An AI predicts 5-year breast cancer risk by reading how previous mammograms evolve

By Zendoric · October 11, 2026.

A team at NYU Langone Health and its Perlmutter Cancer Center has tested an artificial intelligence capable of estimating a woman's risk of developing breast cancer over the next five years, according to a study published in the American Journal of Roentgenology. The key does not lie in a new image: the model analyzes several years of 3D mammograms from the same patient —tomosynthesis technology, which captures breast tissue in layers and is increasingly common in screening programs— to detect how that tissue changes over time, instead of simply reading the latest scan. The system was trained on more than 313,000 mammograms from over 161,000 women, according to the study's own data.

The difference from current methods is one of approach, not just of tooling. The risk models used in the clinic today rely above all on family history and known genetic mutations. Connie Lehman, a breast imaging specialist at Harvard Medical School who did not take part in the study, tells Fox News Digital that many diagnoses arrive without any of those warning signs: "Over and over I would hear: 'how can this be? No one in my family has had breast cancer'". According to Lehman, AI provides a different source of information: clues within the tissue itself that traditional models do not capture.

That nuance has a practical consequence worth underlining: the system can analyze mammograms the patient has already had, with no need for a new test or additional radiation. If a woman gets a high-risk score, the next step would be to talk with her doctor about complementary tests, such as an MRI; if the risk is medium or low, she would continue with routine screening. Lehman is clear on one point: a high score is neither a diagnosis nor a certainty of future disease, and a low score does not eliminate risk or replace regular mammograms.

That said, it is best not to get ahead of ourselves. According to Lehman, the model itself is still a research tool, without FDA (Food and Drug Administration, the U.S. agency that regulates drugs and medical devices) clearance for clinical use. The study was built with data from NYU Langone hospitals and with equipment from a single mammography manufacturer, which limits what can be concluded about its performance in other populations, other screening protocols and other machines. The researchers themselves have announced that they plan to validate it with data from other medical centers and other 3D mammography systems before taking the next step.

In general, the recent history of diagnostic AI shows that this leap —from a promising single-center study to a reliable tool in diverse clinical practice— is the stretch where most projects fall by the wayside: what works with one hospital's tissue, equipment and population does not always hold up when the context changes. That NYU Langone presents it this way, as work in progress and not as a finished product, is a sign of rigor rather than a weakness.

Our reading is that modest, specific advances like this one are precisely the kind of progress that supports the underlying thesis we defend about AI in medicine: not the replacement of the doctor nor the spectacular headline, but ever earlier and more personalized detection of disease, built on data the healthcare system already had and that until now was used once and discarded. If tools of this kind pass external validation, breast cancer screening will stop being uniform —the same protocol for every woman— and will instead be adjusted to each woman's real risk, with fewer surprises like the ones Lehman describes. That is the long, unglamorous road toward the goal that really matters: not just moving diagnosis earlier, but getting closer to a future in which diseases like this one are managed routinely instead of arriving like an unexpected sentence.