Comparing several years of 3D mammograms predicts breast cancer risk better than a single image

Trained on 313,531 examinations of 161,165 women, the NYU-DRP model correctly classifies women at risk in 72% of cases, compared to 56% for the reference clinical tool.

Researchers from NYU Langone Health and the Perlmutter Cancer Center have built a deep learning model, NYU-DRP, which does not analyze an isolated mammogram but the series of annual tomosyntheses of the same patient. The training game involves 313,531 examinations carried out between 2016 and 2020 on 161,165 women who were cancer-free at the time of the scan.

To predict cancer risk at five years, the model correctly classifies higher-risk patients 72% of the time, compared to 70% for a single 3D mammogram approach, 68% for an AI working on 2D images, and 56% for the Tyrer-Cuzick clinical tool.

The main lesson lies in the gap between the first two digits: the repetition of the examinations contains information on the risk that neither breast density alone nor a single image captures. The work, published in the American Journal of Roentgenology, still needs to be validated in other centers and other device manufacturers before any clinical use.