Sci Transl Med/Sci Adv
AI mammogram risk scores outperform existing breast cancer tools
Clinical Takeaway: Image-derived AI risk scores may improve identification of women who'd benefit from prevention or closer screening. Broader validation is needed before clinical adoption.
Current breast cancer risk models rely on lifestyle and family history rather than the mammogram itself and were largely built from White cohorts that limit performance in non-European populations. Two studies evaluated whether AI can sharpen risk assessment overall and for underrepresented groups.
A ten-year AI breast cancer risk model substantially outperformed the standard clinical tools on long-term prediction. In the top 10% of women flagged as highest risk at baseline, the AI model captured 33% of cancers that later developed, versus 20% to 24% for Tyrer-Cuzick, BCSC, and the 5-year Mirai model. Applied to NICE clinical thresholds, it flagged 31% of future cancers as high-risk at baseline in the Swedish KARMA cohort, compared with 7% for Tyrer-Cuzick-v8 and under 1% for BCSC-v3.
The model used image data from all four standard mammogram views combined with age at the scan. It was developed in Sweden's KARMA cohort and validated across 8,696 women including 1,633 with incident breast cancer, most of European ancestry, with external validation in Olmsted County, Minnesota, and the EMBED dataset in Atlanta.
"This stresses the need for further evaluating the model in diverse populations and for its intended use before considering its clinical use," the authors wrote.
A separate image-based score, which summarizes breast-tissue texture features from the mammogram, performed consistently across racial and ethnic groups. It was about as strongly linked to breast cancer risk in non-Hispanic Black, East Asian, South Asian, and Indigenous women as in non-Hispanic White women, with similar score distributions and strong calibration in every group evaluated. The mammogram risk score (MRS) was tested in more than 226,000 women across two North American cohorts.
Together, the findings point toward image-based risk scoring as a way to use existing screening infrastructure for risk stratification, without collecting lifestyle and family history data.
Sources: Eriksson M. Sci Transl Med. 2026 May 20. A long-term image-derived AI-based risk model for primary prevention of breast cancer in individuals at high risk; Jiang S. Sci Adv. 2026 May 20. Generalizability of an AI-based mammogram risk score (MRS) for breast cancer among diverse populations of women