AI tool uses past scans to predict breast cancer risk
A new deep-learning model developed at NYU Langone Health suggests that analyzing a woman's history of 3D mammograms offers a more accurate five-year risk prediction than relying on single scans or traditional genetic assessments.
An artificial intelligence system trained on years of 3D mammography data has demonstrated superior ability in forecasting breast cancer risk over the next five years. The tool, developed by researchers at NYU Langone Health and its Perlmutter Cancer Center, outperforms both single-image AI models and standard clinical risk assessments that rely on family history and genetics.
According to a study published in the American Journal of Roentgenology, the model correctly ranked higher-risk patients 72% of the time. This compares to 70% accuracy for single 3D mammogram models and 56% for the widely used Tyrer-Cuzick risk assessment tool, which does not utilize imaging data. The research highlights a shift toward using longitudinal imaging data to personalize screening protocols.
Longitudinal data improves prediction accuracy
The model, named NYU-DRP, was built using 313,531 annual 3D mammograms from over 161,000 women who did not have breast cancer during the study period from 2016 to 2020. By analyzing how breast tissue changes across multiple screenings, the system captures subtle structural shifts that single snapshots might miss. This approach allows the AI to identify patterns in tissue evolution that are predictive of future malignancy.
Researchers noted that the tool’s advantage lies in its use of digital breast tomosynthesis, or 3D mammography, over standard 2D imaging. The ability to view tissue in three dimensions over time provides a richer dataset for the deep-learning algorithms. This results in a more nuanced risk profile that accounts for individual biological changes rather than just static population averages.
Challenging assumptions about breast density
A key finding challenges the traditional reliance on breast density as a primary risk indicator. While dense tissue is generally associated with higher cancer risk, the study found that the AI model did not strictly follow this rule. In women with extremely dense breasts, the model classified 37.6% as having average risk, and actual cancer cases in this group were only 0.7% over five years.
Conversely, the model identified 15.5% of women with less-dense, fatty breasts as high risk, with actual cases at 2.5%. This discrepancy suggests that density alone is an incomplete metric. The AI appears to detect other structural factors within the tissue that contribute to risk, offering a more personalized assessment than density-based guidelines alone.
Potential for tailored screening strategies
If validated in broader clinical settings, this technology could help physicians tailor screening to individual needs. By accurately identifying women who may benefit from additional testing, the system could reduce unnecessary supplemental procedures for those at lower risk. This precision aims to optimize resource allocation and reduce patient anxiety associated with over-screening.
The study, reported by GN technics/ai (en-US), emphasizes that while the results are promising, further experiments are required. The team plans to test the model on additional datasets to ensure its reliability across diverse populations. Until then, the tool remains a research prototype rather than a standard clinical diagnostic device.
