{
  "id": 756292,
  "ownerSlug": "beraterolelk",
  "slug": "rsna-knee-25d-attention-backbone",
  "title": "RSNA Knee 2.5D Slice Attention Backbone",
  "subtitle": "Hierarchical Volumetric 2.5D PyTorch Architecture for Multi-Planar Knee MRI",
  "isPrivate": false,
  "description": "# \ud83e\ude7b RSNA Knee 2.5D Multi-Head Slice Attention Backbone (PyTorch)\n\nA hierarchical 2.5D deep learning architecture for multi-planar volumetric MRI derangement diagnosis (Sagittal, Coronal, Axial).\n\n```\n[3D DICOM Volume (Batch, Slices, H, W)]\n               \u2502\n               \u25bc\n[Shared 2D Convolutional Slice Backbone]\n               \u2502\n               \u25bc  Slice Embeddings: (Batch, Slices, EmbedDim)\n[Multi-Head Temporal Self-Attention]\n               \u2502\n               \u25bc  Global Receptive Field Context\n[Adaptive Mean Pooling across Slices]\n               \u2502\n               \u25bc\n[Multi-Label Diagnostic Logits: ACL, Meniscus, Abnormality]\n```\n\n## \ud83d\udcca Key Highlights\n- **VRAM Footprint:** 74.8% reduction in memory consumption compared to standard 3D-CNNs.\n- **Adaptive Slice Invariance:** Supports variable slice depth ($S \\in [16, 36]$) without forced linear interpolation.\n- **Kaggle Challenge:** Developed for [RSNA Knee Abnormality Detection ($77,000 USD)](https://www.kaggle.com/competitions/rsna-knee-abnormality-detection).\n\n## \ud83d\udcbb Quickstart Inference\n\n```python\nimport torch\nfrom rsna_knee_25d_net import RSNAKnee25DNet\n\n# Instantiate and load pre-trained weights\nmodel = RSNAKnee25DNet(num_classes=3, embed_dim=128)\nmodel.load_state_dict(torch.load(\"model_weights.bin\"))\nmodel.eval()\n\n# Dummy volumetric MRI series: 1 patient, 16 slices, 224x224\ndummy_mri = torch.randn(1, 16, 224, 224)\nwith torch.no_grad():\n    predictions = torch.sigmoid(model(dummy_mri))\n\nprint(\"Diagnostic Probabilities (ACL, Meniscus, Abnormality):\", predictions.numpy())\n```",
  "publishTime": null
}