Spatiotemporal Asymmetries of Longitudinal Screening Mammograms for Breast Cancer Risk Prediction

Purpose To develop and evaluate a new deep learning-based risk model that explicitly captures bilateral and longitudinal asymmetries on sequential mammograms for predicting breast cancer risk. Materials and Methods In this institutional review board (IRB)-approved retrospective case–control study, the images of sequential mammographic examinations (at least two per patient with interexamination intervals of 12–36 months)—all of which were acquired on Hologic systems—were extracted from the CSAW-CC dataset (406 cancer patients, 6,053 controls) and an independent dataset (293 cancer patients, 297 controls). A novel model, STA-Risk, whose architecture incorporates side encoding, temporal encoding, and customized asymmetry loss, was constructed using these data. Fivefold cross-validation with the individual datasets and joint training with a mixed dataset were performed. Model performance was reported with the concordance index (C-index) and the time-dependent area under the curve (AUC) (1–5 years). Results STA-Risk achieved C-indexes of 0.72 with the CSAW-CC data and 0.73 with the independent cohort data and outperformed all the compared risk models (0.67–0.70 and 0.66–0.72, respectively), with consistently higher AUCs in the 1-to 5-year risk predictions. Ablation studies revealed that all three key components of STA-Risk contributed to this improved performance. Domain shifts were observed, but their effects were mitigated with joint training strategy; the resulting model achieved cross-cohort test C-indexes of 0.75 with the CSAW-CC dataset and 0.67 with the independent dataset. Conclusion The STA-Risk deep learning risk model constructed from spatiotemporal asymmetries detected on longitudinal mammograms improves breast cancer risk prediction over existing models.