A New Breast Cancer Imaging Study Found That Looking at Several Years of 3D Mammograms Can Predict Future Risk Better Than Using Only the Latest Scan—What Researchers Hope to Improve

For millions of women, a mammogram is an annual or biennial appointment that comes with a familiar question: Does anything look concerning right now?

But researchers are exploring whether those images could answer another important question.

What might a woman’s mammograms reveal about her risk of developing breast cancer several years in the future?

A new study from NYU Langone Health suggests that artificial intelligence may be able to identify patterns in breast tissue that become more meaningful when several years of mammograms are examined together.

Instead of analyzing only the most recent scan, researchers developed a computer model that compares information from multiple three-dimensional mammograms taken over time.

The results suggest that the approach can predict a woman’s five-year breast cancer risk more accurately than models relying on a single 3D mammogram or conventional two-dimensional images.

The study, published August 12, 2026, in the American Journal of Roentgenology, analyzed hundreds of thousands of mammograms from more than 161,000 women.

Researchers hope the technology could eventually help doctors identify women who might benefit from additional screening while allowing others to avoid unnecessary testing.

The findings are promising, but the technology remains a research development rather than a replacement for established breast cancer screening recommendations.

And importantly, predicting elevated risk is not the same as diagnosing cancer.

Researchers Analyzed More Than 313,000 Mammograms

The research team at NYU Langone Health and its Perlmutter Cancer Center developed a deep-learning model called NYU-DRP.

The system was designed to estimate a woman’s likelihood of developing breast cancer over the following five years.

To build the model, researchers used more than 313,000 three-dimensional mammography examinations from approximately 161,000 women.

The scans were performed at NYU Langone hospitals between 2016 and 2020.

The researchers focused on women who did not have breast cancer at the time of their screening examinations.

They then evaluated whether the imaging information could help distinguish women who subsequently developed breast cancer from those who did not.

The model used a method known as longitudinal digital breast tomosynthesis.

That technical description refers to examining three-dimensional mammograms collected at different points in time.

Rather than treating each examination as an isolated event, the system looks for patterns across multiple screenings.

According to NYU Langone’s September 10 announcement, the researchers found that combining current and previous mammograms improved the model’s ability to predict future risk.

The work suggests that information already collected during routine breast screening could potentially serve another purpose.

What Is a 3D Mammogram?

Three-dimensional mammography is also called digital breast tomosynthesis, or DBT.

It is a breast imaging technique that takes multiple X-ray images from different angles.

A computer reconstructs those images into thin views of breast tissue, allowing radiologists to examine overlapping structures more clearly.

Conventional two-dimensional mammography produces flat images of the breast.

Both methods are used to look for abnormalities that could indicate cancer.

However, 3D mammography can make it easier to evaluate certain findings because overlapping tissue may be less likely to obscure an area of concern.

Many imaging centers now offer digital breast tomosynthesis as part of routine screening.

The NYU study explored a different use for these images.

Instead of focusing only on whether a current scan showed a suspicious abnormality, researchers investigated whether patterns within the images could help estimate future cancer risk.

That distinction is central to understanding the findings.

The technology is intended to support risk assessment, not independently determine whether a patient currently has breast cancer.

Why Previous Mammograms May Contain Valuable Information

Doctors already compare current mammograms with older images when those records are available.

A radiologist may look for changes in breast tissue, new masses or developing abnormalities.

But artificial intelligence may be able to evaluate imaging patterns that are difficult to quantify through visual inspection alone.

The NYU researchers examined whether those patterns become more informative when several years of images are analyzed together.

Breast tissue can change over time.

Age, hormonal changes, menopause and other factors may affect breast density and tissue appearance.

Some changes are normal.

Others may be associated with changes in breast cancer risk.

By examining a sequence of mammograms, the computer model can evaluate how tissue patterns evolve rather than relying exclusively on a single image.

The researchers believe those changes may contain information that helps distinguish women at higher risk from women at lower risk.

That possibility could make past mammograms more valuable than previously recognized.

The Model Performed Better Than Single-Scan Approaches

The study compared several methods of estimating five-year breast cancer risk.

The strongest results came from the model that analyzed multiple years of 3D mammograms.

