An artificial intelligence model developed by researchers at NYU Langone Health can evaluate breast cancer risk up to five years in advance by tracking structural changes across historical 3D mammograms. The diagnostic tool assesses cumulative tissue variations across multiple annual screenings, moving beyond traditional methods that evaluate only an individual’s most recent scan.
The findings, validated in the American Journal of Roentgenology, establish that analyzing repeated imaging provides earlier detection signals than single-examination reviews. In comparative testing, the system reached a 72% accuracy rate in classifying women who developed breast cancer within five years, outperforming standard single-exam 3D models at 70% and traditional 2D mammography artificial intelligence systems at 68%.
NYU-DRP algorithm tracks tissue changes across consecutive years
Known as NYU-DRP, the deep learning model evaluates digital breast tomosynthesis, commonly referred to as 3D mammography, collected over consecutive examinations. Traditional computer vision tools examine a solitary baseline image to find current lesions, but this new approach maps biological variations and density evolution over long periods. The research team constructed the algorithm by training it on 313,531 annual 3D mammograms from 161,165 women who had no prior diagnosis of breast cancer between 2016 and 2020 at NYU Langone facilities.
“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,” stated Yanqi Xu, lead researcher of the study and postdoctoral fellow in the Department of Radiology at the NYU Grossman School of Medicine.
Working alongside study co-author Yiqiu Shen, assistant professor of radiology at the NYU Grossman School of Medicine, the investigators sought to identify sub-visual radiographic trends that escape conventional visual review. Rather than relying only on isolated dense regions, the system measures how tissue architecture shifts from one appointment to the next, helping flag anomalies before palpable masses form.
Comparison between NYU-DRP and traditional clinical models
Standard preventive oncology relies heavily on clinical risk calculators such as the Tyrer-Cuzick model. These conventional frameworks compute mathematical scores from documented family medical history, identified genetic mutations, age, reproductive milestones, and general breast density classifications. However, these tools do not inspect radiographic scans directly to extract temporal tissue variations.
In a direct comparison cohort involving 432 women, the NYU-DRP model achieved a 67% accuracy rate in forecasting five-year breast cancer outcomes. By comparison, the Tyrer-Cuzick clinical framework registered an accuracy rate of 56% within the identical patient group. The image-trained model identified high-risk patients across diverse categories, demonstrating that breast density alone does not dictate comprehensive risk levels and uncovering vulnerability among women with non-dense tissue.
Key findings from the NYU Langone breast cancer research
- Training cohort size: 313,531 longitudinal 3D mammograms recorded between 2016 and 2020.
- Patient sample: 161,165 women evaluated with no prior history of breast cancer.
- Prediction window: individual probability calculated across a five-year horizon.
- Overall model accuracy: 72% precision for the longitudinal 3D system, compared to 70% for single 3D scans and 68% for 2D exams.
- Direct comparative trial: 67% accuracy for NYU-DRP versus 56% for the Tyrer-Cuzick clinical assessment.
- Routine screening deficit: between 30% and 40% of eligible women currently do not attend recommended routine mammograms.
Medical specialists evaluate the role of predictive imaging
External breast health experts emphasized the clinical necessity of identifying patient susceptibility long before visible tumors develop. Traditional diagnostics often miss patients who lack hereditary indicators, leaving them unaware of underlying risks until regular imaging reveals advanced lesions.
“For decades, as a doctor and breast imaging specialist, I have cared for women who were shocked by a breast cancer diagnosis. Again and again, I heard: ‘How can this be? No one in my family has had breast cancer,'” said Connie Lehman, breast imaging specialist and professor of radiology at Harvard Medical School. Lehman noted that automated imaging analysis can bridge this diagnostic gap by identifying hidden indicators locked inside mammograms that have already been collected.
“A mammogram can tell us more than whether cancer is visible today — it can also provide information about a woman’s risk of developing cancer in the years ahead,” Lehman said. “A higher score can prompt a conversation with your health care provider about additional screening and risk-reduction options. A lower score does not eliminate risk or replace regular mammography.”
Dr. Elisa Port, director of breast surgery at Mount Sinai and author of The Breast Advice, noted that artificial intelligence applications can reveal hidden tumors while helping decrease false-positive callbacks. These computational systems present particular advantages for patients with dense breast tissue, while offering objective diagnostic support that could encourage greater screening attendance among the 30% to 40% of women who currently miss regular appointments.
Clinical status and pending regulatory validation
Because the technology assesses historical imaging files already stored in electronic medical record archives, applying the model does not expose women to additional ionizing radiation or require supplemental appointments. If integrated into clinical workflows, patients identified with elevated five-year scores could consult medical teams regarding advanced screening protocols, including supplemental magnetic resonance imaging or customized surveillance timelines, while average-risk individuals maintain standard testing routines.
Despite the published accuracy metrics, the NYU-DRP system remains an experimental investigative tool and does not possess regulatory clearance from the Food and Drug Administration for hospital deployment. Investigators have not yet announced a scheduled date for regulatory submission, and the algorithm must undergo extensive external validation to confirm that its accuracy holds across different hospital systems and varied imaging equipment manufacturers before entering standard preventive practice.
