B-RAD inference checkpoints

Released inference weights for B-RAD, an interpretable multimodal retrieval-augmented diagnosis pipeline for breast ultrasound. B-RAD turns a single breast ultrasound image into an interpretable BI-RADS assessment by chaining four modules: BI-RADS-aware retrieval, retrieval-guided few-shot detection, foundational segmentation, and region-aware classification.

Paper: npj Digital Medicine (2026)
DOI: 10.1038/s41746-026-03272-3
Code: github.com/snuh-rad-aicon/B-RAD

Files

File	Size	Module	Contents
`b_rad_clip.pt`	1.98 GiB	B-RAD-CLIP	VPT-adapted BiomedCLIP image tower, BioLinkBERT-large text tower, both 512-d projection heads, the BI-RADS ordinal heads, and the learned temperature
`b_rad_full.pt`	1.98 GiB	B-RAD-FULL	the same encoder plus the 5-fold BI-RADS classification ensemble (`mlp_heads`) and its per-fold standard scalers
`b_rad_det.pt`	339 MiB	B-RAD-Det	fold-0 stage-2 detector: ViT-B/16 backbone, OWL-ViT-style detection head, and the four learned class queries

Usage

git clone https://github.com/snuh-rad-aicon/B-RAD && cd B-RAD
conda env create -f environment.yml && conda activate b-rad

bash B-RAD/ckpts/download_weights.sh    # fetches this repo, plus UltraSam.pth
bash B-RAD/scripts/inference_full.sh    # end-to-end on the bundled sample

notebooks/brad_inference_demo.ipynb runs the same pipeline with a visualization of every stage. The segmentation module needs its own OpenMMLab environment; see the repository README.

UltraSam.pth for B-RAD-Seg is not hosted here; download_weights.sh fetches it from the UltraSam authors.

Where each file is used

B-RAD/
├── B-RAD-CLIP/     infer_retrieval.py   --ckpt  b_rad_clip.pt
├── B-RAD-Det/      inference_single.py  --ckpt  b_rad_det.pt
├── B-RAD-Seg/      infer_bbox.py        --ckpt  UltraSam.pth
├── B-RAD-FULL/     infer_birads.py      --ckpt  b_rad_full.pt
└── ckpts/          download_weights.sh  (puts all four here)

Intended use and limitations

Research use only. These weights are not a medical device and must not be used for clinical decision making. They were trained on public breast-ultrasound cohorts plus a single-institution clinical cohort, so performance on other scanners, populations, or acquisition protocols is not characterized.

Citation

@article{jang2026interpretable,
  title     = {Interpretable multimodal retrieval augmented diagnosis for breast ultrasound with multinational clinical validation and reader study},
  author    = {Jang, Han and An, Subin and Jung, Sunghyun and Ji, Hye and Kim, Miran and Lee, Soochahn and Choi, Kyu Sung and Ha, Su Min},
  journal   = {npj Digital Medicine},
  year      = {2026},
  publisher = {Nature Publishing Group UK London},
  doi       = {10.1038/s41746-026-03272-3},
  url       = {https://doi.org/10.1038/s41746-026-03272-3}
}

License

CC BY-NC-SA 4.0, matching the code repository: non-commercial use only, with attribution, and derivatives under the same license. Upstream components retain their original licenses: BiomedCLIP, BioLinkBERT, OWL-ViT, and the timm ViT-B/16 initialization.

The methods behind these weights are covered by Korean patent application 10-2026-0096895. No patent rights are granted by this license. For commercial use or licensing, please contact the corresponding author.

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