Advancements in Shear Wave Elastography with Neural Networks and Multi-Resolution Approaches

Accurate measurement of force-induced tissue motion is widely studied in ultrasound elastography. Especially, in shear wave elastography (SWE), improving the quality of shear wave speed (SWS) estimation by optimized utilization of the detected motion profiles is important. Generally, traditional time delay estimators based on cross-correlation and phase estimation are used for motion estimation and the obtained displacement-time profiles at two tracking locations separated by a fixed distance are used to compute the SWS at a given location. Recently, Convolutional Neural Networks (CNNs) have attracted the attention of researchers and the architectures of neural networks have been modified to enable the network to extract high-frequency information from RF data. MPWC-Net++ was one of the modified networks based on IRRPWC-Net that demonstrated excellent performance on quasi-static elastography solely trained on computer vision images without any training on ultrasound data. Here, we demonstrate the feasibility of adapting these networks to measure particle motion in SWE where the displacement is substantially lower than the quasi-static elastography. When estimating SWS based on the time delay using a pair of spatially lagged motion profiles, the robustness to time-of-flight errors is traded with the spatial resolution. Previously, model-based multi-resolution approaches were applied in simulated elastography data to obtain a noise-robust output with high resolution. Here, we investigate the possibility of improving a multi-resolution approach for experimental shear data with extended flexibility in the model implementation. In this work, we study the combined influence of the developed network for motion estimation along with the improved multi-lag application for SWS reconstruction to advance the performance of SWE. The improvements are demonstrated by comparing with traditional methods.