{
  "id": 745786,
  "ownerSlug": "yunusgmsoy",
  "slug": "rsna-knee-radhead-resnet50",
  "title": "rsna-knee-radhead-resnet50",
  "subtitle": "",
  "isPrivate": false,
  "description": "# RSNA Knee RadImageNet Attention Heads \u2014 Yunus (5-Fold)\n\nTrained 5-Fold RadImageNet Attention Heads on top of the frozen ResNet-50 medical imaging encoder for the RSNA Knee Abnormality Detection competition.\n\n---\n\n## \ud83c\udfd7\ufe0f Architecture & Training Configuration\n\n* **Encoder:** Pre-trained RadImageNet ResNet-50 (1.35M medical scans, frozen).\n* **Head:** `RadHead` \u2014 LayerNorm + Linear projection (2048 to 512) + Learned Plane Embeddings (3 slots) + Slice Position Embeddings (8 slices) + Multi-Head Cross-Attention (8 heads, 512-dim) + Dual Residual Pooling (mean & max).\n* **Supervision:** Trained with confidence-weighted binary cross-entropy loss against the **V5 4-Source Master Consensus Labels** (`report_labels_v5.csv`, 0.8950 Gold Macro-AUC).\n* **Cross-Validation:** 5-Fold Stratified Cross-Validation on all 4,407 training studies.\n* **Validation Performance:** **`0.8251` 5-Fold OOF Macro-AUC** on the 58 official gold ground-truth studies.\n\n---\n\n## \ud83d\udcc1 Included Artifacts\n\n1. `v52_rad_heads_yunus.pt` (60.6 MB): PyTorch weights bundle containing all 5 fold heads with metadata.\n2. `v52_oof_yunus.csv` (4,407 rows \u00d7 13 cols): Full Out-of-Fold probability matrix for meta-stacking and calibration.\n\n---",
  "publishTime": null
}