nnunet-breast-mri

nnU-Net v2 — Breast DCE-MRI Segmentation Weights

Pretrained nnU-Net v2 weights for tissue segmentation on breast dynamic contrast-enhanced MRI (DCE-MRI). The accompanying inference/evaluation code is in code/; the online demo is the Space zhang0319/breast-mri-seg.

Four segmented classes:

Repository layout

tissue/   tissue-study models
region/   region-study models
  └── Dataset0XX_.../nnUNetTrainer_750epochs__nnUNetResEncUNetLPlans__3d_fullres/
        ├── fold_all/checkpoint_final.pth
        ├── plans.json
        ├── dataset.json
        └── dataset_fingerprint.json
code/      full inference + evaluation package (nnunet-breast-inference)
examples/  de-identified example cases (P0–P3 + ground-truth GT + reference predictions)

Model configuration (must match exactly when loading): trainer = nnUNetTrainer_750epochs · plans = nnUNetResEncUNetLPlans · configuration = 3d_fullres · fold = all

Quick start (local, GPU recommended)

# 1) Download this repo
hf download zhang0319/nnunet-breast-mri --local-dir ./nnunet-breast-mri

# 2) Install the code package
cd nnunet-breast-mri/code
pip install -e .

# 3) Edit the config: point models_dir at the downloaded weights (which contain tissue/ and
#    region/), and source_dir at your raw DCE-MRI (one folder per patient: P0.nii.gz is the
#    pre-contrast image, P1/P2/... are post-contrast timepoints).
cp config/inference_template.yaml config/my_site.yaml

# 4) Run inference (you can skip extract-models when the weights are already unpacked)
prepare-inputs    --config config/my_site.yaml
run-inference     --config config/my_site.yaml
format-outputs    --config config/my_site.yaml
build-breast-mask --config config/my_site.yaml

Each input channel is [pre, post, post − pre]. See code/docs/Inference.md for details.

Intended use and limitations

• Intended use: research, reproduction and teaching of breast DCE-MRI segmentation methods.
• Limitations: trained on specific acquisition protocols / populations; performance may drop on other scanners, sequences or populations, and it is sensitive to pre/post misregistration and motion artefacts.
• ⚠️ Not for clinical use: these weights and code are not a medical device and must not be used for clinical diagnosis or any medical decision.

License

The code is Apache-2.0 (see code/LICENSE). By using these weights you acknowledge and accept the intended-use limitations and disclaimer above.