A new artificial intelligence (AI) model that analyses mammogram images has demonstrated significantly greater accuracy than traditional breast density assessments in predicting a woman’s five-year risk of developing breast cancer. The breakthrough findings were presented at the annual meeting of the Radiological Society of North America (RSNA), held from November 30 to December 4 in Chicago.

The study, led by Dr. Constance D. Lehman, professor of radiology at Harvard Medical School, evaluated how an image-only AI model compared with breast density — a commonly used clinical indicator — in forecasting future breast cancer risk.

To develop the model, researchers trained a deep convolutional neural network using 421,499 mammograms collected from 27 medical facilities. After calibration on an independent dataset, the model was tested on a massive cohort of 245,344 screening mammograms performed between 2011 and 2017. The AI system then generated five-year cancer risk scores, categorised into average, intermediate, and high risk based on National Comprehensive Cancer Network (NCCN) thresholds.

Breast density — long considered a key predictor of breast cancer — showed only a modest association with risk. Women with higher density had a hazard ratio of 1.16, indicating a slight increase in cancer likelihood.

In contrast, the AI model provided a far more nuanced and powerful risk stratification. Women flagged as intermediate risk by AI showed more than double the cancer risk (hazard ratio 2.06) compared to the average-risk group, while those in the high-risk category had over four times the risk (hazard ratio 4.49). These associations remained strong even after adjusting for breast density.

When both AI predictions and breast density were analysed together, density added only a minimal increase in predictive value (adjusted hazard ratio 1.12). Meanwhile, the AI model’s risk categories remained nearly unchanged, with adjusted hazard ratios of 2.04 for intermediate risk and 4.47 for high risk.

Dr. Lehman said the findings underscore the transformative potential of AI in preventive oncology. “The model is able to detect changes in the breast tissue that the human eye can’t see,” she noted, highlighting the technology’s ability to capture subtle, early-stage patterns invisible in conventional screenings.

The study marks a significant step forward in personalised breast cancer screening, suggesting that AI-driven risk prediction could eventually help clinicians tailor surveillance strategies, improve early detection, and better protect women at heightened risk.