Generalizable AI-driven cross-domain breast cancer diagnosis based on mammograms by solving Fourier transformation-based jigsaw puzzles

The performance of deep learning models for the early diagnosis of breast cancer based on mammograms often degrades when domain shifts occur. Although certain domain generalization methods can produce images with new styles for alleviating domain shifts, each produced image exhibits a single style. The present study introduces a Fourier transformation-based jigsaw puzzle (F-Jip) method which incorporates a jigsaw puzzle generation (JPG) module, a new style generation (NSG) module, and an enhanced jigsaw puzzle generation (EJPG) process to produce the enhanced jigsaw puzzles with multi-domain information. The enhanced jigsaw puzzles can guide the model to learn domain-invariant features to alleviate the influence of domain shifts for cross-domain breast cancer diagnosis. A leave-one-domain-out cross validation using six datasets which can simulate varying domain shifts encountered in clinical scenarios is used to evaluate the performance of the models. In each fold, the AUC and accuracy of the proposed model are higher than those of several state-of-the-art domain generalization models. Additionally, when compared to BI-RADS assessments of the radiologists in one dataset, the proposed model exhibits better performance for discriminating between the benign and malignant lesions, and further identifies 81.3 % patients with benign lesions in the subgroup analysis of patients with BI-RADS 4 lesions to avoid unnecessary biopsies. The experimental results indicate that the proposed model has the potentials for assisting doctors in diagnosing breast cancer.