n\n if orientation is not None:\n\n try:\n\n row = orientation[:3]\n\n col = orientation[3:]\n\n normal = np.cross(\n row,\n col,\n )\n\n norm = np.linalg.norm(\n normal\n )\n\n if norm > 0:\n\n normal = (\n normal /\n norm\n )\n\n def projection(record):\n\n pos = record.get(\n \"image_position\"\n )\n\n if pos is None:\n\n return float(\n \"inf\"\n )\n\n try:\n\n return float(\n\n np.dot(\n\n np.asarray(\n pos,\n dtype=np.float32,\n ),\n\n normal,\n )\n )\n\n except Exception:\n\n return float(\n \"inf\"\n )\n\n return sorted(\n\n records,\n\n key=projection,\n )\n\n except Exception:\n\n pass\n\n return sorted(\n\n records,\n\n key=lambda x: (\n\n x.get(\n \"instance_number\"\n )\n\n if x.get(\n \"instance_number\"\n ) is not None\n\n else 10**9\n )\n )\n\n\n# ============================================================\n# 19. CACHE SIGNATURE\n# ============================================================\n\ndef make_cache_signature(\n test_csv,\n test_series_csv,\n test_studies,\n):\n\n signature = {\n\n \"version\":\n CACHE_VERSION,\n\n \"study_count\":\n len(test_studies),\n\n \"study_ids\":\n tuple(\n str(x)\n for x in test_studies\n ),\n }\n\n try:\n\n stat = test_csv.stat()\n\n signature[\n \"test_csv_size\"\n ] = stat.st_size\n\n signature[\n \"test_csv_mtime\"\n ] = stat.st_mtime_ns\n\n except Exception:\n\n pass\n\n try:\n\n stat = test_series_csv.stat()\n\n signature[\n \"test_series_csv_size\"\n ] = stat.st_size\n\n signature[\n \"test_series_csv_mtime\"\n ] = stat.st_mtime_ns\n\n except Exception:\n\n pass\n\n return signature\n\n\n# ============================================================\n# 20. LOAD CACHE\n# ============================================================\n\ndef load_lookup_cache(\n cache_file,\n signature,\n):\n\n path = Path(\n cache_file\n )\n\n if not path.exists():\n\n return None\n\n try:\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"CHECKING STUDY LOOKUP CACHE\"\n )\n\n print(\n \"=\" * 70\n )\n\n with open(\n path,\n \"rb\",\n ) as f:\n\n payload = pickle.load(\n f\n )\n\n if not isinstance(\n payload,\n dict,\n ):\n\n return None\n\n if (\n payload.get(\n \"signature\"\n )\n != signature\n ):\n\n print(\n \"Cache signature changed.\"\n )\n\n print(\n \"Rebuilding lookup...\"\n )\n\n return None\n\n lookup = payload.get(\n \"lookup\"\n )\n\n if lookup is None:\n\n return None\n\n print(\n \"Cache found:\"\n )\n\n print(\n f\" {path}\"\n )\n\n print(\n f\"Cached studies: \"\n f\"{len(lookup):,}\"\n )\n\n return lookup\n\n except Exception as exc:\n\n print(\n \"Cache could not be loaded:\"\n )\n\n print(\n f\" {exc}\"\n )\n\n return None\n\n\n# ============================================================\n# 21. SAVE CACHE\n# ============================================================\n\ndef save_lookup_cache(\n cache_file,\n signature,\n lookup,\n):\n\n path = Path(\n cache_file\n )\n\n try:\n\n payload = {\n\n \"signature\":\n signature,\n\n \"lookup\":\n lookup,\n }\n\n temp = path.with_suffix(\n \".tmp\"\n )\n\n with open(\n temp,\n \"wb\",\n ) as f:\n\n pickle.dump(\n\n payload,\n\n f,\n\n protocol=\n pickle.HIGHEST_PROTOCOL,\n )\n\n os.replace(\n temp,\n path,\n )\n\n print()\n print(\n \"Lookup cache saved:\"\n )\n\n print(\n f\" {path}\"\n )\n\n except Exception as exc:\n\n print(\n \"Warning: could not save cache:\"\n )\n\n print(\n f\" {exc}\"\n )\n\n\n# ============================================================\n# 22. DICOM FILE DISCOVERY\n# ============================================================\n\ndef discover_dicom_files(\n series_dir\n):\n\n if not series_dir.is_dir():\n\n return []\n\n files = []\n\n try:\n\n for path in series_dir.iterdir():\n\n if not path.is_file():\n\n continue\n\n if (\n path.suffix.lower()\n == \".dcm\"\n ):\n\n files.append(\n path\n )\n\n continue\n\n if path.suffix == \"\":\n\n files.append(\n path\n )\n\n except Exception:\n\n return []\n\n return sorted(\n\n files,\n\n key=lambda p: p.name,\n )\n\n\n# ============================================================\n# 23. CREATE SERIES LOOKUP\n# ============================================================\n\ndef create_series_lookup_optimized(\n test_series_root,\n test_studies,\n test_series_df,\n test_csv,\n test_series_csv,\n cache_file=CACHE_FILE,\n):\n\n test_studies = [\n\n str(x)\n\n for x in test_studies\n ]\n\n signature = make_cache_signature(\n\n test_csv,\n\n test_series_csv,\n\n test_studies,\n )\n\n cached = load_lookup_cache(\n\n cache_file,\n\n signature,\n )\n\n if cached is not None:\n\n return cached\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"BUILDING TEST SERIES LOOKUP\"\n )\n\n print(\n \"=\" * 70\n )\n\n print(\n f\"Studies: \"\n f\"{len(test_studies):,}\"\n )\n\n print(\n \"Test DICOM root:\"\n )\n\n print(\n f\" {test_series_root}\"\n )\n\n print()\n\n lookup = {}\n\n total_files = 0\n\n valid_files = 0\n\n bad_files = 0\n\n missing_studies = 0\n\n missing_series = 0\n\n start_time = time.time()\n\n metadata = (\n test_series_df.copy()\n )\n\n metadata[\n \"StudyInstanceUID\"\n ] = metadata[\n \"StudyInstanceUID\"\n ].astype(str)\n\n metadata[\n \"SeriesInstanceUID\"\n ] = metadata[\n \"SeriesInstanceUID\"\n ].astype(str)\n\n metadata[\n \"Fluid_Sensitive\"\n ] = pd.to_numeric(\n\n metadata[\n \"Fluid_Sensitive\"\n ],\n\n errors=\"coerce\",\n\n ).fillna(\n 0\n ).astype(\n int\n )\n\n metadata[\n \"Anatomical_Plane\"\n ] = metadata[\n \"Anatomical_Plane\"\n ].astype(str)\n\n for (\n study_index,\n study_id,\n ) in enumerate(\n\n test_studies,\n\n start=1,\n ):\n\n study_dir = (\n\n test_series_root /\n study_id\n )\n\n if not study_dir.is_dir():\n\n missing_studies += 1\n\n lookup[\n study_id\n ] = {}\n\n print(\n\n f\"[{study_index}/\"\n f\"{len(test_studies)}] \"\n f\"WARNING: missing study \"\n f\"{study_id}\"\n )\n\n continue\n\n study_meta = metadata[\n\n metadata[\n \"StudyInstanceUID\"\n ]\n ==\n study_id\n ]\n\n series_lookup = {}\n\n for _, row in (\n study_meta.iterrows()\n ):\n\n series_uid = str(\n\n row[\n \"SeriesInstanceUID\"\n ]\n )\n\n series_dir = (\n\n study_dir /\n series_uid\n )\n\n if not series_dir.is_dir():\n\n flat_candidate = (\n\n test_series_root /\n series_uid\n )\n\n if flat_candidate.is_dir():\n\n series_dir = (\n flat_candidate\n )\n\n else:\n\n missing_series += 1\n\n continue\n\n dcm_files = (\n discover_dicom_files(\n series_dir\n )\n )\n\n if not dcm_files:\n\n missing_series += 1\n\n continue\n\n records = []\n\n for dcm_path in dcm_files:\n\n total_files += 1\n\n record = (\n read_dicom_metadata(\n dcm_path\n )\n )\n\n if \"error\" in record:\n\n bad_files += 1\n\n continue\n\n valid_files += 1\n\n records.append(\n record\n )\n\n if (\n\n total_files\n %\n PROGRESS_EVERY_FILES\n ==\n 0\n\n ):\n\n elapsed = (\n time.time()\n -\n start_time\n )\n\n rate = (\n\n total_files\n /\n max(\n elapsed,\n 1e-6,\n )\n )\n\n print(\n\n f\" Files: \"\n f\"{total_files:,} | \"\n f\"Valid: \"\n f\"{valid_files:,} | \"\n f\"Bad: \"\n f\"{bad_files:,} | \"\n f\"{rate:.1f} files/s\"\n )\n\n if not records:\n\n continue\n\n records = (\n sort_series_records(\n records\n )\n )\n\n plane = str(\n\n row[\n \"Anatomical_Plane\"\n ]\n ).strip()\n\n fluid_sensitive = int(\n\n row[\n \"Fluid_Sensitive\"\n ]\n )\n\n fat_suppression = int(\n\n row.get(\n \"Fat_Suppression\",\n 0,\n )\n )\n\n first = records[0]\n\n series_info = {\n\n \"series_uid\":\n series_uid,\n\n \"series_dir\":\n str(\n series_dir\n ),\n\n \"plane\":\n plane,\n\n \"fluid_sensitive\":\n fluid_sensitive,\n\n \"fat_suppression\":\n fat_suppression,\n\n \"series_description\":\n first.get(\n \"series_description\",\n \"\",\n ),\n\n \"protocol_name\":\n first.get(\n \"protocol_name\",\n \"\",\n ),\n\n \"sequence_name\":\n first.get(\n \"sequence_name\",\n \"\",\n ),\n\n \"pixel_spacing\":\n first.get(\n \"pixel_spacing\",\n [0.5, 0.5],\n ),\n\n \"body_part\":\n first.get(\n \"body_part\",\n \"\",\n ),\n\n \"modality\":\n first.get(\n \"modality\",\n \"\",\n ),\n\n \"records\":\n records,\n\n \"image_count\":\n len(records),\n }\n\n series_lookup[\n series_uid\n ] = series_info\n\n lookup[\n study_id\n ] = series_lookup\n\n if (\n\n study_index == 1\n\n or\n\n study_index\n %\n PROGRESS_EVERY_STUDIES\n ==\n 0\n\n or\n\n study_index\n ==\n len(test_studies)\n\n ):\n\n elapsed = (\n time.time()\n -\n start_time\n )\n\n rate = (\n\n study_index\n /\n max(\n elapsed,\n 1e-6,\n )\n )\n\n remaining = (\n\n len(test_studies)\n -\n study_index\n )\n\n eta = (\n\n remaining\n /\n max(\n rate,\n 1e-6,\n )\n )\n\n series_count = sum(\n\n len(x)\n\n for x in lookup.values()\n )\n\n print()\n\n print(\n\n f\"[{study_index:,}/\"\n f\"{len(test_studies):,}] \"\n f\"studies | \"\n f\"Series: \"\n f\"{series_count:,} | \"\n f\"Files: \"\n f\"{total_files:,} | \"\n f\"Bad: \"\n f\"{bad_files:,} | \"\n f\"ETA: \"\n f\"{eta / 60:.2f} min\"\n )\n\n save_lookup_cache(\n\n cache_file,\n\n signature,\n\n lookup,\n )\n\n elapsed = (\n time.time()\n -\n start_time\n )\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"LOOKUP COMPLETE\"\n )\n\n print(\n \"=\" * 70\n )\n\n print(\n f\"Studies indexed: \"\n f\"{len(lookup):,}\"\n )\n\n print(\n f\"Missing studies: \"\n f\"{missing_studies:,}\"\n )\n\n print(\n f\"Missing series: \"\n f\"{missing_series:,}\"\n )\n\n print(\n f\"Valid DICOMs: \"\n f\"{valid_files:,}\"\n )\n\n print(\n f\"Malformed DICOMs: \"\n f\"{bad_files:,}\"\n )\n\n print(\n f\"Total files: \"\n f\"{total_files:,}\"\n )\n\n print(\n f\"Elapsed: \"\n f\"{elapsed / 60:.2f} minutes\"\n )\n\n if elapsed > 0:\n\n print(\n\n f\"File rate: \"\n f\"{total_files / elapsed:.1f} files/s\"\n )\n\n return lookup\n\n\n# ============================================================\n# 24. CHOOSE SERIES\n# ============================================================\n\ndef choose_series(\n series_lookup,\n plane,\n fluid_preference=None,\n):\n\n if not series_lookup:\n\n return None\n\n candidates = []\n\n for (\n series_uid,\n series,\n ) in series_lookup.items():\n\n if (\n\n str(\n series.get(\n \"plane\",\n \"\",\n )\n ).lower()\n\n !=\n\n str(\n plane\n ).lower()\n\n ):\n\n continue\n\n candidates.append(\n series\n )\n\n if not candidates:\n\n return None\n\n if fluid_preference is not None:\n\n desired = (\n\n 1\n\n if fluid_preference\n\n else 0\n )\n\n fluid_candidates = [\n\n series\n\n for series in candidates\n\n if int(\n series.get(\n \"fluid_sensitive\",\n 0,\n )\n )\n ==\n desired\n ]\n\n if fluid_candidates:\n\n candidates = (\n fluid_candidates\n )\n\n return max(\n\n candidates,\n\n key=lambda x: int(\n\n x.get(\n \"image_count\",\n 0,\n )\n ),\n )\n\n\n# ============================================================\n# 25. READ DICOM PIXELS\n# ============================================================\n\ndef read_dicom_pixels(path):\n\n ds = pydicom.dcmread(\n\n str(path),\n\n force=True,\n )\n\n image = ds.pixel_array\n\n try:\n\n image = apply_modality_lut(\n\n image,\n\n ds,\n )\n\n except Exception:\n\n pass\n\n image = np.asarray(\n\n image,\n\n dtype=np.float32,\n )\n\n if (\n\n str(\n\n getattr(\n\n ds,\n\n \"PhotometricInterpretation\",\n\n \"\",\n )\n ).upper()\n\n ==\n\n \"MONOCHROME1\"\n\n ):\n\n image = (\n\n np.max(image)\n -\n image\n )\n\n image = np.nan_to_num(\n\n image,\n\n nan=0.0,\n\n posinf=0.0,\n\n neginf=0.0,\n )\n\n return image\n\n\n# ============================================================\n# 26. PERCENTILE NORMALIZATION\n# ============================================================\n\ndef normalize_percentile(\n image\n):\n\n image = np.asarray(\n\n image,\n\n dtype=np.float32,\n )\n\n valid = np.isfinite(\n image\n )\n\n if not valid.any():\n\n return np.zeros_like(\n\n image,\n\n dtype=np.float32,\n )\n\n values = image[\n valid\n ]\n\n lo, hi = np.percentile(\n\n values,\n\n [\n 2,\n 98,\n ],\n )\n\n if hi <= lo:\n\n lo = float(\n values.min()\n )\n\n hi = float(\n values.max()\n )\n\n if hi <= lo:\n\n return np.zeros_like(\n\n image,\n\n dtype=np.float32,\n )\n\n image = (\n\n image - lo\n\n ) / (\n\n hi - lo\n )\n\n return np.clip(\n\n image,\n\n 0.0,\n\n 1.0,\n\n ).astype(\n np.float32\n )\n\n\n# ============================================================\n# 27. CENTRAL PHYSICAL CROP\n# ============================================================\n\ndef central_physical_crop(\n image,\n pixel_spacing,\n crop_mm,\n):\n\n h, w = image.shape[:2]\n\n try:\n\n sy = float(\n pixel_spacing[0]\n )\n\n sx = float(\n pixel_spacing[1]\n )\n\n except Exception:\n\n sy = 0.5\n\n sx = 0.5\n\n if sy <= 0:\n\n sy = 0.5\n\n if sx <= 0:\n\n sx = 0.5\n\n crop_h = max(\n\n 1,\n\n int(\n\n round(\n crop_mm / sy\n )\n ),\n )\n\n