Subgroup performance of a commercial digital breast tomosynthesis model for breast cancer detection

AI models show potential to improve breast cancer screening, however detailed subgroup evaluations to uncover the strengths and weaknesses of models are lacking. This study presents a granular evaluation of a commercial AI model for cancer detection on digital breast tomosynthesis (DBT) on a retrospective cohort of 167,860 screening exams in female patients. Performance in distinguishing screen detected cancers (1,368 exams) from negative exams (166,387 exams) is stratified across demographic, imaging, and pathologic subgroups to identify disparities. The overall AUROC is 0.91 and sensitivity is 0.73 with robust performance across demographics. In-situ cancers (AUROC: 0.85, sensitivity: 0.55), calcifications (AUROC: 0.80, sensitivity: 0.66), and dense breast tissue (AUROC: 0.88, sensitivity: 0.63) are associated with lower performance, while masses (AUROC: 0.93, sensitivity: 0.85) and architectural distortions (AUROC: 0.90, sensitivity: 0.83) are associated with higher performance. These results highlight the need for detailed evaluations and vigilance in adopting new clinical tools. Established Artificial intelligence (AI) models have shown potential for improving breast cancer screening outcomes using mammography, however, their performance across various subgroups of patient demographics, imaging characteristics and pathology subtypes remains unclear. In here, the authors find that the application commercial AI model (Lunit INSIGHT DBT) for breast cancer detection on digital breast tomosynthesis (DBT) demonstrates robust performance across demographics, however, this is reduced for Ductal Carcinoma in Situ (DCIS), calcifications, and dense breasts.