Serial 3D Mammograms Rank Risk That Density Scores Miss

NYU-DRP ranked five-year breast cancer risk at an AUC of 0.721 by reading serial 3D mammograms, NYU Langone Health said on September 10, 2026. The deep-learning model beat a single 3D exam, a 2D AI score, and the Tyrer-Cuzick questionnaire on the same task.

Clinics already sort women into extra MRI using density letters and lifetime-risk cutoffs. The NYU model often picked a different group.

Stacked 3D Exams Rank Five-Year Risk

Researchers at NYU Langone Health and Perlmutter Cancer Center trained NYU-DRP on 313,531 yearly 3D mammograms from 161,165 women who did not have breast cancer when they were scanned at NYU Langone hospitals between 2016 and 2020. The tool reads digital breast tomosynthesis, the slice-by-slice 3D mammogram, across more than one year, plus age and breast density.

The five-year risk from longitudinal DBT paper went online in the American Journal of Roentgenology on August 12, 2026, after acceptance on August 3. On an independent test set of 34,570 exams, the serial model’s five-year AUC was 0.721 (95% CI, 0.698 to 0.744). A model that saw only the latest 3D exam reached 0.707. Mirai, a well-known 2D model run on same-day full-field digital mammograms, reached 0.687. Both gaps had p values under.001. NYU described those three AUCs as ranking higher-risk women 72 percent, 70 percent, and 68 percent of the time.

This study finds a deep learning model leveraging longitudinal DBT significantly improves 5-year breast cancer risk prediction, outperforming traditional clinical and mammography-based models. https://t.co/pigwSMYEVz

— AJR (@AJR_Radiology) August 18, 2026

FIVE-YEAR AUC ON THE NYU TESTS

Lead investigator Yanqi Xu, PhD, a postdoctoral fellow in radiology at NYU Grossman School of Medicine, said existing 3D files already hold how tissue has changed across screenings. “Our study shows how AI models like NYU-DRP can be used to reliably determine a woman’s future risk of breast cancer based on existing 3D mammograms, which hold information on how the breast tissue has changed across multiple screenings over time,” Xu said.

Dense Tissue Did Not Match Who Got Cancer

Since September 10, 2024, U.S. mammography sites have had to put a breast density notification in lay summaries sent to patients. The Food and Drug Administration’s wording tells women their tissue is dense or not dense, and that dense tissue both hides cancer on a mammogram and raises the chance of developing it. Reports to doctors still use four categories: almost entirely fatty, scattered fibroglandular density, heterogeneously dense, and extremely dense.

NYU-DRP’s five-year calls did not track those buckets. Senior investigator Yiqiu “Artie” Shen, PhD, an assistant professor of radiology, said repeated 3D exams hold risk information that density and a single mammogram do not fully capture.

DENSITY VERSUS THE MODEL’S FIVE-YEAR CALLS

• Extremely dense tissue: The model classed 37.6 percent of these women as average risk, and actual diagnoses after five years were 0.7 percent.
• Almost entirely fatty tissue: It classed 15.5 percent as high risk, and actual diagnoses after five years were 2.5 percent.
• The patient letter: Heterogeneously dense and extremely dense tissue both get the same “Your breast tissue is dense” paragraph, so the letter does not separate the group with the lowest observed rate here.

Dense tissue can still hide a mass on today’s X-ray, which is why the federal letter also says other imaging may help find cancers in some people with dense tissue. NYU-DRP is scoring who is more likely to be diagnosed over five years, not marking a mass on the current study. A “not dense” letter can sit on an image history the model treats as high risk, and that is the slice density policy does not pull toward extra tests.

The Score That Still Decides Extra MRI

Tyrer-Cuzick does not read mammograms. It uses personal and family history, genetic mutations, breast density, and biopsy results, and clinics often treat a lifetime risk of 20 percent or higher on tools in that family as a reason to add MRI. The American College of Radiology already points high-risk women, including those at or above that lifetime cutoff, toward MRI, and it recommends MRI for women with extremely dense tissue. The 2024 U.S. Preventive Services Task Force found insufficient evidence to recommend extra MRI for otherwise average-risk women on density alone.

