VI Seminar #87: Deep Learning Based Breast Cancer Risk Prediction from Longitudinal Mammograms

Presented by Solveig Thrun, Doctoral Research Fellow at SFI Visual Intelligence

Abstract

Regular mammography screening enables early breast cancer detection, and deep learning–based risk models offer opportunities to personalize screening intervals for high-risk individuals. Recent approaches increasingly leverage longitudinal mammograms, but accurate spatial alignment across time points remains challenging. Misalignment can obscure meaningful tissue changes and degrade predictive performance.

In this study, we systematically evaluate alignment strategies for longitudinal risk modeling, including image-based registration, feature-level alignment with and without regularization, and implicit alignment methods.Using two large-scale mammography datasets, we compare these approaches across predictive performance metrics and deformation field quality. Image-based registration consistently outperforms feature-based and implicit methods, yielding more accurate and temporally consistent predictions while producing smooth, anatomically plausible deformation fields. Although regularization improves deformation quality for feature-level alignment, it negatively impacts risk prediction performance. Applying image-based deformation fields within the feature space achieves the best overall results.

These findings highlight the importance of image-based deformation fields for spatial alignment in longitudinal breast cancer risk modeling, supporting more accurate and robust personalized screening strategies.