TOPLINE

A deep learning model using longitudinal digital breast tomosynthesis (DBT) examinations improved 5-year breast cancer risk prediction compared with full-field digital mammography (FFDM)-based AI and clinical risk models. Among women with extremely dense breasts, the model identified an average-risk subset with low observed cancer incidence, whereas among women with fatty breasts, it identified a high-risk subset with elevated cancer incidence.

METHODOLOGY

• Standard breast cancer risk assessment tools rely mainly on personal and family history and typically predict risk with only modest accuracy. DBT has largely replaced older FFDM for cancer screening in the US, but prior studies have mainly focused on cancer detection, short-term risk estimation, or synthetic 2D projections generated from DBT rather than native 3D volumes. Whether longitudinal DBT can improve long-term breast cancer risk prediction has remained largely unexplored.
• To address this, researchers developed a DBT-based risk prediction (DRP) model using a retrospective cohort of 313,335 DBT examinations from 161,077 women (mean age, 58.5 years), comprising 267,319 training, 11,446 validation, and 34,570 hold-out test examinations. The model integrated longitudinal DBT examinations (up to 10 per patient), patient age, and breast density to estimate 2- to 5-year risk for breast cancer.
• The model's performance was evaluated in an independent test set of 34,570 examinations using area under the curve (AUC), time-dependent concordance index (C-index), and integrated Brier score. It was compared with a single-timepoint DBT model, the Mirai model (using same-day FFDM), and the Tyrer-Cuzick model.
• To compare the new model with the standard Tyrer-Cuzick model, a matched case-control cohort of 432 examinations (216 cancer-positive, 216 cancer-negative) was used. Control individuals were matched 1:1 by age (within 3 years) and exact Breast Imaging Reporting and Data System (BI-RADS) density category.
• Examinations were classified into three risk groups based on predicted 5-year risk: average risk (less than 1.7%), increased risk (between 1.7% and 2.999%), and high risk (3.0% or higher).

TAKEAWAY

• In the independent test set, the longitudinal DRP model achieved a 5-year AUC of 0.721 (95% CI, 0.698-0.744) and a C-index of 0.712 (95% CI, 0.688-0.726), surpassing the single-timepoint DRP model, which had an AUC of 0.707 (95% CI, 0.683-0.730; P < .001) and a C-index of 0.704 (95% CI, 0.682-0.726), as well as the Mirai model, which had an AUC of 0.687 (95% CI, 0.663-0.710) and a C-index of 0.684 (95% CI, 0.661-0.707; P < .001 for both).
• In the matched case-control cohort, the 5-year AUC was 0.676 (95% CI, 0.626-0.726) for the DRP model, compared with 0.563 (95% CI, 0.509-0.619; P < .001) for the Tyrer-Cuzick model.
• In subgroup analyses, 5-year AUCs varied by age, ranging from 0.678 (95% CI, 0.642-0.714) among women in their sixties to 0.749 (95% CI, 0.691-0.803) among women in their forties. Predicted risk increased with age: Women aged 40-49 years had an average predicted 5-year risk of 1.58%, compared with 3.31% for women aged 70 or older (P < .001).
• Among examinations of women with extremely dense breasts (BI-RADS D; n = 1877), the model classified 39.7% as average risk, with an observed 5-year cancer incidence of 0.8% (6/746). Among examinations of women with fatty breasts (BI-RADS A; n = 2605), the model classified 14.8% as high risk, with an observed 5-year cancer incidence of 2.6% (10/386). Across all density groups, examinations classified as having average risk showed the lowest 5-year cancer incidence (0.6%-0.8%).

IN PRACTICE

"[T]his study suggests that longitudinal DBT examinations improve individualized long-term breast cancer risk prediction," the study authors wrote. "These findings indicate that longitudinal DBT provides complementary volumetric and temporal information beyond existing clinical and FFDM-based approaches," they further added.

SOURCE

The study, led by Yanqi Xu, MSc, Center for Data Science, New York University, New York City, was published online in the American Journal of Roentgenology.

LIMITATIONS

The model was developed and validated at a single academic health system using DBT examinations from one vendor, which may have limited generalizability to other populations and DBT systems. The model did not include genetic, reproductive, or surgical history information, and the comparison with the Tyrer-Cuzick model involved a small case-control cohort.

DISCLOSURES

The study was supported partly by grants from the National Science Foundation, the National Institutes of Health, the Milstein Pilot Project Fund, the Shifrin-Myers Breast Cancer Discovery Fund, and the Manhasset Women's Coalition Against Breast Cancer. The authors reported no conflicts of interest.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.