crop_w = max(\n\n 1,\n\n int(\n\n round(\n crop_mm / sx\n )\n ),\n )\n\n crop_h = min(\n crop_h,\n h,\n )\n\n crop_w = min(\n crop_w,\n w,\n )\n\n y0 = max(\n\n 0,\n\n (h - crop_h) // 2,\n )\n\n x0 = max(\n\n 0,\n\n (w - crop_w) // 2,\n )\n\n return image[\n\n y0:\n y0 + crop_h,\n\n x0:\n x0 + crop_w,\n ]\n\n\n# ============================================================\n# 28. RESIZE\n# ============================================================\n\ndef resize_image(\n image,\n size,\n):\n\n if (\n\n image.shape[0]\n ==\n size\n\n and\n\n image.shape[1]\n ==\n size\n\n ):\n\n return image.astype(\n np.float32\n )\n\n return cv2.resize(\n\n image,\n\n (\n size,\n size,\n ),\n\n interpolation=\n cv2.INTER_AREA,\n\n ).astype(\n np.float32\n )\n\n\n# ============================================================\n# 29. BUILD VOLUME\n# ============================================================\n\ndef build_volume(\n series_lookup,\n total_slices=TOTAL_SLICES,\n):\n\n volume = np.zeros(\n\n (\n\n total_slices,\n\n IMAGE_SIZE,\n\n IMAGE_SIZE,\n ),\n\n dtype=np.float32,\n )\n\n valid = np.zeros(\n\n total_slices,\n\n dtype=bool,\n )\n\n cursor = 0\n\n for (\n\n plane,\n fluid_flag,\n count,\n\n ) in SERIES_SLOTS:\n\n count = min(\n\n int(count),\n\n total_slices - cursor,\n )\n\n if count <= 0:\n\n break\n\n fluid_preference = (\n\n None\n\n if fluid_flag == -1\n\n else bool(fluid_flag)\n )\n\n series = choose_series(\n\n series_lookup,\n\n plane,\n\n fluid_preference,\n )\n\n if series is None:\n\n cursor += count\n\n continue\n\n records = series.get(\n\n \"records\",\n\n [],\n )\n\n if not records:\n\n cursor += count\n\n continue\n\n n = len(records)\n\n if n == 1:\n\n indices = np.zeros(\n\n count,\n\n dtype=int,\n )\n\n else:\n\n indices = (\n\n np.linspace(\n\n 0,\n\n n - 1,\n\n count,\n )\n\n .round()\n\n .astype(int)\n )\n\n for (\n\n local_idx,\n src_idx,\n\n ) in enumerate(indices):\n\n dst_idx = (\n\n cursor\n +\n local_idx\n )\n\n if dst_idx >= total_slices:\n\n break\n\n record = records[\n int(src_idx)\n ]\n\n try:\n\n image = (\n read_dicom_pixels(\n record[\"path\"]\n )\n )\n\n image = (\n normalize_percentile(\n image\n )\n )\n\n spacing = (\n\n record.get(\n\n \"pixel_spacing\",\n\n [\n 0.5,\n 0.5,\n ],\n )\n )\n\n image = (\n central_physical_crop(\n\n image,\n\n spacing,\n\n CROP_MM,\n )\n )\n\n image = (\n resize_image(\n\n image,\n\n IMAGE_SIZE,\n )\n )\n\n volume[\n dst_idx\n ] = image\n\n valid[\n dst_idx\n ] = True\n\n except Exception:\n\n pass\n\n cursor += count\n\n return (\n volume,\n valid,\n )\n\n\n# ============================================================\n# 30. MAKE SLICE CENTERS\n# ============================================================\n\ndef make_centers(\n valid_mask,\n k,\n lo,\n hi,\n):\n\n valid_indices = np.where(\n valid_mask\n )[0]\n\n if len(valid_indices) == 0:\n\n return (\n\n np.linspace(\n\n 0,\n\n TOTAL_SLICES - 1,\n\n k,\n )\n\n .round()\n\n .astype(int)\n )\n\n first = int(\n valid_indices[0]\n )\n\n last = int(\n valid_indices[-1]\n )\n\n lo_idx = int(\n\n round(\n\n first\n\n +\n\n lo\n *\n (\n last - first\n )\n )\n )\n\n hi_idx = int(\n\n round(\n\n first\n\n +\n\n hi\n *\n (\n last - first\n )\n )\n )\n\n lo_idx = max(\n\n first,\n\n min(\n lo_idx,\n last,\n ),\n )\n\n hi_idx = max(\n\n lo_idx,\n\n min(\n hi_idx,\n last,\n ),\n )\n\n centers = (\n\n np.linspace(\n\n lo_idx,\n\n hi_idx,\n\n k,\n )\n\n .round()\n\n .astype(int)\n )\n\n return centers\n\n\n# ============================================================\n# 31. CREATE 2.5D RGB WINDOWS\n# ============================================================\n\ndef create_windows(\n volume,\n centers,\n):\n\n windows = []\n\n for center in centers:\n\n center = int(\n center\n )\n\n prev_idx = max(\n\n 0,\n\n center - 1,\n )\n\n next_idx = min(\n\n volume.shape[0] - 1,\n\n center + 1,\n )\n\n rgb = np.stack(\n\n [\n\n volume[\n prev_idx\n ],\n\n volume[\n center\n ],\n\n volume[\n next_idx\n ],\n ],\n\n axis=0,\n )\n\n rgb = np.clip(\n\n rgb,\n\n 0.0,\n\n 1.0,\n\n ).astype(\n np.float32\n )\n\n rgb = (\n\n rgb\n\n -\n\n IMAGENET_MEAN[\n :,\n None,\n None,\n ]\n\n ) / (\n\n IMAGENET_STD[\n :,\n None,\n None,\n ]\n )\n\n windows.append(\n rgb\n )\n\n if not windows:\n\n return np.empty(\n\n (\n\n 0,\n\n 3,\n\n IMAGE_SIZE,\n\n IMAGE_SIZE,\n ),\n\n dtype=np.float32,\n )\n\n return np.stack(\n\n windows,\n\n axis=0,\n )\n\n\n# ============================================================\n# 32. MODEL\n# ============================================================\n\nclass CheckpointFreeKneeModel(\n nn.Module\n):\n\n def __init__(\n self,\n arch=DEFAULT_ARCH,\n num_classes=NUM_CLASSES,\n pretrained=False,\n ):\n\n super().__init__()\n\n print()\n print(\n \"Creating CoAtNet:\"\n )\n\n print(\n f\" Architecture: {arch}\"\n )\n\n print(\n f\" Pretrained: {pretrained}\"\n )\n\n # ----------------------------------------------------\n # IMPORTANT:\n #\n # This call is offline-safe when pretrained=False.\n #\n # We NEVER let timm attempt an online download.\n # ----------------------------------------------------\n\n self.backbone = (\n timm.create_model(\n\n arch,\n\n pretrained=pretrained,\n\n num_classes=num_classes,\n\n in_chans=3,\n )\n )\n\n self.num_classes = (\n num_classes\n )\n\n def forward(\n self,\n x,\n ):\n\n return self.backbone(x)\n\n\n# ============================================================\n# 33. FIND LOCAL CHECKPOINTS\n# ============================================================\n\ndef find_local_model_files():\n\n candidates = []\n\n # --------------------------------------------------------\n # Explicit checkpoint first.\n # --------------------------------------------------------\n\n if MODEL_CHECKPOINT is not None:\n\n path = Path(\n MODEL_CHECKPOINT\n )\n\n if path.is_file():\n\n candidates.append(\n path\n )\n\n # --------------------------------------------------------\n # Search Kaggle input for common model files.\n # --------------------------------------------------------\n\n kaggle_input = Path(\n \"/kaggle/input\"\n )\n\n if kaggle_input.exists():\n\n patterns = [\n\n \"*.pth\",\n\n \"*.pt\",\n\n \"*.bin\",\n\n \"*.safetensors\",\n ]\n\n for pattern in patterns:\n\n try:\n\n candidates.extend(\n\n kaggle_input.glob(\n f\"**/{pattern}\"\n )\n )\n\n except Exception:\n\n pass\n\n # --------------------------------------------------------\n # Search working directory.