In independent testing, the researchers reported the following approximate results:

Risk prediction approach	Area under the curve
Multiple years of 3D mammograms	0.72
Most recent 3D mammogram alone	0.71
AI model using 2D mammograms	0.69

The measure used is called the area under the receiver operating characteristic curve, commonly abbreviated AUC.

It describes how well a model distinguishes between people who develop a condition and those who do not.

A score of 0.5 is equivalent to random discrimination, while a score of 1.0 represents perfect discrimination.

The longitudinal 3D model achieved an AUC of approximately 0.72.

That was better than the single-scan and 2D models, although the difference between the two 3D approaches was modest.

The researchers reported that the improvements were statistically significant.

These results are encouraging, but they do not mean that the model can predict with certainty which individual woman will develop breast cancer.

What Does the 72% Figure Actually Mean?

The study’s results have sometimes been described as correctly ranking women at higher risk approximately 72% of the time.

That wording requires explanation.

It does not mean that the system can identify 72 out of every 100 future breast cancer cases.

It also does not mean that a woman receiving a high-risk result has a 72% chance of developing cancer.

Instead, the AUC statistic measures how well the model separates future cancer cases from noncases across different risk thresholds.

One way to understand it is to imagine selecting two women from the study: one who later developed breast cancer and one who did not.

An AUC of 0.72 means the model would assign the higher risk score to the woman who developed cancer approximately 72% of the time.

That represents meaningful discrimination, but it is far from perfect.

Some women who later develop cancer will receive relatively low scores.

Some women who never develop cancer will receive relatively high scores.

That is why risk prediction models are generally intended to support medical decisions rather than make them independently.

Breast Density Was an Important Part of the Research

Breast density refers to the proportion of fibrous and glandular tissue compared with fatty tissue in the breast.

Dense breast tissue is common.

It is also relevant to breast cancer screening for two reasons.

First, women with dense breasts have a higher breast cancer risk than women with less dense breast tissue, although density is only one factor among many.

Second, dense tissue can make cancer harder to detect on mammograms because both dense tissue and many cancers appear white on the images.

The NYU researchers investigated whether their model could identify risk differences that breast density alone might miss.

Their findings suggest that it may.

Some women with extremely dense breasts were classified as having relatively low or average risk.

Meanwhile, some women with predominantly fatty breasts were classified as having elevated risk.

That suggests breast density does not tell the entire story.

Nearly 40% of Women With Extremely Dense Breasts Were Classified as Average Risk

One notable finding involved women whose mammograms showed extremely dense breast tissue.

The researchers examined a subgroup of 1,877 women with extremely dense breasts.

The model classified approximately 39.7% of those women as average risk.

Among that group, the observed five-year breast cancer incidence was approximately 0.8%.

The finding suggests that not all women with extremely dense breasts have the same level of risk.

Although breast density remains important, other imaging characteristics may provide additional information.

This could eventually help clinicians make more individualized decisions about supplemental screening.

However, the finding does not mean that women with dense breasts should stop receiving recommended mammograms or decline additional testing solely because an experimental AI model assigns them a lower score.

Clinical decisions must consider the patient’s overall risk profile, available evidence and current medical guidelines.

Some Women With Fatty Breasts Were Classified as High Risk

The researchers also examined women with predominantly fatty breast tissue.

This group is generally considered to have lower risk associated with breast density.

But the AI model identified a subset whose imaging patterns suggested higher risk.

Among 2,605 examinations involving women with fatty breasts, approximately 14.8% were classified as high risk.

The observed five-year breast cancer incidence in that group was approximately 2.6%.

This finding is significant because it suggests that lower breast density does not automatically mean a woman has low overall risk.

Other imaging features may carry information that is not captured by a simple density classification.

For researchers, the results support the possibility of moving beyond broad categories toward more individualized risk estimates.

For patients, the important lesson is that breast density is only one part of breast cancer risk assessment.

How Does AI Identify Patterns That May Be Difficult to See?

Artificial intelligence systems used in medical imaging are trained to identify statistical patterns in large collections of images.

Deep-learning models can analyze combinations of visual features that may not be obvious to human observers.

In breast imaging, those features may include tissue texture, density patterns and structural relationships within the images.

The NYU model evaluates information from multiple examinations rather than relying solely on the latest scan.

Researchers believe this allows the system to capture changes that may be associated with future risk.

