Reconsidering Spatial Alignment for Longitudinal Breast Cancer Risk Prediction

Solveig Thrun, Stine Hansen, Zijun Sun, Nele Blum, Suaiba Amina Salahuddin, XIN WANG, Kristoffer Knutsen Wickstrøm, Elisabeth Wetzer, Robert Jenssen, Maik Stille, Michael Kampffmeyer

Keywords: Breast Cancer Risk Prediction, Longitudinal Image Registration, Mammography Alignment, Deformation Field Evaluation

Abstract: Regular mammography screening is key for early breast cancer detection, and deep learning enables personalized screening strategies. However, misalignment across time points can obscure subtle tissue changes and degrade risk prediction performance. This study provides insights into the impact of different alignment strategies, namely image-based registration, feature-level alignment, and implicit methods, on risk prediction using two large-scale mammography datasets, offering guidance for future research and methodological development. Results show that our newly proposed image-based registration model outperforms others, improving accuracy and yielding anatomically plausible deformations, underscoring the importance of precise alignment in longitudinal risk modeling.

Submission Number: 34