Dual-stream architecture with asymmetric clinical safety evaluation for colorectal histopathology — checkpoints and embeddings

De-identified artifacts for Nasser-eddine et al., "Dual-stream architecture with asymmetric clinical safety evaluation for colorectal histopathology" (submitted to Medical Image Analysis). They reproduce the paper's headline results without the raw whole-slide images (patient material, available on reasonable request only), for the three best configurations by ACSS (F4 selected, S4 and F6). Per model M (F4, S4, F6): M_head.pth is the stage-2 classifier head, and external_cache_M.pt holds the frozen 512-d and 4096-d embeddings of the additional 91-slide cohort. A single shared test_embeddings.npz holds the 4096-d embeddings of the 6894 held-out test patches, shared across F4/S4/F6, which use the same frozen encoder. Class/label order in the .npz is 0=ADE, 1=NOR, 2=CAR. Applying the head to the test embeddings reproduces the patch-level results (accuracy 92.37/93.39/92.67 percent; ACSS 95.92/95.88/95.56). Aggregating the 91-slide cache with a Gated Attention MIL (5-fold cross-validation over 5 seeds) reproduces the slide-level results (F4: 92.8 +/- 1.1 percent, zero CAR->NOR out of 205 carcinoma predictions). A self-contained notebook and the code are at https://github.com/oddissea/dualstream-crc