\n # --------------------------------------------------------\n\n working = Path(\n \"/kaggle/working\"\n )\n\n if working.exists():\n\n patterns = [\n\n \"*.pth\",\n\n \"*.pt\",\n\n \"*.bin\",\n\n \"*.safetensors\",\n ]\n\n for pattern in patterns:\n\n try:\n\n candidates.extend(\n\n working.glob(\n f\"**/{pattern}\"\n )\n )\n\n except Exception:\n\n pass\n\n # --------------------------------------------------------\n # Remove duplicates.\n # --------------------------------------------------------\n\n unique = []\n\n seen = set()\n\n for path in candidates:\n\n try:\n\n key = str(\n path.resolve()\n )\n\n except Exception:\n\n key = str(path)\n\n if key in seen:\n\n continue\n\n seen.add(key)\n\n unique.append(\n path\n )\n\n return unique\n\n\n# ============================================================\n# 34. INSPECT CHECKPOINT\n# ============================================================\n\ndef checkpoint_looks_compatible(\n checkpoint_path\n):\n\n try:\n\n checkpoint = torch.load(\n\n checkpoint_path,\n\n map_location=\"cpu\",\n\n weights_only=False,\n )\n\n except Exception as exc:\n\n print()\n print(\n \"Could not inspect checkpoint:\"\n )\n\n print(\n f\" {checkpoint_path}\"\n )\n\n print(\n f\" {exc}\"\n )\n\n return False\n\n state_dict = None\n\n if isinstance(\n checkpoint,\n dict,\n ):\n\n # Common formats.\n for key in [\n\n \"state_dict\",\n\n \"model_state_dict\",\n\n \"model\",\n\n \"weights\",\n ]:\n\n if key in checkpoint:\n\n candidate = (\n checkpoint[key]\n )\n\n if isinstance(\n candidate,\n dict,\n ):\n\n state_dict = (\n candidate\n )\n\n break\n\n if state_dict is None:\n\n # Could itself be a state_dict.\n if all(\n\n isinstance(\n k,\n str,\n )\n\n for k in checkpoint.keys()\n\n ):\n\n state_dict = checkpoint\n\n if state_dict is None:\n\n print(\n \"Checkpoint format not recognized.\"\n )\n\n return False\n\n # --------------------------------------------------------\n # Look for a classifier tensor.\n # --------------------------------------------------------\n\n possible_heads = []\n\n for key, value in (\n state_dict.items()\n ):\n\n if not torch.is_tensor(\n value\n ):\n\n continue\n\n if value.ndim != 2:\n\n continue\n\n if (\n value.shape[0]\n ==\n NUM_CLASSES\n ):\n\n possible_heads.append(\n (\n key,\n tuple(\n value.shape\n ),\n )\n )\n\n print()\n print(\n \"Checkpoint inspection:\"\n )\n\n print(\n f\" File: {checkpoint_path}\"\n )\n\n print(\n f\" Candidate 12-class heads:\"\n f\" {len(possible_heads)}\"\n )\n\n for key, shape in (\n possible_heads[:10]\n ):\n\n print(\n f\" {key}: {shape}\"\n )\n\n return len(\n possible_heads\n ) > 0\n\n\n# ============================================================\n# 35. LOAD CHECKPOINT STATE\n# ============================================================\n\ndef load_checkpoint_state(\n model,\n checkpoint_path,\n):\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"LOADING LOCAL CHECKPOINT\"\n )\n\n print(\n \"=\" * 70\n )\n\n print(\n f\"Checkpoint:\"\n )\n\n print(\n f\" {checkpoint_path}\"\n )\n\n checkpoint = torch.load(\n\n checkpoint_path,\n\n map_location=\"cpu\",\n\n weights_only=False,\n )\n\n state_dict = None\n\n if isinstance(\n checkpoint,\n dict,\n ):\n\n for key in [\n\n \"state_dict\",\n\n \"model_state_dict\",\n\n \"model\",\n\n \"weights\",\n ]:\n\n if key in checkpoint:\n\n candidate = (\n checkpoint[key]\n )\n\n if isinstance(\n candidate,\n dict,\n ):\n\n state_dict = (\n candidate\n )\n\n break\n\n if state_dict is None:\n\n if all(\n\n isinstance(\n k,\n str,\n )\n\n for k in checkpoint.keys()\n\n ):\n\n state_dict = checkpoint\n\n if state_dict is None:\n\n raise RuntimeError(\n\n \"Could not locate a state_dict \"\n \"inside the checkpoint.\"\n )\n\n # --------------------------------------------------------\n # Remove DataParallel prefix.\n # --------------------------------------------------------\n\n cleaned = {}\n\n for key, value in (\n state_dict.items()\n ):\n\n if key.startswith(\n \"module.\"\n ):\n\n key = key[\n len(\"module.\") :\n ]\n\n cleaned[key] = value\n\n state_dict = cleaned\n\n # --------------------------------------------------------\n # Try strict loading first.\n # --------------------------------------------------------\n\n try:\n\n result = model.load_state_dict(\n\n state_dict,\n\n strict=True,\n )\n\n print(\n \"Checkpoint loaded with \"\n \"strict=True.\"\n )\n\n return True\n\n except Exception as exc:\n\n print()\n print(\n \"Strict checkpoint loading failed:\"\n )\n\n print(\n str(exc)[:3000]\n )\n\n # --------------------------------------------------------\n # Try non-strict loading.\n # --------------------------------------------------------\n\n try:\n\n result = model.load_state_dict(\n\n state_dict,\n\n strict=False,\n )\n\n print()\n print(\n \"Checkpoint loaded with \"\n \"strict=False.\"\n )\n\n print(\n f\"Missing keys: \"\n f\"{len(result.missing_keys)}\"\n )\n\n print(\n f\"Unexpected keys: \"\n f\"{len(result.unexpected_keys)}\"\n )\n\n return True\n\n except Exception as exc:\n\n print()\n print(\n \"Non-strict checkpoint loading \"\n \"also failed:\"\n )\n\n print(\n str(exc)[:3000]\n )\n\n return False\n\n\n# ============================================================\n# 36. LOAD MODEL — STRICTLY OFFLINE\n# ============================================================\n\ndef load_model():\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"LOADING OFFLINE MODEL\"\n )\n\n print(\n \"=\" * 70\n )\n\n # --------------------------------------------------------\n # STEP A\n # Look for an actual trained local checkpoint.\n # --------------------------------------------------------\n\n local_files = (\n find_local_model_files()\n )\n\n if local_files:\n\n print()\n print(\n \"Local model files found:\"\n )\n\n for path in local_files[:20]:\n\n print(\n f\" {path}\"\n )\n\n # ----------------------------------------------------\n # Prefer explicit MODEL_CHECKPOINT.\n # ----------------------------------------------------\n\n ordered = []\n\n if MODEL_CHECKPOINT is not None:\n\n explicit = Path(\n MODEL_CHECKPOINT\n )\n\n if explicit in local_files:\n\n ordered.append(\n explicit\n )\n\n for path in local_files:\n\n if path not in ordered:\n\n ordered.append(\n path\n )\n\n # ----------------------------------------------------\n # Try compatible checkpoints.