AI uses past 3D mammograms to predict breast cancer risk

An artificial intelligence model that analyzes changes across multiple 3D mammograms may predict a woman’s five-year breast cancer risk more accurately than models based on a single mammogram. Researchers at NYU Langone Health and Perlmutter Cancer Center say the approach could eventually support more personalized breast cancer screening.

The deep-learning model, called NYU-DRP, uses longitudinal digital breast tomosynthesis (DBT), comparing 3D mammograms taken during annual screenings over several years. The study was published in the American Journal of Roentgenology. NYU-DRP correctly ranked women at higher risk 72 percent of the time. A model analyzing only the latest 3D mammogram achieved 70 percent, while an AI model based on 2D mammograms reached 68 percent.

Changes over time

Researchers developed NYU-DRP using 313,531 annual 3D mammograms from 161,165 women without breast cancer who underwent imaging at NYU Langone hospitals between 2016 and 2020. According to the researchers, repeated mammograms contain information about how breast tissue changes over time. These longitudinal patterns appear to provide additional information about future cancer risk beyond what can be extracted from a single scan.

The researchers also compared NYU-DRP with Tyrer-Cuzick, an established risk-assessment tool based on factors including age, family and personal medical history, genetic mutations, breast density and biopsy results. For this analysis, they compared 432 women, half of whom developed breast cancer within five years and were matched with women of similar age and background who did not. NYU-DRP correctly ranked five-year risk 67 percent of the time, compared with 56 percent for Tyrer-Cuzick. Across the broader study population, which was followed until 2025, fewer than 3 percent of women developed breast cancer.

Beyond breast density

The findings also suggest that breast density alone does not fully reflect individual risk. Dense breast tissue is an established breast cancer risk factor, but NYU-DRP identified substantial variation within density categories. Among women with extremely dense breasts, the model classified 37.6 percent as having average risk. Their actual five-year cancer rate was 0.7 percent. Conversely, 15.5 percent of women with less-dense, fatty breasts were classified as high risk, with a five-year cancer rate of 2.5 percent.

Senior investigator Yiqiu “Artie” Shen said the results indicate that repeated 3D mammograms contain risk information that cannot be captured completely by breast density or a single mammogram.

Personalized screening

The researchers envision a future in which AI-assisted risk prediction could help determine which women may benefit from additional screening while reducing unnecessary supplemental examinations for those at lower risk. However, the model is not yet ready to guide screening decisions. The team plans to evaluate NYU-DRP prospectively and investigate whether its use affects health outcomes and breast cancer detection.

External validation will also be important. The researchers intend to test the model using data from other academic medical centers and different mammography systems. All imaging in the current study was performed using equipment from a single manufacturer.

The results therefore provide evidence that longitudinal 3D mammography may improve AI-based risk assessment, but further testing across more diverse populations and clinical settings is needed before the technology could support personalized screening in routine care.

AI-supported mammography

A large Swedish randomized controlled trial involving more than 100,000 women, earlier this year found that AI-supported mammography screening can improve early detection of clinically relevant breast cancers while reducing radiologists’ workload. In the MASAI trial, AI-assisted screening resulted in 12% fewer interval cancers, diagnosed between screening rounds, compared with standard double reading by radiologists.

During two years of follow-up, the AI group had 1.55 interval cancers per 1,000 women, versus 1.76 in the control group. Researchers also reported 16% fewer invasive cancers, 21% fewer large tumors and 27% fewer aggressive subtypes. Overall, 81% of cancers in the AI group were detected during screening, compared with 74% under standard screening, while false-positive rates remained similar.

Earlier analyses showed a 44% reduction in radiologists’ screen-reading workload and a 29% increase in cancer detection. Researchers stress that AI supports rather than replaces radiologists, with human assessment remaining essential.