BREST-AI: three-year breast cancer risk from screening mammograms

This release provides the BREST-AI 16-bit, four-view model developed at the University of Warwick with NHS and academic collaborators. It estimates an examination-level three-year risk score from a screening mammogram. The model uses a ResNeXt-50 encoder and attentional feature fusion.

Paper: Wei J, Zhou X, Santeramo R, et al. BREST-AI: Three-year breast cancer risk prediction from screening mammograms. npj Breast Cancer (2026). doi:10.1038/s41523-026-01069-x.

Code and preprocessing: github.com/gmontana/BREST.

Download and run

git clone https://github.com/gmontana/BREST.git
cd BREST
python -m pip install -r requirements-published-cuda.lock
python scripts/download_weights.py --output-dir models
python scripts/patientLevel-inference.py \
    --metadata_csv /data/examinations.csv \
    --image_root_dir /data/processed_png \
    --model_checkpoint models/brest-ai-v1.safetensors \
    --final_csv_path /data/predictions.csv

Use Python 3.11 on Linux. The locked environment pins PyTorch 2.1.0 / CUDA 11.8, Kornia 0.7.1 and their dependencies. Use this same installation for CPU inference with --device cpu.

The repository supplies the model definition and checks the downloaded weights against a fixed SHA-256 checksum. No remote Python code needs to be executed from this model repository. GPU inference is selected when available; add --device cpu for CPU inference.

Inputs and output

Use the repository's documented DICOM preprocessing to produce breast-aligned 2048 × 2048 grayscale 16-bit PNGs. Select paper-hologic for 16-bit rendering with Lanczos resizing, or paper-siemens to reproduce the original Siemens route of 8-bit rendering expanded to uint16 followed by NumPy bilinear resizing. Both profiles use a fixed background seed of 42. Inference divides pixels by 65535, repeats the grayscale channel three times, resizes to 1792 × 1792 using the original Kornia bilinear antialiased resize, and normalises with mean 0.0856 and standard deviation 0.1687 per channel. Right-breast flipping belongs to DICOM preprocessing; inference does not flip the prepared images again.

Each examination requires one image for each of left CC, right CC, left MLO and right MLO, in that network order. The CSV supplies ClientID, path, ViewPosition and ImageLaterality; EpisodeID distinguishes repeat examinations. The loader validates and orders these views. Outcome labels are optional for prediction. For an original evaluation CSV whose rows are already ordered, --preserve_row_order reproduces the original loader without sorting historical view tags.

The output is the sigmoid of one examination-level risk logit. Population calibration is required before interpreting it as an absolute three-year probability. Its value is not the probability that an MRI will detect a cancer immediately.

Evaluation and scope

The published retrospective UK Hologic case–control evaluation included 7,596 women (1,899 cases); BREST-AI AUC was 0.751 (95% CI 0.738–0.764), compared with 0.700 (0.687–0.714) for MIT's Mirai. The paper also reports Siemens and US external evaluations. Acquisition differences can affect performance; evaluate the model and calibration in the intended population.

Release contents and privacy

The weights are stored as architecture-validated model tensors in safetensors format. The release includes the input configuration, checksums, source-checkpoint hash and the software licence. Patient images, patient-level metadata, predictions, training logs, optimiser state and institutional file paths are excluded. Examples in the code repository are synthetic metadata.

These file-level controls do not constitute a differential-privacy guarantee for the learned parameters. The model was trained using controlled-access research data; those data are obtained separately through their custodians.

Licence and citation

The University of Warwick BREST software licence covers the code and neural-network weights. It permits academic research and non-commercial, non-clinical evaluation. Retain the complete licence and copyright notice on redistribution. Commercial licensing enquiries: ventures@warwick.ac.uk. Research enquiries: g.montana@warwick.ac.uk.

@article{wei2026brest,
  title = {BREST-AI: Three-year breast cancer risk prediction from screening mammograms},
  author = {Wei, Jiefei and Zhou, Xinyu and Santeramo, Ruggiero and Al-Attar, Miaad and Muthyala, Sreenivas and Gilbert, Fiona and Brentnall, Adam R. and Montana, Giovanni},
  journal = {npj Breast Cancer},
  year = {2026},
  doi = {10.1038/s41523-026-01069-x}
}

Downloads last month: 23