AI model may guide supplemental breast imaging decisions

A deep-learning mammography model may help identify which women are most likely to benefit from supplemental breast imaging, according to a study published in JAMA Network Open.

The study compared the Mirai deep-learning breast cancer risk model with BI-RADS breast density assessments for estimating 2 outcomes: future breast cancer within 5 years and false-negative screening results. The authors noted that federal density-notification rules have increased the need for more precise risk tools because breast density alone applies to a large share of women and is subjectively assessed.

Researchers analyzed 123,091 screening mammograms from 67,019 women at 5 sites within a large academic health system between 2009 and 2018. The median patient age was 58, and 41.4% of mammograms were from women with dense breasts, according to the published study summary.

Mirai classified 5-year breast cancer risk as low, intermediate, or high. The risk categories were defined as less than 1.7%, 1.7% to 3%, and greater than 3%, respectively. The study then compared those risk assignments with breast density categories and follow-up cancer outcomes.

The model outperformed density alone for predicting 5-year breast cancer risk. Diagnostic Imaging reported an area under the receiver operating characteristic curve of 0.71 for the deep-learning model, compared with 0.53 for breast density classification.

Cancer incidence increased across the model’s risk categories. ResearchGate’s indexed version of the article reports cancer incidence of 1.0% in the low-risk group, 2.7% in the intermediate-risk group, and 6.2% in the high-risk group. The association between density and cancer incidence was less pronounced.

False-negative screening results also increased across Mirai risk groups, especially among women with dense breasts. The study found that adding breast density to the deep-learning model did not improve performance, suggesting that density-related imaging features may already be captured in the model’s risk estimates.

The authors said current binary density-based policies may miss some high-risk women with nondense breasts while sending some low-risk women with dense breasts toward additional imaging. They wrote that more precise risk stratification tools are needed “to achieve more personalized and equitable screening.”

The findings point to a possible role for AI-based mammography risk tools in deciding who should receive supplemental imaging, rather than relying on density status alone. That question has become more important as patients receive breast-density notifications and clinicians weigh the benefits, false positives, costs, and access limits of additional imaging.

Further work will be needed to determine how deep-learning risk models should be incorporated into clinical pathways, reimbursement rules, and shared decision-making for supplemental breast imaging.