SideAge-RiskNet: Predicted Mammographic-Age Features for Future Breast Cancer Risk and Laterality Prediction

Abstract

Whether breast-specific mammographic aging provides information beyond chronological age for future cancer prediction remains unclear. We propose SideAge-RiskNet, a framework integrating predicted mammographic-age features with breast-specific representations for multi-year breast cancer risk and conditional laterality prediction. A side-aware DenseNet121 trained on never-cancer reference examinations estimates left and right breast ages. Age-adjusted residuals and bilateral age descriptors are subsequently fused with corresponding image representations, while clinical variables contribute through an image--clinical risk branch. The framework jointly predicts cumulative cancer risk over 1--5 years and the subsequently affected breast among unilateral cancer cases. Across 4,480 EMBED and 248 external examinations, pre-cancer scans demonstrated pronounced bilateral age asymmetry. SideAge-RiskNet established state-of-the-art performance in 1--5-year risk assessment over Mirai and AsymMirai with mean risk AUROCs of 81.73% on EMBED and 72.83% on CSMC. Furthermore, it enabled conditional cancer laterality prediction, achieving AUROCs of 73.26% and 82.53%, respectively. This demonstrates the critical utility of bilateral aging phenotypes in localized breast cancer risk surveillance.