Predicting 5-Year Breast Cancer Risk from Longitudinal Digital Breast Tomosynthesis: A Single-center Retrospective Study

Authors

Abstract

Key ResultsIn an independent test set, a longitudinal DBT-based deep learning model achieved higher 5-year risk discrimination than the FFDM-based Mirai model (AUC, 0.720 vs 0.687; p < 0.001) and, in a matched case-control cohort, higher discrimination than the Tyrer-Cuzick model (AUC, 0.676 vs 0.567; p < 0.001). ImportanceLongitudinal DBT-based AI improves individualized breast cancer risk assessment beyond FFDM-based AI and clinical risk models. BackgroundImaging-based breast cancer risk prediction models primarily use full-field digital mammography (FFDM). As digital breast tomosynthesis (DBT) has become a predominant screening modality in the United States, its potential for long-term breast cancer risk prediction remains under-explored. ObjectiveTo develop and evaluate a deep learning model that uses longitudinal DBT exams to predict long-term breast cancer risk. MethodsThis retrospective study included 313,531 DBT exams from 161,165 women (mean age, 58.5 {+/-} 11.7 years) between January 2016 and August 2020 at Institute A. A risk prediction (DRP) model was developed to estimate 2-5 year breast cancer risk using longitudinal DBT exams, patient age and breast density. Model performance was compared with a single-time point DBT model, the Mirai model using same-day FFDM, and the Tyrer-Cuzick model using the area under the receiver operating characteristic curve (AUC), time-dependent concordance index, and integrated Brier score. ResultsIn an independent test set (n = 34,580), the longitudinal DRP model achieved a 5-year AUC of 0.720 (95% CI, 0.703-0.738), improving on the single time point DRP model (AUC, 0.706; 95% CI, 0.687-0.724; p < 0.001) and the Mirai model (AUC, 0.687; 95% CI, 0.668-0.705; p < 0.001). In a matched case-control cohort (n=432), the DRP model achieved a 5-year AUC of 0.676 (95% CI, 0.626-0.727), compared with 0.567 (95% CI, 0.514-0.621; p < 0.001) for the Tyrer-Cuzick model. The model reclassified 37.6% (705/1,877) of women with extremely dense breasts as average risk, with a 5-year cancer incidence of 0.7% (5/705), and identified 15.5% (404/2,605) of women with fatty breasts as high risk, with a 5-year cancer incidence of 2.5% (10/404). ConclusionA deep learning model using longitudinal DBT examinations improved long-term breast cancer risk prediction compared with FFDM-based and clinical risk models. Clinical ImpactsLongitudinal DBT-based risk prediction may enable dynamic risk assessment using screening images, supporting personalized screening strategies and more targeted use of supplemental imaging.

Tags

Topics