BCRBench: A Breast Cancer Risk Benchmark

Purpose: Image-based breast cancer risk prediction holds strong potential for personalized screening intervals, yet no standardized benchmark exists for fair comparison across methods. Existing implementations rely on separate codebases with inconsistent preprocessing and evaluation pipelines, limiting reproducibility. We present href{https://github.com/sot176/BCRBench}{BCRBench}, an open-source repository unifying state-of-the-art models within a single reproducible framework for single- and multi-timepoint risk prediction, providing a structured foundation for future research and standardized development in image-based breast cancer risk modeling. Comprehensive documentation is available at https://sot176.github.io/BCRBench.

Material and Methods: BCRBench is structured in a modular manner, enabling easy extensibility through a unified interface that ensures consistent benchmarking across models while enabling integration of new datasets, models, and evaluation protocols. Within this framework, we currently integrate two datasets, EMBED and CSAW-CC, and five models, Mirai, VMRA-MaR, OA-BreaCR, ImgFeatAlign, and LMV-Net, into a unified preprocessing, training, and evaluation pipeline implemented in PyTorch. BCRBench supports image, breast, and patient-level risk prediction from standard mammographic views (CC and MLO). Notably, three models (VMRA-MaR, ImgFeatAlign, and LMV-Net) build upon the Mirai encoder, and the framework optionally supports end-to-end fine-tuning. Performance is evaluated using AUC and C-index metrics, with additional support for subgroup analyses based on breast density and cancer subtype.

Results: Standardized training and evaluation reduced implementation variability across codebases, enabling consistent and reproducible benchmarking. All five models were successfully reimplemented and achieved competitive performance on EMBED and CSAW-CC. Longitudinal models outperformed single-timepoint approaches, highlighting the value of multi-timepoint information. The modular architecture also enabled integration of additional models, demonstrating the framework’s extensibility.

Conclusion: BCRBench harmonizes preprocessing, training, inference, and evaluation across single-timepoint and longitudinal breast cancer risk prediction models, enabling fair and reproducible benchmarking. By lowering the barrier to integrating and evaluating new methods, the framework aims to accelerate progress toward clinically reliable personalized risk-based mammographic screening.