The purpose of the study was to develop a Breast Imaging Reporting and Data System (BI-RADS®) breast density deep learning (DL) model in a multi-site setting for synthetic 2D mammography (SM) images derived from 3D DBT exams using FFDM images and limited SM data. The DL model was originally trained on a large dataset of FFDM images from an academic medical center in the midwestern United States. The model demonstrated closer agreement with the reporting radiologists than previous academic or commercial breast density products (Ours: 82%; Harvard [Lehman et al. 2019]: 77%; NYU [Wu et al. 2017]: 77%; Volpara [Brandt et al. 2016]: 57%; Quantra [Brandt et al. 2016]: 56%). Without adaptation, the model also demonstrated close agreement with the original reporting radiologists for SM images from two different sites (Site 1: 79%; Site 2: 76%). Performance further improved with adaptation using only 500 SM images from each site (Site 1: 81%; Site 2: 80%). These results suggest that the DL model can generalize to different institutions and from FFDM to SM images, which could lead to more consistent estimates of breast density for women.
Read more about how our BI-RADS breast density DL model demonstrated strong performance on mammography and tomosynthesis-based images from two institutions without modification and improved using a few images for site-specific localization.
