G-Cascade-w34 — Two-stage cascade, S2 focal weight w=34
Reliability run 2026-08-31. MedSigLIP-448 ring-norm features. Seeds 42-46, frozen Source-D test (1,754 frames).
Mean test: NML R 0.319±0.074, NML P 0.292, MT R 0.859, NROI R 0.940, MCC 0.704±0.025.
- S1: frame MLP (lesion vs NROI), unweighted FocalLoss(g=3, ls=0.1)
- S2: per-frame MLP (MT vs NML) on lesion frames, WeightedFocal w=[1,34], ls=0.1
- No lesion aggregation. One variable changed vs unweighted cascade.
Per seed: seed<S>/{models/*.pt, metrics.json, preds_test.npz}. Harness: g_harness.py, script: run_cascade_w34.py.
Paths in the code are sanitized placeholders (, ).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support