Physics in the Loop: Integrating Biomechanics-Derived Training Data into a Neural Ordinary Differential Equation-Based Deformable Registration Framework

Wensi Wu, Yifan Wu, Analise M. Sulentic, James Gee, Alison Marie Pouch, Matthew A. Jolley

Keywords: Biomechanics knowledge, deformable image registration, heart valve dynamics

Abstract: Image registration of moving heart valves has been fundamentally challenging due to large leaflet deformations occurring over a short period of time with limited temporal image resolution. In this work, we propose integrating mechanics-derived data augmentation into deep learning-based registration frameworks to enhance the accuracy of the registration of heart valve motion. A finite element (FE) analysis is employed to generate additional physically realistic image frames of heart valves, which are then integrated with a neural ordinary differential equation-based deformable registration method to facilitate the registration process. We observe that augmenting the cardiac image sequence with FE-simulated frames better preserves the dynamic anatomy of the heart valves and outperforms traditional registration-based methods alone.

Submission Number: 9