On a matched set of 432 women, half of whom later developed breast cancer, NYU-DRP’s five-year AUC was 0.676 against 0.563 for Tyrer-Cuzick. NYU described that pair as ranking higher-risk women 67 percent versus 56 percent of the time. Follow-up for the wider study ended in 2025.

If future experiments in other women with breast cancer prove successful, then AI-assisted 3D mammograms like NYU-DRP could help physicians better tailor screening to a woman’s actual risk, by identifying those women who may benefit from additional screening while avoiding unnecessary supplemental tests for those at lower risk.

Laura Heacock, MD, associate professor of radiology, NYU Langone Health

That is the policy fight the AUC table does not show. MRI slots, contrast mammography, and screening ultrasound are finite, and payers already argue over who qualifies. If a serial 3D score holds up outside NYU, the women added to that queue would not be the same women density letters and questionnaire cutoffs name today.

Why One 3D Exam Leaves Signal Behind

A single 3D exam still ranked risk at 0.707, only 0.014 behind the serial model on the large test set. The extra lift came from feeding prior years, so the network could see how tissue changed, not only how it looked on the latest visit. Xu’s point is that those priors already sit in the archive. Annual screening from age 40, now the common U.S. advice for average-risk women, is what fills that archive.

The design also uses age and density as inputs, so NYU-DRP is not pretending density is useless. It is refusing to let density stand in for the image record. Shen put it in plainer terms: the future-risk signal in stacked 3D files is not the same as a density category or one exam.

That is a different job from AI that flags a possible cancer on the study a radiologist is reading today. Detection tools try to cut missed cancers on this visit. NYU-DRP tries to rank who is more likely to have a diagnosis in two to five years, which is the window used to argue for shorter intervals or extra modalities.

Mirai Was the 2D Comparison Model

Mirai, built by researchers at MIT and Massachusetts General Hospital on standard 2D digital mammograms, has been the yardstick for image-based risk scores. In its original multi-hospital tests it posted C-indices of 0.76 at MGH, 0.81 at Karolinska, and 0.79 at Chang Gung, figures from different countries and different film, not from NYU’s 3D stack. Later groups have run Mirai on 2D screening programs in the United Kingdom and elsewhere, with five-year AUCs often in the mid-0.60s depending on the cohort.

NYU did not retrain Mirai on tomosynthesis. It ran the 2D model on same-day full-field digital mammograms and compared that score with NYU-DRP’s serial 3D score. The 0.687 versus 0.721 gap is a same-cohort result, and it is also a reminder that most published mammography AI still lives on 2D pictures even as 3D units now sit in most U.S. rooms.

Adam Yala’s group designed Mirai to estimate risk at several time points and to stay stable across machines. NYU-DRP’s claim is narrower and more local: year-to-year 3D change at one health system, on one vendor’s files, ranked five-year diagnoses a bit better than that 2D baseline.

The Study Used Only Hologic Scanners

Every exam in the current tests came from equipment made by Hologic Inc. of Marlborough, Massachusetts. Shen said the group plans to share the tool with other academic centers and to check it on 3D systems from other makers. Until that happens, a hospital running GE HealthCare, Siemens Healthineers, or Fujifilm tomosynthesis cannot treat the 0.721 figure as its own.

The installed base is large enough that the vendor limit is not a footnote. As of September 1, 2026, the FDA’s MQSA tally listed 9,128 certified facilities and 13,575 accredited DBT units at 8,657 of those sites, alongside 14,218 accredited 2D digital units.

THE SCREENING LOAD AROUND THE MODEL

• Procedures: FDA facilities reported 45,025,781 annual mammography procedures as of September 1, 2026.
• 2026 diagnoses: The American Cancer Society estimates 321,910 new invasive breast cancer cases in women, plus 60,730 cases of DCIS, and 42,140 deaths.
• Lifetime odds: About 1 in 8 women in the U.S., or about 13 percent, will develop breast cancer, and more than 4 million survivors live in the country.
• Early-stage survival: For localized invasive disease diagnosed from 2015 to 2021, five-year relative survival is above 99 percent in SEER figures the cancer society publishes.

The median age at diagnosis is 63. Incidence has been rising about 1 percent a year, and a little faster under age 50. Death rates have fallen 44 percent through 2023 since 1989. Ranking who still needs extra tests is how a system with that volume tries not to MRI everyone with dense tissue and skip everyone without it.

