We are sharing an open-source meta-repository for deep learning screening mammography classifiers. The resource is intended to promote standardization and reproducibility in research on machine learning applications to mammography for detection and diagnosis of breast cancer.
The growing number of machine learning models for screening mammography has made implementation and comparison of their performance on different datasets challenging. The meta-repository gives researchers straightforward tools to compare how a given model performs across datasets and how various models perform on the same set of data, be it public or private.
At the time of publication, the meta-repository contains five state-of-the-art machine learning models for breast cancer exam classification. Researchers can use their own data representative of particular target populations to compare how different models perform.
Developers can submit their own models to the meta-repository in order to make their work available to the machine learning research community and to compare the performance of their models with that of other models.
The meta-repository is a cross-platform solution and works on various hardware setups. Containerization of deep learning models with Docker allows users to generate predictions on most machines, with or without GPU acceleration.
Meta-repository of screening mammography classifiers.
Please cite this work if you are using the meta-repository in your research.
Neural Networks Improve Radiologists’ Performance in Breast Cancer Screening.
IEEE Trans Med Imaging. 2020 Apr;39(4):1184-1194. doi: 10.1109/TMI.2019.2945514
An interpretable classifier for high-resolution breast cancer screening images utilizing weakly supervised localization.
Med Image Anal. 2021 Feb;68:101908. doi: 10.1016/j.media.2020.101908
Weakly-supervised high-resolution segmentation of mammography images for breast cancer diagnosis.
Proc Mach Learn Res. 2021. Medical Imaging with Deep Learning. openreview.net/forum?id=nBT8eNF7aXr
Questions about this resource may be directed to Krzysztof Geras or raised as issues on Github.