The addition of digital breast tomosynthesis (DBT) to digital mammography has been shown to improve cancer detection rates [1-4] and lower recall rates [2-7] but prolongs reading time almost two-fold [1, 8, 9], compared to digital mammography alone. Therefore, an artificial intelligence (AI) system (ProFound AI™ ver 2.0; iCAD, Inc., Nashua, NH/US) based on deep learning was developed to detect malignant soft tissue and calcific lesions in DBT exams. The algorithm provides outlines of detected lesions in the DBT slices and calibrated lesion-level and case-level scores from 0-100 to indicate the algorithm’s confidence that a finding or case is malignant. The AI outlines and scores were designed to be used concurrently by radiologists while interpreting DBT exams.

This study evaluated the specific impact of concurrent use of the AI system with DBT in a large, retrospective reader study based on mammographic appearance and histopathology. Subgroup analyses assessed reader performance for cancer cases with soft tissue lesions (with or without calcifications) versus cases with only calcifications and for invasive carcinomas versus ductal carcinomas in situ (DCIS). For non-cancer cases, the subgroups consisted of cases with suspicious, but non-malignant soft tissue lesions (with or without calcifications), cases with only suspicious, non-malignant calcifications and cases without any suspicious lesions that were interpreted at the case acquisition sites as Breast Imaging Reporting and Data System (BI-RADS) 1 or 2. Non-cancer cases with suspicious lesions included cases with biopsy-proven benign lesions and cases with lesions shown not to warrant biopsy based on additional imaging.