Domain generalization in deep learning based mass detection in mammography: A large-scale multi-center study

作者：

Highlights：

• Best practices on domain generalization in deep learning based breast cancer detection.

• Eight deep learning based detection methods, including Transformer-based architectures.

• Analysis of the domain shift present in six digital mammography datasets.

• Comparison of mass and breast attributes, highlighting the biases of detection.

• The single-source domain generalization pipeline boosted the AUC from 0,79 to 0,89.

摘要

•Best practices on domain generalization in deep learning based breast cancer detection.•Eight deep learning based detection methods, including Transformer-based architectures.•Analysis of the domain shift present in six digital mammography datasets.•Comparison of mass and breast attributes, highlighting the biases of detection.•The single-source domain generalization pipeline boosted the AUC from 0,79 to 0,89.

论文关键词：Domain generalization,Digital mammography,Breast cancer,Transformer-based detection,Transfer learning,Data augmentation

论文评审过程：Received 26 January 2022, Revised 7 August 2022, Accepted 19 August 2022, Available online 24 August 2022, Version of Record 12 September 2022.

论文官网地址：https://doi.org/10.1016/j.artmed.2022.102386