Artificial Intelligence Versus Radiologist False-Positives on Digital Breast Tomosynthesis Examinations in a...

BACKGROUND. Insights into the nature of false-positive findings flagged by contemporary mammography artificial intelligence (AI) systems could inform the potential use of AI to reduce false-positive recall rates. OBJECTIVE. The purpose of this study was to compare AI and radiologists in terms of characteristics of false-positive digital breast tomosynthesis (DBT) examinations in a breast cancer screening population. METHODS. This retrospective study included 2977 women (mean age, 55 years) who were participating in an observational population-based screening study and underwent 3183 screening DBT examinations from January 2013 to June 2017. A commercial AI tool analyzed DBT examinations. Positive examinations were defined as having an elevated-risk result for AI and as having been assigned BI-RADS category 0 for interpreting radiologists. False-positive examinations were defined as the absence of a breast cancer diagnosis within 1 year. Radiologists rereviewed the imaging for AI-flagged false-positive findings. RESULTS. The false-positive rate was 10% for both AI (304/3183) and radiologists (308/3183). Of 541 total false-positive examinations, 233 (43%) were false-positives for AI only, 237 (44%) were false-positives for radiologists only, and 71 (13%) were false-positives for both. AI-only versus radiologist-only false-positives were associated with greater mean patient age (60 vs 52 years, p < .001), lower frequency of dense breasts (24% vs 57%, p < .001), and greater frequencies of a personal history of breast cancer (13% vs 2%, p < .001), prior breast imaging studies (95% vs 78%, p < .001), and prior breast surgical procedures (37% vs 11%, p < .001). The false-positive examinations included 932 AI-only flagged findings, 315 radiologist-only flagged findings, and 49 flagged findings concordant between AI and radiologists. AI-only flagged findings were most commonly benign calcifications (40%), asymmetries (13%), and benign postsurgical change (12%); radiologist-only flagged findings were most commonly masses (47%), asymmetries (19%), and indeterminate calcifications (15%). Of 18 concordant flagged findings that were biopsied, 44% yielded high-risk lesions. CONCLUSION. Imaging and patient-level differences were observed between AI and radiologist false-positive DBT examinations. Although only a small fraction of false-positive examinations overlapped between AI and radiologists, concordant flagged findings had a high rate of representing high-risk lesions. CLINICAL IMPACT. The findings may help guide strategies for using AI to improve DBT recall specificity. In particular, concordant findings may represent an enriched subset of actionable abnormalities.