Low-Rank Tensor Regularized Network for Accurate Cardiac Motion Estimation in Echocardiography Videos

Accurate estimation of myocardial motion from echocardiography videos is essential for assessing cardiac function and diagnosing cardiovascular diseases. Conventional approaches typically estimate motion between consecutive frame pairs, neglecting the long-term spatiotemporal dynamics across cardiac cycles. This limitation often leads to inconsistent motion estimation. To address this issue, we propose a novel framework that jointly incorporates spatial structures and temporal dependencies to improve myocardial motion estimation. Specifically, we construct cost volumes from motion features extracted at each frame and propose a learnable spatio-temporal tensor low-rank (STLR) module to exploit the inherent low-rank properties of cardiac motion, thereby capturing dominant motion patterns and enhancing temporal continuity. In addition, a GRU-based recurrent refinement module integrates local motion cues from cost volumes with global motion context derived from the STLR module, enabling progressive refinement of the motion field. Experimental results demonstrate that our method achieves superior accuracy and consistency compared to the state-of-the-art methods, highlighting its potential for improving echocardiography motion analysis in clinical practice.