What Years of 3D Mammograms Can Reveal About Future Risk

A mammogram usually answers a question about the present: is there something here that needs a closer look? Researchers at NYU Langone Health are using those images to investigate a question about the future: who is more likely to develop breast cancer over the next five years?

The difference matters. A clearer picture of risk could eventually help clinicians identify women who need additional screening—and avoid unnecessary extra testing for others. The useful AI here isn’t offering reassurance from a chatbot. It is looking for information in medical images that patients have already had taken.

The history matters

In a September 10 announcement, NYU described a deep-learning model called NYU-DRP. It uses 3D mammograms taken over time, together with age and breast density, to estimate future risk. This is not a one-mammogram magic trick: the study specifically tested whether a screening history adds information beyond the latest examination alone.

The peer-reviewed study, published online August 12 in the American Journal of Roentgenology, examined 313,335 mammography examinations from 161,077 women at one health institution. The images were collected between January 2016 and August 2020. Researchers then evaluated how well their model distinguished women who subsequently developed cancer from those who did not.

Better ranking, not a crystal ball

In the independent test set, the model’s five-year risk-ranking score—called AUC—was 0.721, compared with 0.707 for the model using one 3D examination and 0.687 for Mirai, an AI model using 2D mammograms. Both improvements were statistically significant.1

That score measures how well a system separates future cancer cases from non-cases. Broadly, 0.721 means it would rank a randomly selected future case above a randomly selected non-case about 72% of the time. It does not mean it correctly diagnoses 72% of patients, or that someone given a high score has a 72% chance of cancer.

The researchers also compared their model with Tyrer-Cuzick, a widely used clinical risk calculator. In a separate matched group of 432 women, the AI scored 0.676 versus 0.563 for Tyrer-Cuzick. That larger gap belongs to this smaller comparison—not to the entire study population.1

One intriguing finding was that breast density did not tell the whole story. The model classified some examinations of extremely dense breasts as average risk, while identifying some examinations of fatty breasts as high risk. The cancer incidence observed in those groups supported looking beyond density alone.1 That is the practical appeal of personalization: people who share one risk factor do not necessarily share the same overall risk.

The next test is what happens to patients

This was a retrospective study: researchers tested predictions against existing images and later outcomes. It did not show that using the model to choose screening schedules prevents deaths, reduces unnecessary biopsies, or safely lets anyone skip a mammogram.

All imaging equipment in this study came from one manufacturer, Hologic. NYU says the team plans to test the approach at other academic health centers, with other manufacturers’ images, and prospectively—to see what happens when the tool helps guide care.2

Those are meaningful next steps, not small technicalities. A better prediction becomes a better medical tool only when it helps people.

Still, there is something hopeful here: the possibility that years of routine screening could yield more than a series of snapshots. With careful validation, they could help clinicians build a more individual picture of risk—and give each woman a better-informed conversation about what comes next.