Researchers at NYU Langone Health’s Perlmutter Cancer Center have built an AI system that looks at a woman’s mammograms not as a single snapshot but as a timeline, and the approach is proving better at forecasting who is actually likely to develop breast cancer in the next five years. The study, published in the American Journal of Roentgenology in September 2026, could reshape how doctors decide which women need more frequent scans, supplemental MRI, or earlier screening start ages, rather than applying the same schedule to everyone.
How the tool works
The model, called NYU-DRP, was trained on longitudinal digital breast tomosynthesis, meaning 3D mammograms collected from the same patients across multiple annual visits, rather than a single exam. By comparing how breast tissue patterns change year over year, the deep-learning system builds a personalized risk trajectory instead of a one-time score. The NYU Langone research team, whose findings were also detailed in a release from NYU Langone News and distributed via PR Newswire, tested the model against two alternatives: an AI tool that used only the single most recent 3D mammogram, and a separate AI model built on older-style 2D mammograms.
The numbers behind the claim
NYU-DRP correctly ranked women who went on to develop breast cancer within five years as higher-risk 72 percent of the time. The single-scan 3D tool got it right 70 percent of the time, and the 2D-based AI model was accurate 68 percent of the time. The gap is modest in absolute terms, but researchers say it is clinically meaningful at population scale, because even a few percentage points of improved risk stratification can mean thousands of additional high-risk women get flagged for supplemental screening — and, just as importantly, can mean lower-risk women are not routed into unnecessary extra scans, biopsies, and the anxiety that comes with them.
Why risk-based screening is the real target
Right now, most breast cancer screening guidelines are built around age and, in some cases, family history or genetic markers like BRCA mutations, with relatively little use of a woman’s own imaging history to fine-tune her schedule. The NYU team’s stated goal is to move toward what radiologists call risk-stratified screening: instead of every woman over 40 getting an annual mammogram on the same clock, a tool like NYU-DRP could identify which women should be screened more often, switch to MRI, or start screening earlier, while others might safely extend the interval between scans. That shifts the emphasis of AI in mammography away from simply reading a single image faster and toward using a patient’s imaging history as a forecasting tool.
Where the evidence still needs to grow
The NYU Langone team was explicit that this is an early, single-institution result. Its own release states that if future experiments in other groups of women with breast cancer confirm the finding, then AI-assisted longitudinal 3D mammography could help physicians tailor screening more precisely. That is a real caveat: the improvement over existing 2D and single-scan 3D tools has not yet been validated across other hospital systems, other imaging equipment, or more diverse patient populations, all of which have tripped up prior mammography AI tools that performed well in a single dataset but faltered elsewhere. Radiologists outside the study have separately cautioned, in coverage of other 2026 mammography AI research, that risk models trained predominantly on one health system’s patients can carry demographic blind spots that only show up once deployed more broadly.
What’s next
NYU Langone researchers say the next step is testing NYU-DRP prospectively, tracking new patients in real time rather than analyzing historical scans, and comparing outcomes across multiple hospital systems with different scanner hardware. If those studies hold up, radiology groups could begin using longitudinal AI risk scores to justify insurance coverage for supplemental MRI in women flagged as higher risk, a coverage gap that currently frustrates many patients whose risk falls just short of guideline thresholds for extra screening. For now, the tool remains a research finding rather than a bedside product, but it points to where AI in mammography is heading: away from single-exam pattern matching and toward tracking a patient’s own imaging history over time.
Photo: cottonbro studio / PEXELS via Pexels
