A temporal attention-based model for breast cancer risk prediction from consecutive mammographic screenings

Abstract

Prediction of breast cancer risk is essential for personalized risk assessment to guide appropriate early prevention and breast screening strategies. While various deep learning-based models have been developed to predict risk based on four-view mammograms from a single screening session, they remain limited in capturing longitudinal dependencies across multiple screenings. To address this gap, we propose a novel approach that leverages temporal dependencies in consecutive mammograms for breast cancer risk prediction. Our method employs a temporal attention-based encoder to capture progressive changes in breast tissue across three time points for improving breast cancer risk assessment. In addition, we propose a multi-view mammogram feature fusion to integrate information from four standard mammographic views to enhance feature representation. Experimental analysis on a mammographic dataset collected in Singapore demonstrates the effectiveness of our model for breast cancer risk prediction. Comparative evaluations against state-of-the-art methods using the same dataset show that our model achieves superior performance, as evidenced by a higher area under the receiver operating characteristic curve (AUC).