How PDE Motion Models Boost Image Reconstruction in Dynamic CT

Dynamic inverse problems in imaging struggle with undersampled data and unrealistic motion. Neural fields provide a lightweight, smooth representation but often miss motion detail. This study shows that combining neural fields with explicit PDE-based motion regularizers (like optical flow) significantly improves 2D+time CT reconstruction. Results demonstrate that neural fields not only outperform grid-based solvers but also generalize effectively to higher resolutions, offering a powerful path forward for medical and scientific imaging.