Downloads · 30 days
0
happytommy/breast-us-cad-weights
breast-us-cad-weights is a machine learning model from happytommy. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Research prototype — not for clinical use. Not a medical device.
Downloads · 30 days
0
Access
Public
Updated Sep 7, 2026
Repo size
5.2 GB
Likes
0
Public
Click a slice to open those files.
.pt5.2 GB · 100%
From the Hugging Face model README
Research prototype — not for clinical use. Not a medical device.
Frozen artifacts of the BreastUS-CAD demo (Space): five ViT-B/16 cross-validation fold checkpoints (BUS-BRA official patient-level folds), a U-Net (efficientnet-b0 encoder) lesion segmentation demo model, the temperature-scaling parameter, and the frozen decision threshold. Trained on the public BUS-BRA dataset (Gómez-Flores et al., Medical Physics 2024). These root-level files are the ONLY artifacts the Space and the frozen-v1 external validation use.
Checkpoints of the pre-registered follow-up studies (repository
data/external_protocol.md, Amendment 4; results in RESULTS.md):
v2/v2_biomedclip_fold{1-5}.pt + v2/v2_biomedclip_{calibration,operating_point}.json
— Q2: same v1 protocol with the backbone initialised from the BiomedCLIP
ViT-B/16 image encoder. Pre-registered criterion NOT MET (verdict:
"domain pretraining reduces shift" is NOT CLAIMED).v2/v2_loco_{breast,busi,gdph,sysucc}.pt + per-run
_calibration.json / _operating_point.json — Q1: leave-one-cohort-out
multi-source single models (BUS-BRA folds 1–4 + 85% of the three other
external cohorts). Criterion met via the AUC branch only; see RESULTS.md
for the caveats stated alongside.v2/pretrained/biomedclip_vitb16_timm.pt — the BiomedCLIP vision tower
re-exported into timm vit_base_patch16_224 layout (initialisation only;
derived from microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224,
subject to its licence).None of the v2 files is used by the demo. They are published so that the v2 results can be reproduced at inference level from the same frozen keep-lists, which an independent audit (2026-09-06) found impossible while they existed only on the author's disk.
CHECKSUMS.txt lists the SHA-256 of every file in this repo in
sha256sum format (paths as in the repository's models/ directory;
v2_*/pretrained/* entries live under v2/ here).
inference.resolve_weight in the repository resolves both layouts.