AI Model Uses Mammogram History to Predict Breast Cancer Risk

Researchers at NYU Langone Health have developed an artificial intelligence tool that offers a sharper look at breast cancer risk by analyzing a woman’s entire history of 3D mammograms. Unlike current methods that rely on a single scan or static patient data, this system, known as NYU-DRP, tracks how breast tissue changes over time. According to a study published in the American Journal of Roentgenology, this longitudinal approach correctly identified women at higher risk 72 percent of the time, outperforming both single-3D and 2D AI models.

The primary advantage of this tool is its ability to move beyond one-time snapshots. By learning from hundreds of thousands of annual scans, the model captures subtle developments in tissue structure that might be missed in an isolated image. This could allow doctors to tailor screening schedules more precisely, potentially reducing unnecessary tests for low-risk patients while increasing vigilance for those showing early signs of risk.

Longitudinal data improves risk accuracy

The study, reported by GN technics/ai, compared NYU-DRP against the Tyrer-Cuzick method, a widely used statistical model that relies on family history, age, and genetic factors. In a head-to-head comparison involving 432 women, the AI model was correct in 67 percent of cases, while the statistical method was accurate only 56 percent of the time. This suggests that the visual evolution of breast tissue over several years holds predictive power that traditional risk calculators do not fully capture.

The model was trained on a massive dataset of 313,531 yearly 3D mammograms from over 161,000 women who did not have breast cancer at the time of testing. This extensive data allowed the algorithm to identify complex patterns in tissue density and structure that correlate with future cancer development, providing a more nuanced risk assessment than standard clinical tools.

Breast density alone is insufficient

A significant finding from the research is that breast density, often considered the primary indicator of risk, is not a reliable standalone predictor. The study showed that among women with extremely dense breasts, the AI model classified nearly 38 percent as having average risk, and their actual cancer rate after five years was less than 1 percent. Conversely, the model flagged 15.5 percent of women with fatty breasts as high risk, and this group had a cancer incidence of 2.5 percent.

These results challenge the assumption that dense tissue automatically equates to high danger. Instead, the tool identifies specific structural changes that occur over time, which may be more indicative of future issues than density alone. This distinction is crucial for avoiding both over-diagnosis in low-risk dense-breast patients and under-diagnosis in others.

Future validation and clinical implementation

While the results are promising, the team emphasizes that further validation is required. The next step involves testing the tool on women who have already developed breast cancer to ensure its predictive power holds up in a broader clinical context. Researchers also plan to share the model with other academic health centers to cross-check the data, ensuring that the AI’s findings are robust and not specific to the NYU Langone patient population.

If future trials succeed, this technology could shift how screening is approached. Rather than a one-size-fits-all annual checkup, doctors might use these insights to customize the frequency and type of imaging for each patient. This personalized approach aims to maximize detection rates while minimizing the stress and cost associated with unnecessary supplemental tests.