Improved Breast Cancer Detection with Artificial Intelligence in a Real-World Digital Breast Tomosynthesis Screening Program.

Objective: The purpose of this study is to compare radiologists' breast cancer screening performance before and after the implementation of an artificial intelligence (AI) detection system for digital breast tomosynthesis (DBT).

Materials and methods: This retrospective study included 4 radiologists reading DBT screening mammograms across 3 clinical sites during 2 distinct time periods. The pre-AI time period from September 1, 2018 to August 31, 2019 included 10,322 standard DBT interpretations with a computer-aided detection system. The post-AI from January 1 to March 18, 2020 and May 4 to December 31, 2020 included 6,407 DBT interpretations with concurrent use of a deep learning AI support system. Endpoints included cancer detection rate (CDR), abnormal interpretation rate (AIR), and positive predictive values for cancer among screenings with abnormal interpretation (PPV1) and biopsies performed (PPV3). Estimates and 95% confidence intervals (CIs) for each radiologist were calculated for each time point and the difference across time periods.

Results: The CDR per 1000 exams increased from 3.7 without AI to 6.1 with AI (difference 2.4, P = .008, 95% CI: 0.6, 4.2). The AIR was 8.2% without AI and 6.5% with AI (difference -1.7, P < .001, 95% CI: -2.5, -0.8). The PPV1 increased from 4.2% to 8.8% with AI implementation (difference 4.6, P < .001, 95% CI: 3.0, 6.3) and PPV3 increased from 32.3% to 56.5% with AI support (difference 24.2, P = .033, 95% CI: 2.0, 46.4).

Conclusion: Real-world interpretation of DBT after implementation of an AI detection system resulted in increased CDR, reduced AIR, and significantly increased PPV1 and PPV3.