Unsupervised 4D Cardiac Motion Tracking with Spatiotemporal Optical Flow Networks

TL;DR

This work tackles the challenge of unsupervised 4D cardiac motion tracking from noisy echocardiography by introducing a 3D optical flow network that ingests a reference frame and a sequence to estimate motion fields without ground-truth annotations. It enforces spatiotemporal coherence through a loss framework combining $L_{Temporal\_Consistency}$ and $L_{rec}$, expressed as $L_{ST} = \gamma L_{Temporal\_Consistency} + \beta L_{rec}$, and leverages forward/backward interpolation cycles to encourage temporal consistency. The model extends TV-L1 principles to 3D, uses a compact 9-layer CNN plus a 3D warp, and demonstrates competitive accuracy and speed (about 0.3s per frame) on synthetic STRAUS data, outperforming several baselines while avoiding annotated motion fields. The results establish a first end-to-end unsupervised deep-learning solution for 4D cardiac motion tracking and point toward adaptation to real data with future evaluation metrics for practical clinical impact.

Abstract

Cardiac motion tracking from echocardiography can be used to estimate and quantify myocardial motion within a cardiac cycle. It is a cost-efficient and effective approach for assessing myocardial function. However, ultrasound imaging has the inherent characteristics of spatially low resolution and temporally random noise, which leads to difficulties in obtaining reliable annotation. Thus it is difficult to perform supervised learning for motion tracking. In addition, there is no end-to-end unsupervised method currently in the literature. This paper presents a motion tracking method where unsupervised optical flow networks are designed with spatial reconstruction loss and temporal-consistency loss. Our proposed loss functions make use of the pair-wise and temporal correlation to estimate cardiac motion from noisy background. Experiments using a synthetic 4D echocardiography dataset has shown the effectiveness of our approach, and its superiority over existing methods on both accuracy and running speed. To the best of our knowledge, this is the first work performed that uses unsupervised end-to-end deep learning optical flow network for 4D cardiac motion tracking.

Figures (6)

• Figure 1: Overview of our Cardiac Motion Tracking with Spatialtemporal Networks framework: The optical flow network takes the reference frame, and an image sequence as input and outputs reconstructed sequence. Three loss functions are illustrated with the orange arrows and equations. Notice that, the 3D sequence contains both forward time sequence and reverse time sequence will be discussed in section \ref{['section4']}.
• Figure 2: The Temporal-Consistency loss for motion estimation.
• Figure 3: Results of MSE and EPE on three testing dataset
• Figure 4: MSE for related motion tracking methods and our result.
• Figure 5: LADDIST: Illustrated frame is the 30th of 34 frames. Heterogeneous motion caused by ischemia is captured.