CM3070 breast cancer detection

Coursework checkpoints for a weight-shared two-view EfficientNet-B3 at 576x448, scored on 291 paired CBIS-DDSM JPEG exams. These are not a clinical device.

The code, notebooks, and report are at https://github.com/cheeseballs22/CM3070_Breast_Cancer_Detection.

Exam models

Rank by test ROC-AUC.

patch_classifier/stage2/best_val.pt is the patch exam model, phase 1 epoch 7. Flip-TTA ROC-AUC 0.851.
direct_classifier/stage2/best_val.pt is the direct classifier exam model, phase 1 epoch 3. Flip-TTA ROC-AUC 0.830.

Other checkpoints

patch_classifier/stage1/best_val.pt is patch Stage 1.
patch_classifier/design_b/best_val.pt is the 300x300 three-class head. It is not the exam model.
direct_classifier/stage1/best_val.pt is direct classifier Stage 1.

Download into the weights/ folder of the GitHub repo:

hf download ruiner/CM3070_Breast_Cancer_Detection --local-dir weights

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support