\n # ----------------------------------------------------\n\n for checkpoint_path in ordered:\n\n if not checkpoint_path.is_file():\n\n continue\n\n try:\n\n if not checkpoint_looks_compatible(\n checkpoint_path\n ):\n\n continue\n\n model = (\n CheckpointFreeKneeModel(\n\n arch=DEFAULT_ARCH,\n\n num_classes=NUM_CLASSES,\n\n pretrained=False,\n )\n )\n\n loaded = (\n load_checkpoint_state(\n\n model,\n\n checkpoint_path,\n )\n )\n\n if loaded:\n\n model = model.to(\n DEVICE\n )\n\n model.eval()\n\n print()\n print(\n \"SUCCESS:\"\n )\n\n print(\n \"Using local trained \"\n \"RSNA checkpoint.\"\n )\n\n return model\n\n except Exception as exc:\n\n print()\n print(\n \"Checkpoint attempt failed:\"\n )\n\n print(\n f\" {checkpoint_path}\"\n )\n\n print(\n f\" {exc}\"\n )\n\n # --------------------------------------------------------\n # STEP B\n #\n # IMPORTANT:\n #\n # We DO NOT call:\n #\n # timm.create_model(..., pretrained=True)\n #\n # because that can attempt Hugging Face access.\n #\n # Instead, we create the architecture locally.\n # --------------------------------------------------------\n\n print()\n print(\n \"No compatible local RSNA checkpoint found.\"\n )\n\n print()\n print(\n \"Creating CoAtNet with pretrained=False.\"\n )\n\n print(\n \"This guarantees no Hugging Face download.\"\n )\n\n model = (\n CheckpointFreeKneeModel(\n\n arch=DEFAULT_ARCH,\n\n num_classes=NUM_CLASSES,\n\n pretrained=False,\n )\n )\n\n model = model.to(\n DEVICE\n )\n\n model.eval()\n\n print()\n print(\n \"WARNING:\"\n )\n\n print(\n \"The current model has randomly initialized \"\n \"weights.\"\n )\n\n print(\n \"It is NOT a trained RSNA 12-label model.\"\n )\n\n print()\n print(\n \"The pipeline is now completely offline.\"\n )\n\n return model\n\n\n# ============================================================\n# 37. PREDICT WINDOWS\n# ============================================================\n\n@torch.inference_mode()\ndef predict_windows(\n model,\n windows,\n batch_size=16,\n):\n\n if len(windows) == 0:\n\n return np.zeros(\n\n NUM_CLASSES,\n\n dtype=np.float32,\n )\n\n predictions = []\n\n for start in range(\n\n 0,\n\n len(windows),\n\n batch_size,\n ):\n\n batch = torch.from_numpy(\n\n windows[\n\n start:\n start + batch_size\n ]\n ).to(\n\n DEVICE,\n\n non_blocking=True,\n )\n\n try:\n\n if DEVICE.type == \"cuda\":\n\n with torch.autocast(\n\n device_type=\"cuda\",\n\n dtype=torch.float16,\n ):\n\n logits = model(\n batch\n )\n\n else:\n\n logits = model(\n batch\n )\n\n except Exception:\n\n logits = model(\n batch\n )\n\n if logits.ndim == 1:\n\n logits = logits.unsqueeze(\n 0\n )\n\n probs = torch.sigmoid(\n\n logits.float()\n )\n\n predictions.append(\n\n probs.detach()\n .cpu()\n .numpy()\n )\n\n predictions = np.concatenate(\n\n predictions,\n\n axis=0,\n )\n\n return predictions.mean(\n\n axis=0\n\n ).astype(\n np.float32\n )\n\n\n# ============================================================\n# 38. TTA FUSION\n# ============================================================\n\ndef fuse_tta(\n predictions,\n weights,\n):\n\n predictions = np.asarray(\n\n predictions,\n\n dtype=np.float32,\n )\n\n weights = np.asarray(\n\n weights,\n\n dtype=np.float32,\n )\n\n weight_sum = weights.sum()\n\n if weight_sum <= 0:\n\n weights = np.ones_like(\n weights\n )\n\n weight_sum = weights.sum()\n\n weights = (\n\n weights\n /\n weight_sum\n )\n\n probabilities = np.clip(\n\n predictions,\n\n 1e-6,\n\n 1.0 - 1e-6,\n )\n\n logits = np.log(\n\n probabilities\n\n /\n\n (\n 1.0\n -\n probabilities\n )\n )\n\n fused_logits = np.sum(\n\n logits\n *\n weights[:, None],\n\n axis=0,\n )\n\n fused_probs = (\n\n 1.0\n\n /\n\n (\n\n 1.0\n +\n np.exp(\n -fused_logits\n )\n )\n )\n\n direct_probs = np.sum(\n\n probabilities\n *\n weights[:, None],\n\n axis=0,\n )\n\n fused = (\n\n 0.60\n *\n fused_probs\n\n +\n\n 0.40\n *\n direct_probs\n )\n\n return np.clip(\n\n fused,\n\n 0.0,\n\n 1.0,\n\n ).astype(\n np.float32\n )\n\n\n# ============================================================\n# 39. PREDICT ONE STUDY\n# ============================================================\n\ndef predict_study(\n model,\n series_lookup,\n):\n\n volume, valid = (\n build_volume(\n\n series_lookup,\n\n TOTAL_SLICES,\n )\n )\n\n valid_count = int(\n valid.sum()\n )\n\n if valid_count == 0:\n\n print(\n \"WARNING: study contains \"\n \"no valid DICOM slices.\"\n )\n\n return np.full(\n\n NUM_CLASSES,\n\n 0.5,\n\n dtype=np.float32,\n )\n\n tta_predictions = []\n\n tta_weights = []\n\n for config in TTA_CONFIGS:\n\n centers = make_centers(\n\n valid,\n\n config[\"k\"],\n\n config[\"lo\"],\n\n config[\"hi\"],\n )\n\n windows = create_windows(\n\n volume,\n\n centers,\n )\n\n prediction = (\n predict_windows(\n\n model,\n\n windows,\n )\n )\n\n tta_predictions.append(\n prediction\n )\n\n tta_weights.append(\n config[\"weight\"]\n )\n\n del windows\n\n if DEVICE.type == \"cuda\":\n\n torch.cuda.empty_cache()\n\n result = fuse_tta(\n\n tta_predictions,\n\n tta_weights,\n )\n\n del volume\n\n del valid\n\n return result\n\n\n# ============================================================\n# 40. FINAL RANK CALIBRATION\n# ============================================================\n\ndef final_rank_calibration(\n predictions,\n):\n\n predictions = np.asarray(\n\n predictions,\n\n dtype=np.float32,\n ).copy()\n\n if len(predictions) <= 1:\n\n return np.clip(\n\n predictions,\n\n 0.0,\n\n 1.0,\n )\n\n for j, label in enumerate(\n LABELS\n ):\n\n values = predictions[\n :,\n j,\n ]\n\n order = np.argsort(\n\n np.argsort(\n values\n )\n )\n\n ranks = (\n\n order.astype(\n np.float32\n )\n\n /\n\n max(\n\n len(values) - 1,\n\n 1,\n )\n )\n\n weight = (\n\n TARGET_RANK_WEIGHTS.get(\n\n label,\n\n 0.0,\n )\n )\n\n predictions[\n\n :,\n\n j\n\n ] = (\n\n (\n\n 1.0\n -\n weight\n )\n\n *\n\n values\n\n +\n\n weight\n\n *\n\n ranks\n )\n\n return np.clip(\n\n predictions,\n\n 0.0,\n\n 1.0,\n )\n\n\n# ============================================================\n# 41. SAMPLE SUBMISSION\n# ============================================================\n\ndef load_sample_submission(\n root\n):\n\n path = (\n\n root /\n\n \"sample_submission.csv\"\n )\n\n if path.is_file():\n\n return pd.read_csv(\n path\n )\n\n matches = list(\n\n root.glob(\n\n \"**/sample_submission.csv\"\n )\n )\n\n if matches:\n\n return pd.read_csv(\n matches[0]\n )\n\n return None\n\n\n# ============================================================\n# 42. VALIDATE SUBMISSION\n# ============================================================\n\ndef validate_submission(\n submission,\n sample_submission=None,\n):\n\n if submission.empty:\n\n raise ValueError(\n \"Submission is empty.