An evidential deep learning framework for assessment of mammograms

Abstract In this study, we present an evidential deep learning framework called MV-DEFEAT, incorporating the strength of Dempster-Shafer evidential theory and subjective logic, for various mammogram assessment tasks, including mammogram density assessment, BIRADS scoring, and mammogram finding as normal/benign/malignant. The framework combines evidence from multiple mammogram’s views to mimic a radiologist’s decision-making process. We conducted experiments on two open-source digital mammogram datasets, VinDr-Mammo: A large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography and the Digital Database for Screening Mammography (DDSM), which include 5,000 and 1,975 mammogram examinations, respectively. Our results showed that MV-DEFEAT achieves weighted macro-average area under the receiver operating characteristic curves (AUCs) of 98.92 ± 0.007 and 86.53 ± 0.051 for mammogram density assessment on the VinDr-Mammo and DDSM datasets, respectively, which are considerably better than a state-of-the-art multi-view deep learning method with AUCs of 86.38 ± 0.258 and 70.62 ± 0.261. In addition, MV-DEFEAT demonstrated strong generalization for mammogram density assessment on two independent and unseen datasets, the Chinese Mammography Database (CMMD) and Virtual Tissue Biobank (VTB). Furthermore, we demonstrated the efficient feature representation of the MV-DEFEAT that could be transferred to a breast cancer stage classification task using the unseen Kuopio Osteoporosis Risk Factor and Prevention study (OSTPRE) dataset. Our results enhance trust in using deep learning algorithms to assess mammograms in clinical settings.