Deep Learning-Based Displacement Tracking for Post-Stroke Myofascial Shear Strain Quantification

Shear strain measurements derived from ultrasound imaging of tissue displacements have the potential to be a useful biomarker of muscle health, myofascial stiffness, and associated pain. However, displacement tracking techniques that were previously implemented to discern this potential are limited by displacement estimation noise and lengthy runtimes. This work investigates the ability of a deep learning-based displacement tracking method applied to myofascial shear strain quantification to resolve these two limitations. A pre-trained recurrent allpairs field transforms (RAFT) network was deployed on in vivo shoulder muscle data acquired from two stroke patients. Results were compared to a conventional window-based tracking approach. RAFT reduced the estimation variance by factors of 354-667 relative to the conventional approach, which manifests as smoother lateral displacement images. When each post-displaced image was warped by the estimated displacement, RAFT reduced the maximum residual between the pre- and post-displaced images by 19-23%, when compared to the conventional approach. In addition, RAFT completed displacement tracking of consecutive image pairs in 0.25 seconds, which is one of the greatest advantages of RAFT for the proposed clinical application. Results are promising to quantify myofascial stiffness and monitor muscle health with real-time speed (e.g., 4 Hz).