Breast cancer risk assessment may be entering a new era. Researchers at NYU Grossman School of Medicine have developed an artificial intelligence model that analyzes longitudinal 3D mammography images to predict five-year breast cancer risk, achieving significantly better performance than both established clinical risk calculators and existing AI tools that examine only single imaging timepoints.

The model, which uses routinely collected digital breast tomosynthesis (DBT) examinations along with age and breast density data, achieved an area under the curve (AUC) of 0.721 for five-year risk prediction—outperforming the widely used Tyrer-Cuzick clinical model (AUC 0.563) and Mirai, a 2D mammography-based AI system (AUC 0.687). The research was published online August 12, 2026, in the American Journal of Roentgenology.

Mining Temporal Patterns in Breast Tissue

What distinguishes this approach is its ability to detect subtle tissue changes over time. By incorporating up to 10 prior DBT examinations, the AI system identifies temporal patterns that single-timepoint assessments miss. When researchers compared the longitudinal model to one using only the current examination, performance improved from an AUC of 0.707 to 0.721.

The study analyzed more than 313,000 DBT examinations from approximately 161,000 women at a single academic health system. Lead researchers Yiqiu Shen, PhD, Assistant Professor in the Department of Radiology, and Yanqi Xu, PhD, from NYU's Center for Data Science, designed the model to work with imaging data already collected during routine screening—requiring no additional tests or radiation exposure.

Reclassifying Risk Beyond Density Alone

Perhaps most clinically significant is the model's ability to refine risk assessment beyond breast density categories. Among women with extremely dense breasts—traditionally considered higher risk—approximately 40% were reclassified as average risk, with an observed five-year cancer incidence of just 0.8%. Conversely, among women with fatty breasts, roughly 15% were identified as high risk, with a 2.6% observed incidence.

This granular risk stratification could reshape screening protocols. The American College of Radiology recognizes risk-stratified screening, but existing models like Gail and Tyrer-Cuzick rely on broad clinical factors rather than detailed imaging data. The NYU model's ability to integrate temporal imaging analysis suggests potential for more personalized screening intervals and supplemental imaging decisions.

Validation Challenges Ahead

The researchers acknowledge important limitations. The study was retrospective, conducted at a single center using a single vendor's DBT equipment, and observed performance variation across racial and ethnic subgroups. No external validation has been performed, which remains a critical barrier before clinical deployment.

For healthcare AI companies, the findings create competitive pressure to incorporate temporal analysis and 3D modality compatibility. For providers, the approach could eventually enable workflow integration without disrupting existing screening protocols. For patients, validated deployment could mean fewer unnecessary interventions for lower-risk individuals while intensifying surveillance for those at genuinely elevated risk.

The shift from snapshot to longitudinal imaging analysis represents a fundamental evolution in how AI approaches breast cancer risk—one that leverages the rich temporal data already embedded in routine screening histories.

Sources

  1. NYU Study: AI Model Using Longitudinal 3D Mammography Outperforms Standard Risk Tools for 5-Year Breast Cancer Prediction — MedicalResearch.com