Other Hospitals Have Not Repeated the Test

The paper is a single-center retrospective study. It has not been cleared as a clinical device, and NYU has not said it is in routine use. Shen’s next step is to watch how a longitudinal DBT program affects who later develops cancer, then to cross-check the weights on outside data. Funding came from National Science Foundation grant 1922658, National Institutes of Health grant R01EB036530, the Milstein Pilot Project Fund, the Shifrin-Myers Breast Cancer Discovery Fund, and the Manhasset Women’s Coalition Against Breast Cancer.

WHAT WE KNOW

• The ranking: Serial Hologic 3D exams plus age and density beat a single 3D exam, Mirai on 2D, and Tyrer-Cuzick on NYU’s test sets.
• The density split: A large share of extremely dense exams looked average-risk to the model, with a 0.7 percent five-year diagnosis rate, while a slice of fatty exams looked high-risk, with a 2.5 percent rate.
• The letters: Federal dense/not-dense wording has been required in patient summaries since September 10, 2024.

WHAT IS UNCONFIRMED

• Other vendors: No published NYU-DRP result exists yet on GE, Siemens, or Fujifilm 3D files.
• Other hospitals: External validation at other academic centers has been promised, not shown.
• Clinic effect: There is no trial yet showing that acting on the score cuts late diagnoses or spare unnecessary MRI.

Co-investigators included Jungkyu Park, PhD; Felicia Pasadyn, MA; Qi Lei, PhD; Alana A. Lewin, MD; Krzysztof J. Geras, PhD; Linda Moy, MD; and Freya R. Schnabel, MD. The public argument over density letters already ran in 2024. This model lands on a quieter question: once those letters are in the envelope, which image histories actually belong in the extra-test line.

Frequently Asked Questions

What Is NYU-DRP and What Images Does It Read?

NYU-DRP is a deep-learning risk model that estimates two- to five-year breast cancer risk from serial digital breast tomosynthesis exams, patient age, and breast density. It was built at NYU Langone Health on screening 3D mammograms stored from visits in 2016 through 2020, not on a new kind of scanner, and it outputs a ranking rather than a biopsy decision.

How Did NYU-DRP Compare With the Tyrer-Cuzick Score?

Tyrer-Cuzick is a clinical lifetime-risk calculator that never opens the image files; NYU compared five-year discrimination on 432 matched women and reported AUCs of 0.676 for NYU-DRP and 0.563 for Tyrer-Cuzick. Many high-risk clinics still use a 20 percent lifetime cutoff on Tyrer-Cuzick-style tools to justify screening MRI, which is a different timescale than the five-year image score.

Does Dense Breast Tissue Mean High Risk on NYU-DRP?

No. In the NYU tests the model labeled 37.6 percent of women with extremely dense breasts as average risk, and only 0.7 percent of that group had cancer within five years, while 15.5 percent of women with almost entirely fatty breasts were labeled high risk and 2.5 percent of that group were diagnosed. Density still matters for whether a mass is hidden on X-ray, which is a separate detection problem from this five-year ranking.

Can Clinics Run NYU-DRP on Any 3D Mammogram Machine?

Not on the evidence published so far. All study exams were acquired on Hologic systems, and the authors said they still need to test files from other 3D manufacturers that already sell tomosynthesis in the United States, including GE HealthCare, Siemens Healthineers, and Fujifilm. A site whose archive is another vendor’s format would be running an untested transfer.

How Rare Was a Later Cancer in the NYU-DRP Cohort?

Fewer than 3 percent of the women tested developed breast cancer by the time follow-up ended in 2025, which is why a five-year AUC near 0.72 is a ranking of a relatively uncommon event, not a diagnosis. The training exams themselves were taken in women who did not have breast cancer at those visits, and the independent test set used for the 0.721 figure held 34,570 examinations.

Disclaimer: This article is news reporting on a published radiology study and related screening rules. It is informational only and is not medical advice, a diagnosis, or a recommendation to start, skip, or change mammograms, MRI, ultrasound, or any other test. Risk scores, density letters, and survival figures describe groups in studies and registries, not any one reader’s odds. Talk with a qualified physician or breast-imaging specialist before making screening decisions, and treat the numbers here as the figures those sources listed on the dates cited, which can change with new data or new validation of the model.