undefined
A newly developed deep-learning artificial intelligence tool that examines women’s past and recent annual 3D mammograms proves more effective at forecasting a five-year breast cancer risk than traditional assessment models or single-scan AI methods. Researchers at NYU Langone Health and its Perlmutter Cancer Center created the deep-learning model, designated as NYU-DRP, by evaluating multiple years of digital breast tomosynthesis scans, commonly referred to as longitudinal DBT. The findings were published online in the American Journal of Roentgenology.
According to the study data, the NYU-DRP model correctly ranked individuals at higher risk 72 percent of the time. For comparison, single 3D mammograms and AI-assisted 2D testing accurately predicted higher-risk cases 70 percent and 68 percent of the time, respectively. Study lead investigator Yanqi Xu, PhD, a postdoctoral research fellow in the Department of Radiology at NYU Grossman School of Medicine and Perlmutter Cancer Center, noted that the research highlights how modern algorithms can reliably utilize existing 3D imaging archives to uncover subtle shifts in breast tissue over time.
To build the NYU-DRP model, investigators analyzed 313,531 yearly 3D mammograms originating from 161,165 patients without breast cancer who underwent imaging at NYU Langone facilities between 2016 and 2020. When evaluated against the widely recognized Tyrer-Cuzick lifetime risk assessment tool, the AI model demonstrated superior predictive capability. The NYU-DRP correctly identified individuals at higher risk after five years in 67 percent of cases, whereas the Tyrer-Cuzick method achieved 56 percent accuracy. Unlike imaging-based AI, the Tyrer-Cuzick assessment relies on personal and family medical history, genetic factors, and breast density.
The study also revealed surprising insights regarding breast tissue density. While dense breast tissue is traditionally linked to an elevated cancer risk, the researchers found that tissue density alone does not fully correlate with predicted outcomes. Among patients with extremely dense breasts, the NYU-DRP system classified 37.6 percent as average risk, whereas actual diagnosed cases over five years stood at 0.7 percent. Conversely, the model flagged 15.5 percent of patients with less dense, fatty breasts as high risk, matching an actual five-year incidence rate of 2.5 percent.
Study senior investigator Yiqiu Artie Shen, PhD, an assistant professor of radiology at NYU Grossman School of Medicine, emphasized that serial 3D mammograms contain predictive details regarding future breast cancer risk that cannot be captured by breast density measurements or isolated screenings alone. Co-investigator Laura Heacock, MD, an associate professor of radiology, added that pending successful validation in broader populations, such AI-guided tools could help clinicians tailor screening schedules more precisely, directing supplemental tests to those who truly need them while sparing lower-risk patients from unnecessary procedures.
Looking ahead, the research team intends to utilize the longitudinal DBT program to monitor patient breast health prospectively and cross-verify the algorithm using external data from independent academic institutions and various imaging equipment manufacturers. All imaging data utilized in the initial study were gathered exclusively from equipment produced by Hologic Inc. Funding for the research was supplied by the National Science Foundation, the National Institutes of Health, and several specialized cancer discovery and pilot project funds.
Source: News-Medical.net