{
  "id": 737010,
  "ownerSlug": "riachk",
  "slug": "smart-rsna-mri-detection-tool",
  "title": "Smart RSNA MRI Detection Tool",
  "subtitle": "Faster Reads & Sharp Precision For Every Knee Scan",
  "isPrivate": false,
  "description": "# Model Summary\n\nThis model is a 2.5D vision system designed for the RSNA Knee Abnormality Detection competition. It processes multi-plane MRI volumes (sagittal, coronal, axial) using a PyTorch vision backbone paired with Multiple Instance Learning (MIL) attention heads. It predicts 12 diagnostic target abnormalities simultaneously across full knee MRI series.\n\n## Usage\n\nThe model accepts multi-slice 2D DICOM series preprocessed into 3D volume tensors.\n\nInput Shape: (batch_size, channels, depth, height, width) \u2014 e.g., (1, 3, 32, 224, 224) per plane.\nOutput Shape: (batch_size, 12) representing probability scores for the 12 knee abnormality classes.\n\nKnown failures include severely corrupted DICOM headers or missing MRI view sequences.\n\n```python\nimport torch\n\nmodel = torch.load(\"model.pth\")\nmodel.eval()\n\n# Example input tensor (batch, channels, depth, height, width)\nx = torch.randn(1, 3, 32, 224, 224)\nwith torch.no_grad():\n    predictions = model(x)\n\n## System\n\nThis model operates as an inference engine inside a Kaggle evaluation pipeline. Inputs are DICOM image files from the competition dataset, and downstream outputs are formatted into submission.csv for competition scoring.\n\n## Implementation requirements\n\n* Hardware: Trained using NVIDIA GPUs (T4 / P100 / V100). Inference executes within Kaggle Notebook GPU environments under the 9-hour runtime limit.\n * Software: Python 3.10+, PyTorch, PyTorch Image Models (timm), PyDICOM, NumPy, Pandas.\n\n# Model Characteristics\n\n## Model initialization\n\nFine-tuned from pre-trained vision weights (DINOv2 / EfficientNet) with custom-initialized MIL classification heads.\n\n## Model stats\n\n* Size: ~100M - 300M parameters (~400MB - 1.2GB weight file size).\n * Latency: ~0.2 - 0.5 seconds per MRI volume on GPU.\n\n## Other details\n\nThe model is not quantized or pruned. Standard PyTorch FP16 mixed precision is used for speed during inference.\n\n# Data Overview\n\n## Training data\n\nTrained on the official RSNA Knee Abnormality Detection dataset comprising multi-planar MRI series with associated multi-label abnormality annotations. Preprocessing includes 3D slice windowing, resizing, and intensity normalization.\n\n## Demographic groups\n\nData originates from anonymized clinical imaging studies provided by RSNA contributing medical centers.\n\n## Evaluation data\n\nEvaluated using a stratified 5-fold cross-validation split based on study IDs to prevent data leakage between train and validation sets.\n\n# Evaluation Results\n\n## Summary\n\nAchieves strong log-loss and AUC performance across all 12 target classes on internal 5-fold cross-validation.\n\n## Subgroup evaluation results\n\nEvaluated across sagittal, coronal, and axial views to ensure robust prediction regardless of view completeness.\n\n## Fairness \n\nEvaluated using standard multi-label cross-entropy loss and ROC-AUC metrics across all target categories.\n\n## Usage limitations\n\nIntended strictly for research and Kaggle competition evaluation. Not certified for clinical medical diagnosis.\n\n## Ethics\n\nTrained exclusively on fully anonymized, publicly released competition data in compliance with medical data privacy guidelines.",
  "publishTime": null
}