Features of AI-Missed Cancers on Mammograms

Background: Mammographic sensitivity is 75% to 85%, but this decreases with dense breast parenchyma, perception and interpretation errors, and poor positioning. Initial efforts to use conventional computer-aided detection (CAD) in the 1990s did not improve the diagnostic performance of radiologists due to its false-positive rate and poor specificity. Artificial intelligence (AI) algorithms have been developed as an alternative to conventional CAD, improving the diagnostic performance of radiologists with AI assistance and as standalone programs. AI can miss breast cancers; however, with false-negative rates (FNRs) reported at 19.4%. AI misses more commonly in asymptomatic patients, those with extremely dense breasts, and ductal carcinomas in situ. There is little information on the features of invasive breast cancers missed by AI on mammograms. Objective: To evaluate the FNR of AI mammogram evaluation according to molecular subtype and to investigate the features of AI-missed cancers more...