What AI-CAD Detects—and Misses—in Real-World Breast Cancer Screening

Abstract

AI for Breast Cancer Screening in Mammography (AI-STREAM) is a prospective multicenter trial in Korea aimed at assessing the real-world impact of AI-based computer-aided diagnosis (CAD) in population-based breast cancer screening for women aged 40 years (1).The performance of a system incorporating AI-CAD with the standard single reading was preliminarily evaluated (2).Chang et al. (3) then retrospectively examined the characteristics of cancers detected or missed using AI-CAD for secondary analysis.They analyzed 148 breast cancers in 24,543 women finding that the overall positive predictive value 1 of AI-CAD was 8.7% (133 of 1,535).The abnormality scores of AI-CAD were lower in patients presenting with mammographic asymmetry (p = 0.001) and the luminal A subtype (p = 0.032) than in patients presenting mass, microcalcifications or distortion, and other breast subtypes.AI-CAD detected 3.4% (5/148) of the cancers missed by radiologists.Most of the missed diagnoses occurred in cases of dense breasts with subtle findings, including two asymmetries, one architectural distortion, and one mass.All of these missed diagnoses were invasive ductal carcinomas, with 80% classified as the luminal A subtype.In contrast, AI-CAD missed 8.1% (12/148) of the cases that were detected by a radiologist at recall.False-negative findings also predominantly occurred

Ähnliche Arbeiten