Refining breast cancer risk stratification for selection of women at intermediate to high risk for supplemental MRI screening using artificial intelligence.

10564 Background: Supplemental MRI screening for women at increased breast cancer risk can detect additional cancers that are occult on mammography. However, its adoption is limited by restricted availability and high false-positive rates. Improved risk stratification is therefore needed to optimize the use of supplemental MRI screening. This study aimed to evaluate the feasibility of a mammography-based artificial intelligence (AI) system for stratifying women at intermediate to high risk for supplemental MRI screening. Methods: This retrospective study included consecutive women who underwent abbreviated breast MRI following negative mammography between January 2020 and December 2024. The study cohort comprised women at intermediate to high breast cancer risk, including those with known genetic predisposition, personal history of breast cancer, prior benign or high-risk breast biopsy, or family history of breast cancer. A commercially available AI system was applied to all mammograms, generating both exam-level and breast-level scores representing the likelihood of a malignancy on a scale of 0 to 100. MRI outcomes, false-positive examinations, cancer detection, and positive predictive value (PPV) were evaluated. Univariate analyses were performed to explore associations between breast density and AI-based selection results. Results: In 1,588 women (median age, 51 years; range, 24-87), 1,804 screening mammograms and MRI examinations were performed. Of these examinations, 220 resulted in recall, yielding 24 mammographically occult breast cancers (12 ductal carcinoma in situ and 12 invasive cancers) diagnosed within one year and 166 false-positive examinations. The AI system achieved an area under the curve (AUC) of 0.694 (95% CI, 0.499-0.888) in women without a personal history of breast cancer and 0.611 (95% CI, 0.430-0.792) in women with a personal history of breast cancer. Using AI-based selection, 239 of 1,588 women (14.9%) were selected for supplemental MRI, resulting in an 81% reduction in false-positive examinations (31 vs 166) and an improvement in PPV to 20.5% (8 of 39). 6 of 8 detected cancers (75%) were invasive. Breast density was not associated with the AI-based selection strategy. Conclusions: Mammography-based AI substantially reduced false-positive supplemental MRI examinations while preferentially selecting women at higher risk for invasive breast cancers, without compromising cancer detection. These findings suggest that AI can refine risk stratification and optimize selection of women to improve the efficiency of supplemental MRI screening.