\"\n )\n\n id_col = (\n submission.columns[0]\n )\n\n expected_columns = [\n\n id_col\n\n ] + LABELS\n\n missing = [\n\n col\n\n for col in expected_columns\n\n if col not in submission.columns\n ]\n\n if missing:\n\n raise ValueError(\n\n f\"Missing columns: {missing}\"\n )\n\n prediction_values = (\n\n submission[\n LABELS\n ].to_numpy()\n )\n\n if not np.isfinite(\n prediction_values\n ).all():\n\n raise ValueError(\n\n \"Submission contains \"\n \"NaN or infinite values.\"\n )\n\n if (\n\n prediction_values.min()\n <\n 0\n\n or\n\n prediction_values.max()\n >\n 1\n\n ):\n\n raise ValueError(\n\n \"Predictions must be \"\n \"between 0 and 1.\"\n )\n\n if sample_submission is not None:\n\n expected = list(\n\n sample_submission.columns\n )\n\n actual = list(\n\n submission.columns\n )\n\n if expected != actual:\n\n raise ValueError(\n\n \"\\nSubmission columns do not \"\n \"match sample_submission.\\n\"\n\n f\"Expected: {expected}\\n\"\n\n f\"Actual: {actual}\"\n )\n\n if (\n\n len(submission)\n\n !=\n\n len(sample_submission)\n\n ):\n\n raise ValueError(\n\n \"Submission row count does \"\n \"not match sample_submission.\"\n )\n\n id_col = expected[0]\n\n actual_ids = (\n\n submission[\n id_col\n ].astype(str)\n )\n\n expected_ids = (\n\n sample_submission[\n id_col\n ].astype(str)\n )\n\n if not actual_ids.equals(\n expected_ids\n ):\n\n raise ValueError(\n\n \"Submission IDs/order do \"\n \"not match sample_submission.\"\n )\n\n print(\n \"Submission validation: PASSED\"\n )\n\n\n# ============================================================\n# 43. SANITY CHECK LOOKUP\n# ============================================================\n\ndef sanity_check_lookup(\n lookup,\n test_studies,\n):\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"LOOKUP SANITY CHECK\"\n )\n\n print(\n \"=\" * 70\n )\n\n total_series = 0\n\n total_slices = 0\n\n studies_with_data = 0\n\n for study_id in test_studies:\n\n study = lookup.get(\n\n str(study_id),\n\n {},\n )\n\n series_count = len(\n study\n )\n\n slice_count = sum(\n\n int(\n\n s.get(\n\n \"image_count\",\n\n 0,\n )\n )\n\n for s in study.values()\n )\n\n total_series += (\n series_count\n )\n\n total_slices += (\n slice_count\n )\n\n if slice_count > 0:\n\n studies_with_data += 1\n\n print(\n\n f\"Study {study_id}: \"\n\n f\"{series_count} series | \"\n\n f\"{slice_count} slices\"\n )\n\n print()\n\n print(\n\n f\"Studies with DICOMs: \"\n\n f\"{studies_with_data}/\"\n\n f\"{len(test_studies)}\"\n )\n\n print(\n\n f\"Total series: \"\n\n f\"{total_series:,}\"\n )\n\n print(\n\n f\"Total DICOM slices: \"\n\n f\"{total_slices:,}\"\n )\n\n if (\n\n studies_with_data\n\n ==\n\n 0\n\n ):\n\n raise RuntimeError(\n\n \"\\nNO TEST DICOM DATA FOUND.\\n\"\n\n \"\\n\"\n\n \"Check that the competition input \"\n \"contains test_series/.\\n\"\n )\n\n print(\n \"Lookup sanity check: PASSED\"\n )\n\n\n# ============================================================\n# 44. BUILD TEST LOOKUP\n# ============================================================\n\ndef build_lookup_for_test(\n root,\n test_df,\n id_col,\n):\n\n test_series_root = (\n locate_test_series(\n root\n )\n )\n\n (\n test_series_df,\n test_series_csv,\n ) = (\n\n load_test_series_csv(\n root\n )\n )\n\n test_studies = (\n\n test_df[\n id_col\n ]\n\n .astype(str)\n\n .tolist()\n )\n\n lookup = (\n\n create_series_lookup_optimized(\n\n test_series_root,\n\n test_studies,\n\n test_series_df,\n\n root / \"test.csv\",\n\n test_series_csv,\n\n CACHE_FILE,\n )\n )\n\n sanity_check_lookup(\n\n lookup,\n\n test_studies,\n )\n\n return lookup\n\n\n# ============================================================\n# 45. OFFLINE STATUS CHECK\n# ============================================================\n\ndef print_offline_status():\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"OFFLINE STATUS CHECK\"\n )\n\n print(\n \"=\" * 70\n )\n\n print(\n \"Hugging Face offline:\",\n os.environ.get(\n \"HF_HUB_OFFLINE\"\n )\n )\n\n print(\n \"Transformers offline:\",\n os.environ.get(\n \"TRANSFORMERS_OFFLINE\"\n )\n )\n\n print(\n \"Xet disabled:\",\n os.environ.get(\n \"HF_HUB_DISABLE_XET\"\n )\n )\n\n print()\n\n print(\n \"No internet download will be \"\n \"attempted by this pipeline.\"\n )\n\n\n# ============================================================\n# 46. MAIN\n# ============================================================\n\ndef main():\n\n total_start = time.time()\n\n # --------------------------------------------------------\n # OFFLINE STATUS\n # --------------------------------------------------------\n\n print_offline_status()\n\n # --------------------------------------------------------\n # STEP 1\n # --------------------------------------------------------\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"STEP 1 — LOCATING DATA\"\n )\n\n print(\n \"=\" * 70\n )\n\n root = (\n locate_competition_root()\n )\n\n # --------------------------------------------------------\n # STEP 2\n # --------------------------------------------------------\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"STEP 2 — READING TEST CSV\"\n )\n\n print(\n \"=\" * 70\n )\n\n (\n test_df,\n id_col,\n test_csv,\n ) = load_test_csv(\n root\n )\n\n # --------------------------------------------------------\n # STEP 3\n # --------------------------------------------------------\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"STEP 3 — BUILDING TEST SERIES LOOKUP\"\n )\n\n print(\n \"=\" * 70\n )\n\n lookup = (\n build_lookup_for_test(\n\n root,\n\n test_df,\n\n id_col,\n )\n )\n\n # --------------------------------------------------------\n # STEP 4\n # --------------------------------------------------------\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"STEP 4 — LOADING OFFLINE MODEL\"\n )\n\n print(\n \"=\" * 70\n )\n\n model = load_model()\n\n # --------------------------------------------------------\n # STEP 5\n # --------------------------------------------------------\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"STEP 5 — RUNNING OFFLINE INFERENCE\"\n )\n\n print(\n \"=\" * 70\n )\n\n predictions = []\n\n study_ids = (\n\n test_df[\n id_col\n ]\n\n .astype(str)\n\n .tolist()\n )\n\n inference_start = time.time()\n\n for (\n\n index,\n\n study_id,\n\n ) in enumerate(\n\n study_ids,\n\n start=1,\n ):\n\n series_lookup = (\n\n lookup.get(\n\n study_id,\n\n {},\n )\n )\n\n try:\n\n prediction = (\n\n predict_study(\n\n model,\n\n series_lookup,\n )\n )\n\n except Exception as exc:\n\n print()\n\n print(\n\n f\"WARNING: inference failed \"\n f\"for study {study_id}\"\n )\n\n print(\n f\"Reason: {exc}\"\n )\n\n prediction = np.full(\n\n NUM_CLASSES,\n\n 0.5,\n\n dtype=np.float32,\n )\n\n predictions.append(\n prediction\n )\n\n if (\n\n index == 1\n\n or\n\n index\n %\n PROGRESS_EVERY_STUDIES\n ==\n 0\n\n or\n\n index\n ==\n len(study_ids)\n\n ):\n\n elapsed = (\n\n time.time()\n -\n inference_start\n )\n\n rate = (\n\n index\n /\n max(\n elapsed,\n 1e-6,\n )\n )\n\n remaining = (\n\n len(study_ids)\n -\n index\n )\n\n eta = (\n\n remaining\n /\n max(\n rate,\n 1e-6,\n )\n )\n\n print(\n\n f\"[{index:,}/\"\n f\"{len(study_ids):,}] \"\n\n f\"studies | \"\n\n f\"{rate:.3f} studies/s | \"\n\n f\"ETA: \"\n f\"{eta / 60:.2f} min\"\n )\n\n gc.collect()\n\n predictions = np.vstack(\n predictions\n )\n\n # --------------------------------------------------------\n # STEP 6\n # --------------------------------------------------------\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"STEP 6 — FINAL RANK CALIBRATION\"\n )\n\n print(\n \"=\" * 70\n )\n\n calibrated = (\n\n final_rank_calibration(\n\n predictions\n )\n )\n\n # --------------------------------------------------------\n # STEP 7\n # --------------------------------------------------------\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"STEP 7 — CREATING OFFLINE SUBMISSION\"\n )\n\n print(\n \"=\" * 70\n )\n\n submission = pd.DataFrame(\n\n calibrated,\n\n columns=LABELS,\n )\n\n submission.insert(\n\n 0,\n\n id_col,\n\n test_df[\n id_col\n ].values,\n )\n\n sample_submission = (\n\n load_sample_submission(\n root\n )\n )\n\n validate_submission(\n\n submission,\n\n sample_submission,\n )\n\n submission.to_csv(\n\n OUTPUT_FILE,\n\n index=False,\n )\n\n print()\n print(\n \"Submission saved:\"\n )\n\n print(\n f\" {OUTPUT_FILE}\"\n )\n\n print()\n\n print(\n submission.head()\n )\n\n # --------------------------------------------------------\n # Prediction statistics\n # --------------------------------------------------------\n\n print()\n print(\n \"Prediction statistics:\"\n )\n\n print()\n\n stats = pd.DataFrame({\n\n \"label\":\n LABELS,\n\n \"min\":\n calibrated.min(\n axis=0\n ),\n\n \"mean\":\n calibrated.mean(\n axis=0\n ),\n\n \"max\":\n calibrated.max(\n axis=0\n ),\n })\n\n print(\n\n stats.to_string(\n index=False\n )\n )\n\n # --------------------------------------------------------\n # STEP 8\n # --------------------------------------------------------\n\n total_elapsed = (\n\n time.time()\n -\n total_start\n )\n\n print()\n print(\n \"=\" * 70\n )\n\n print(\n \"PIPELINE COMPLETE\"\n )\n\n print(\n \"=\" * 70\n )\n\n print(\n\n f\"Studies: \"\n f\"{len(submission):,}\"\n )\n\n print(\n\n f\"Labels: \"\n f\"{NUM_CLASSES}\"\n )\n\n print()\n\n print(\n \"Output:\"\n )\n\n print(\n f\" {OUTPUT_FILE}\"\n )\n\n print()\n\n print(\n \"Total runtime:\"\n )\n\n print(\n f\" {total_elapsed / 60:.2f} minutes\"\n )\n\n print()\n\n print(\n \"OFFLINE SUBMISSION READY.\"\n )\n\n return submission\n\n\n# ============================================================\n# 47. RUN\n# ============================================================\n\nif __name__ == \"__main__\":\n\n submission = main()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-14T08:01:49.952926Z","iopub.execute_input":"2026-09-14T08:01:49.953244Z","iopub.status.idle":"2026-09-14T08:08:08.035595Z","shell.execute_reply.started":"2026-09-14T08:01:49.953218Z","shell.execute_reply":"2026-09-14T08:08:08.0347Z"},"_kg_hide-input":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#

Final Image Visualization

","metadata":{}},{"cell_type":"code","source":"# FINAL IMAGE VISUALIZATIONS\n\nimport os\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\n\n# ============================================================\n# BASIC VALIDATION\n# ============================================================\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL IMAGE VISUALIZATIONS\")\nprint(\"=\" * 70)\n\n# ------------------------------------------------------------\n# Validate submission\n# ------------------------------------------------------------\n\nif \"submission\" not in globals():\n raise NameError(\n \"The `submission` DataFrame does not exist. \"\n \"Please run the prediction/submission cells first.\"\n )\n\nif \"LABELS\" not in globals():\n raise NameError(\n \"The `LABELS` variable is not defined. \"\n \"Please run the label-definition cell first.\"\n )\n\nif len(submission) == 0:\n raise ValueError(\n \"The submission DataFrame is empty.\"\n )\n\n\n# ============================================================\n# AUTOMATICALLY FIND ID COLUMN\n# ============================================================\n\n# Common ID column names used in RSNA-style datasets.\npossible_id_columns = [\n \"StudyInstanceUID\",\n \"study_id\",\n \"studyId\",\n \"StudyID\",\n \"id\",\n \"ID\",\n \"patient_id\",\n \"PatientID\",\n]\n\nid_col = None\n\nfor candidate in possible_id_columns:\n\n if candidate in submission.columns:\n\n id_col = candidate\n break\n\n\n# ------------------------------------------------------------\n# Fallback:\n# use the first column that is NOT a prediction label\n# ------------------------------------------------------------\n\nif id_col is None:\n\n non_label_columns = [\n col\n for col in submission.columns\n if col not in LABELS\n ]\n\n if len(non_label_columns) > 0:\n\n id_col = non_label_columns[0]\n\n\n# ------------------------------------------------------------\n# Stop with a clear message if no ID column exists\n# ------------------------------------------------------------\n\nif id_col is None:\n\n raise ValueError(\n \"Could not identify the submission ID column.