CardioODE-Track: Learning Bidirectional Cardiac Motion With Neural ODEs.

Abstract

Accurate and temporally consistent cardiac motion tracking from dynamic imaging sequences is essential for quantitative functional assessment and disease analysis. However, existing learning-based approaches often struggle to preserve temporal coherence over long sequences and rarely exploit the intrinsic temporal structure of cardiac motion. In this article, we propose CardioODE-Track, a unified unsupervised framework for cardiac motion tracking. The framework first employs a frame-aware U-Net to estimate initial interframe (INF) motion fields, which are subsequently refined via continuous-time deformation modeling using neural ordinary differential equations (ODEs). This formulation enables smooth and incremental motion evolution across adjacent frames. To further enforce long-range temporal coherence, we introduce a bidirectional Lagrangian motion formulation that jointly estimates forward and backward motion flows under a bidirectional consistency constraint, explicitly capturing the temporal symmetry of cardiac deformation. Extensive experiments on the ACDC, M&Ms, and CAMUS datasets demonstrate that CardioODE-Track achieves more accurate, stable, and physiologically consistent cardiac motion tracking than state-of-the-art methods across imaging modalities. Code is available at https://github.com/ZhuCY1001/CardioODE-Track.