But that does not mean the computer can identify a future cancer that already exists in every patient.

The model is detecting patterns associated with statistical risk.

It is not predicting the exact location, timing or biological characteristics of a cancer that might develop.

The distinction matters because AI results can sound more certain than they actually are.

A model may help estimate risk without providing a definitive answer about an individual’s future health.

Could This Change How Often Women Receive Mammograms?

One long-term goal is to make breast cancer screening more personalized.

Current screening recommendations rely heavily on age and established risk factors.

Some women also receive additional imaging because of family history, genetic mutations, prior breast conditions or other circumstances.

Researchers hope improved risk prediction could eventually help determine which women might benefit from more intensive surveillance.

For example, a woman identified as having substantially elevated risk might be considered for supplemental imaging or closer monitoring, depending on her complete clinical circumstances.

Conversely, better risk assessment might help reduce unnecessary additional testing among women whose overall risk is lower.

But these possibilities require more evidence.

The NYU study did not establish that changing screening schedules based on this model improves health outcomes.

It also did not demonstrate that women classified as lower risk can safely skip recommended mammograms.

For now, the technology represents a potential tool for future screening strategies rather than a reason for patients to change their existing schedules.

Why Earlier Identification of Risk Matters

Breast cancer outcomes are generally better when the disease is detected before it has spread to distant organs.

That is one reason routine screening remains important.

But screening and risk prediction serve different purposes.

Screening attempts to detect cancer that may already be present.

Risk prediction estimates the likelihood that cancer may develop over a future period.

If doctors can identify elevated risk more accurately, they may be able to offer appropriate patients additional surveillance, counseling or preventive options.

For some women, that could include discussing supplemental breast MRI.

For others, it could involve a more detailed assessment of family history or genetic risk.

The ultimate goal is not simply to produce a better computer score.

It is to determine whether more accurate risk estimates can help doctors make decisions that improve patient outcomes.

That is the next challenge for researchers.

A Separate 2026 Study Also Found That Breast Cancer Risk Scores Change Over Time

The NYU research is part of a broader effort to understand what repeated mammograms can reveal.

Another study published in Radiology in 2026 examined how AI-generated breast cancer risk scores changed over several years.

Researchers analyzed 158,807 mammograms from 54,014 women.

They found that women who later developed breast cancer often showed increasing AI-generated risk scores in the years before diagnosis.

Women who did not develop cancer generally had more stable scores.

According to the Radiological Society of North America, differences in risk patterns could be detected several years before diagnosis in the study population.

The research used conventional 2D mammograms and evaluated changes in risk scores over time.

It was a separate investigation from NYU’s longitudinal 3D mammography study.

Together, the findings support growing scientific interest in evaluating breast cancer risk as something that may change over time rather than remain fixed.

However, neither study proves that changing patient care based on these AI scores will reduce breast cancer deaths.

That remains an important question for future research.

Why Family History Doesn’t Tell the Whole Story

Many women think of breast cancer risk primarily in terms of family history.

Having a mother, sister or daughter with breast cancer can increase risk.

Inherited genetic variants, including certain BRCA1 and BRCA2 mutations, can also substantially increase risk.

But many women diagnosed with breast cancer have no known strong family history or inherited mutation.

That means relying only on family history would miss many women who eventually develop the disease.

Traditional risk assessment tools consider several factors, which may include age, reproductive history, breast density, family history and previous breast biopsy results.

AI-based imaging tools could potentially add another layer of information.

The NYU model suggests that the mammograms themselves may contain useful clues about future risk.

The researchers hope those clues can eventually complement established clinical risk assessment rather than replace it.

What Are the Limitations of the Study?

Although the results are promising, several limitations are important.

First, the study was retrospective.

Researchers analyzed previously collected mammograms and outcomes rather than prospectively assigning patients to different screening strategies.

Second, the images came from one healthcare system.

Even though the dataset was large, results from one institution may not fully represent patients and imaging practices across the country.

Third, the improvement over the single 3D mammogram model was relatively small.

The longitudinal approach achieved an AUC of approximately 0.72 compared with approximately 0.71 for the single-scan model.

That difference may be statistically significant, but researchers still need to determine how much it would improve real-world medical decisions.

