Particle Velocity Tracking in Shear Wave Elastography Using a Mamba-Based Spatiotemporal Network

In shear wave elastography (SWE), speckle tracking is employed to obtain the particle velocity from IQ or RF data. Conventional phase shift speckle tracking methods in SWE compute phase changes between adjacent firings, enabling estimation of local particle velocity. Although computationally efficient, these methods are highly sensitive to noise, as they rely on only one adjacent firing for each sample or patch. Deep learning methods have improved particle velocity estimation, with recent approaches using 3D convolutions over limited firings. However, these multi-resolution architectures are restricted to short temporal windows, limiting their ability to model long-range dependencies and generalize across varying conditions. To address this, we introduce Lorast, a long-range spatiotemporal velocity tracking network. This network disentangles spatiotemporal feature extraction by explicitly computing the phase difference between firings as a form of known operator inside the network architecture. Spatial features are extracted using 2D convolutions, while temporal dependencies are modeled with Mamba, a recent sequence modeling architecture outperforming transformers. This separation eliminates the need to learn joint spatiotemporal representations. We evaluate the performance of the compared methods on simulation data, and experimental phantom data.