Abstract
Screening digital breast tomosynthesis (DBT) aims to identify breast cancer early when treatment is most effective, leading to reduced mortality. In addition to early detection, the information contained within DBT images may also inform subsequent risk stratification and guide risk-reducing management. Using transfer learning, we refined a model in the Joanne Knight Breast Health Cohort at Washington University, a cohort of 5,066 women with DBT screening (mean age, 54.6), among whom 105 were diagnosed with breast cancer (26 ductal carcinoma in situ). We applied the model to external data from the Emory Breast Imaging Dataset, a cohort of 7,017 women free from cancer (mean age, 55.4), among whom 111 pathology-confirmed breast cancer cases were diagnosed more than 6 months after initial DBT (17 ductal carcinoma in situ). We obtained a 5-year AUC of 0.75 [95% confidence interval (CI), 0.73-0.78] in the internal validation. The model validated in external data gave an AUC of 0.72 (95% CI, 0.69-0.75). The AUC was unchanged when age and Breast Imaging-Reporting and Data System density were added to the model with synthetic DBT images. The model significantly outperforms the Tyrer-Cuzick model, with a 5-year AUC of 0.56 (95% CI, 0.54-0.58; P < 0.01). Our model extends risk prediction applications to synthetic DBT, provides 5-year risk estimates, and is readily calibrated to national risk strata for clinical translation and guideline-driven risk management. The model could be implemented within any digital mammography program. Prevention Relevance: We develop and externally validate a 5-year risk prediction model for breast cancer using synthetic DBT and demonstrate clinical utility by calibrating to the national risk strata as defined in breast cancer risk management guidelines.