Monday, November 30 | 9:30 a.m. to 10:30 a.m. | M3-SSBR03-7 | Room S406A
Image-based risk assessment has recently been incorporated into National Comprehensive Cancer Network (NCCN) screening guidelines, but the best thresholds for putting it into practice remain unclear.
Leslie Lamb, MD, from Mass General Brigham in Boston will present research that included consecutive patients who underwent bilateral screening mammography at six sites from 2009 to 2022.
The researchers calculated risk with an image-only, deep-learning five-year model (MIRAI), Tyrer-Cuzick version 8, and the National Cancer Institute’s Breast Cancer Risk Assessment Tool (BCRAT) five-year and lifetime scores. The team assessed cancer detection rate (CDR), abnormal interpretation rate, positive predictive value, sensitivity, and specificity at NCCN thresholds (1.7% and 3.0% five-year risk, and above 20% lifetime risk), at AI percentile thresholds (top 5% to 25%), and at volume-matched thresholds.
The analysis covered 295,570 exams from 96,691 patients, with 2,033 cancers (0.7%). In this session, attendees will hear that AI outperformed both traditional models at every threshold. At the 1.7% threshold, AI flagged more exams as increased risk (66.2% vs. 41.5% for Tyrer-Cuzick and 40.6% for BCRAT) and had higher CDRs (8.5 vs. 6.5 and 6.0 per 1,000). At the 3.0% threshold, AI flagged 24.5% of exams versus 15.7% and 10.3%, with CDRs of 18.0 per 1,000 versus 7.3 and 5.5. Gains held in both dense and non-dense breasts, and AI kept higher CDRs when screening volume was matched. AI also better identified low-risk women, and the top 5% of AI-ranked exams had a CDR of 49.9 per 1,000.
Attend this session to learn more.