\\n\"\n f\"Available columns: {list(submission.columns)}\"\n )\n\n\nprint()\nprint(f\"Submission rows : {len(submission):,}\")\nprint(f\"ID column : {id_col}\")\nprint(f\"Prediction cols : {len(LABELS)}\")\n\n\n# ============================================================\n# VALIDATE LABEL COLUMNS\n# ============================================================\n\nmissing_labels = [\n label\n for label in LABELS\n if label not in submission.columns\n]\n\nif len(missing_labels) > 0:\n\n raise ValueError(\n \"The following LABELS are missing from submission:\\n\"\n f\"{missing_labels}\\n\\n\"\n f\"Available columns:\\n\"\n f\"{list(submission.columns)}\"\n )\n\n\n# ============================================================\n# OUTPUT DIRECTORY\n# ============================================================\n\nOUTPUT_DIR = \"/kaggle/working\"\n\nos.makedirs(\n OUTPUT_DIR,\n exist_ok=True,\n)\n\n\n# ============================================================\n# VISUALIZATION 1\n# Prediction Distribution\n# ============================================================\n\nprint()\nprint(\"=\" * 70)\nprint(\"VISUALIZATION 1 — PREDICTION DISTRIBUTION\")\nprint(\"=\" * 70)\n\n\nplt.figure(\n figsize=(12, 6)\n)\n\n\nfor label in LABELS:\n\n values = pd.to_numeric(\n submission[label],\n errors=\"coerce\",\n ).dropna()\n\n plt.hist(\n values,\n bins=30,\n alpha=0.35,\n label=label,\n )\n\n\nplt.xlabel(\n \"Predicted probability\"\n)\n\nplt.ylabel(\n \"Number of studies\"\n)\n\nplt.title(\n \"Distribution of Predicted Knee Abnormality Probabilities\"\n)\n\nplt.legend(\n bbox_to_anchor=(1.02, 1),\n loc=\"upper left\",\n fontsize=9,\n)\n\nplt.tight_layout()\n\n\nviz1_path = os.path.join(\n OUTPUT_DIR,\n \"prediction_distribution.png\",\n)\n\n\nplt.savefig(\n viz1_path,\n dpi=180,\n bbox_inches=\"tight\",\n)\n\nplt.show()\n\nplt.close()\n\n\nprint(\n f\"Saved: {viz1_path}\"\n)\n\n\n# ============================================================\n# VISUALIZATION 2\n# Prediction Heatmap\n# ============================================================\n\nprint()\nprint(\"=\" * 70)\nprint(\"VISUALIZATION 2 — PREDICTION HEATMAP\")\nprint(\"=\" * 70)\n\n\n# ------------------------------------------------------------\n# Number of studies to display\n# ------------------------------------------------------------\n\nn_show = min(\n 30,\n len(submission),\n)\n\n\n# ------------------------------------------------------------\n# Extract prediction values\n# ------------------------------------------------------------\n\nheatmap_data = (\n submission[LABELS]\n .iloc[:n_show]\n .apply(\n pd.to_numeric,\n errors=\"coerce\",\n )\n .fillna(0.0)\n .values\n)\n\n\n# ------------------------------------------------------------\n# Extract study IDs\n# ------------------------------------------------------------\n\nstudy_ids = (\n submission[id_col]\n .iloc[:n_show]\n .astype(str)\n .tolist()\n)\n\n\n# ============================================================\n# CREATE HEATMAP\n# ============================================================\n\nplt.figure(\n figsize=(14, 8)\n)\n\n\nimage = plt.imshow(\n heatmap_data,\n aspect=\"auto\",\n interpolation=\"nearest\",\n)\n\n\nplt.colorbar(\n image,\n label=\"Predicted probability\",\n)\n\n\nplt.xticks(\n range(len(LABELS)),\n LABELS,\n rotation=45,\n ha=\"right\",\n)\n\n\nplt.yticks(\n range(n_show),\n study_ids,\n)\n\n\nplt.xlabel(\n \"Knee abnormality\"\n)\n\nplt.ylabel(\n \"Study ID\"\n)\n\nplt.title(\n \"Knee Abnormality Prediction Heatmap\"\n)\n\nplt.tight_layout()\n\n\nviz2_path = os.path.join(\n OUTPUT_DIR,\n \"prediction_heatmap.png\",\n)\n\n\nplt.savefig(\n viz2_path,\n dpi=180,\n bbox_inches=\"tight\",\n)\n\nplt.show()\n\nplt.close()\n\n\nprint(\n f\"Saved: {viz2_path}\"\n)\n\n\n# ============================================================\n# VISUALIZATION 3\n# Mean Prediction by Abnormality\n# ============================================================\n\nprint()\nprint(\"=\" * 70)\nprint(\"VISUALIZATION 3 — MEAN PREDICTION BY LABEL\")\nprint(\"=\" * 70)\n\n\nmean_predictions = (\n submission[LABELS]\n .apply(\n pd.to_numeric,\n errors=\"coerce\",\n )\n .mean()\n .sort_values(\n ascending=False\n )\n)\n\n\nplt.figure(\n figsize=(12, 6)\n)\n\n\nbars = plt.bar(\n mean_predictions.index,\n mean_predictions.values,\n)\n\n\nplt.xlabel(\n \"Knee abnormality\"\n)\n\nplt.ylabel(\n \"Mean predicted probability\"\n)\n\nplt.title(\n \"Mean Predicted Probability by Knee Abnormality\"\n)\n\nplt.xticks(\n rotation=45,\n ha=\"right\",\n)\n\n\n# ------------------------------------------------------------\n# Print values above bars\n# ------------------------------------------------------------\n\nfor bar, value in zip(\n bars,\n mean_predictions.values,\n):\n\n plt.text(\n bar.get_x()\n + bar.get_width() / 2,\n bar.get_height(),\n f\"{value:.3f}\",\n ha=\"center\",\n va=\"bottom\",\n fontsize=9,\n )\n\n\nplt.tight_layout()\n\n\nviz3_path = os.path.join(\n OUTPUT_DIR,\n \"mean_prediction_by_label.png\",\n)\n\n\nplt.savefig(\n viz3_path,\n dpi=180,\n bbox_inches=\"tight\",\n)\n\nplt.show()\n\nplt.close()\n\n\nprint(\n f\"Saved: {viz3_path}\"\n)\n\n\n# ============================================================\n# FINAL OUTPUT SUMMARY\n# ============================================================\n\nprint()\nprint(\"=\" * 70)\nprint(\"FINAL OUTPUTS\")\nprint(\"=\" * 70)\n\n\nprint()\nprint(\"1. Submission:\")\nprint(f\" {OUTPUT_FILE}\")\n\n\nprint()\nprint(\"2. Lookup cache:\")\nprint(f\" {CACHE_FILE}\")\n\n\nprint()\nprint(\"3. Prediction distribution:\")\nprint(f\" {viz1_path}\")\n\n\nprint()\nprint(\"4. Prediction heatmap:\")\nprint(f\" {viz2_path}\")\n\n\nprint()\nprint(\"5. Mean prediction by label:\")\nprint(f\" {viz3_path}\")\n\n\nprint()\nprint(\"=\" * 70)\nprint(\"VISUALIZATION COMPLETE\")\nprint(\"=\" * 70)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-09-14T08:28:25.569786Z","iopub.execute_input":"2026-09-14T08:28:25.570645Z","iopub.status.idle":"2026-09-14T08:28:28.070967Z","shell.execute_reply.started":"2026-09-14T08:28:25.570568Z","shell.execute_reply":"2026-09-14T08:28:28.070053Z"},"_kg_hide-input":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}