Obaidullahmiakhil/ich-phase2-models

ICH Phase-2 — trained models & artifacts

RSNA-2019 winner-style cascade (2D CNN -> per-slice embeddings -> BiGRU) with three novelties (conformal risk control, cross-slice CAM-consistency, long-tail loss). Uploaded 2026-07-17.

Contents

• checkpoints/ — CNN + BiGRU heads (cnn_best.pt, seq_winner_best.pt, seq_gate_best.pt, seq_sub_best.pt)
• scores/ — val/test score arrays (*.npz) consumed by Step 3 (conformal + manuscript)
• embeddings/ — per-slice CNN embeddings (emb.npy, emb_ids.json) for regenerating sequence models
• run_config.json, class_stats.json — full config + data statistics

Test-set results (macro-AUC)

• BiGRU winner reproduction: 0.933 | our cascade: 0.932 | 2D-CNN baseline: 0.856
• Epidural (rare class) AUC: 0.924