Fourth, a risk score is only useful if clinicians know how to act on it appropriately.

A model that identifies more women as high risk could lead to additional imaging, expense, anxiety or false-positive findings.

Researchers must evaluate those potential consequences alongside any benefits.

Finally, the study did not show that using NYU-DRP in routine care reduces advanced breast cancer diagnoses or mortality.

Those outcomes would require additional clinical research.

Will Women Need Extra Mammograms for This Technology?

One appealing aspect of the research is that the model uses mammograms that have already been collected.

In principle, an AI system could analyze current and previous screening images without requiring an entirely new imaging appointment solely for risk prediction.

That could make the approach easier to integrate into existing breast imaging workflows.

However, practical challenges remain.

Previous mammograms may be stored at different facilities.

Images may have been produced using different equipment or imaging protocols.

Some women may have incomplete screening histories.

Healthcare systems would also need to address software validation, clinical oversight, patient privacy and appropriate use of risk estimates.

The model’s performance could vary when applied to patients whose imaging histories differ from those in the research dataset.

Those issues would need to be addressed before widespread clinical adoption.

What Should Women Do With Their Mammogram History Now?

The findings reinforce the value of keeping previous mammograms accessible.

Radiologists already use prior images to identify changes and compare suspicious findings.

Women who change imaging centers may benefit from making sure their new provider can obtain earlier mammograms.

That doesn’t mean patients need to collect and interpret the images themselves.

The important step is helping the medical team access relevant records when needed.

Women can also ask their healthcare providers whether their breast density, family history or other factors suggest a need for individualized risk assessment.

Those conversations can be especially useful when screening reports identify dense breast tissue or when a patient has additional risk factors.

The NYU model is not a standard clinical test that every woman can request today.

But the study highlights why past imaging records may remain useful well beyond the date of the original examination.

What Are the Current Breast Cancer Screening Recommendations?

For women at average risk, the U.S. Preventive Services Task Force recommends screening mammography every two years from ages 40 through 74.

The American Cancer Society recommends that women ages 45 through 54 receive annual mammograms, with options for women ages 40 through 44 to begin annual screening.

Women age 55 and older may transition to screening every two years or continue annually, depending on their preferences and health.

Women at higher risk may need different screening strategies, including MRI.

Because professional recommendations differ, screening decisions should be made in consultation with a healthcare professional.

The new AI research does not replace these recommendations.

Women should continue following appropriate screening guidance rather than waiting for emerging technology to become available.

Why This Research Matters Beyond Artificial Intelligence

Artificial intelligence is often discussed as a way to make medical technology faster or more automated.

But the central idea behind this study is more specific.

Researchers are investigating whether information collected during ordinary healthcare appointments can reveal changes that become meaningful over time.

A mammogram may contain more than a snapshot of breast tissue on one particular day.

When combined with previous examinations, it may help reveal patterns associated with future cancer risk.

That possibility could eventually support more individualized screening.

But translating the research into clinical practice will require careful validation.

Doctors need to know not only whether a model produces a statistically better score, but whether using that score helps patients.

The most important measure of success will be improved health outcomes, not simply improved algorithm performance.

The Next Step Is Proving That Better Predictions Lead to Better Care

The NYU study found that examining multiple years of 3D mammograms improved five-year breast cancer risk prediction compared with using only the latest scan.

The researchers evaluated more than 313,000 examinations from approximately 161,000 women and found that their longitudinal model performed better than the comparison approaches.

The results suggest that previous mammograms may contain useful information about how breast cancer risk changes over time.

But the study does not mean artificial intelligence can predict exactly who will develop breast cancer.

It does not eliminate the need for routine mammography.

And it does not establish that women should receive fewer or more scans based solely on an experimental model.

Instead, it points toward a future in which screening recommendations may become more closely tailored to each woman’s individual risk.

For patients, the immediate takeaway is straightforward.

Continue recommended breast cancer screening, make previous imaging available to your healthcare team and discuss any personal risk factors with a medical professional.

For researchers, the challenge is to determine whether these more sophisticated predictions can eventually lead to earlier diagnoses, fewer unnecessary procedures and better outcomes.

The promise isn’t simply that computers may become better at reading mammograms.

It’s that years of imaging history could help doctors make better decisions about a woman’s future breast health.

Original research and medical sources