Comparison of matrix and tensor completion methods for cardiac ultrasound imaging

Cardiac ultrasound requires high frame rates to capture the dynamic motion of tissues. Developing high frame rate imaging sequences necessitates the reduction of sampling requirements, specifically reducing the number of pulse echo events to decrease the capture time of each frame (each event governed by the round-trip time of flight). It has been previously demonstrated that spatial and temporal similarities in the echo data can be exploited to intentionally under-sample the data (missing up to 75% of beams) and robustly reconstruct missing beams using matrix-based tensor completion, an optimization based on rank minimization. We believe that this approach may miss the full multidimensional information contained in a multi-frame image tensor, especially local information. In this work, we evaluate the matrix-based low rank completion approach versus alternative tensor completion methods for cardiac data missing up to 87.5% of beams, an 8x acceleration for data acquisition. We also perform initial investigation into the rank and smoothness properties of the data that may impact the ability to perform tensor completion. We demonstrate that motion compensation is one possible post-processing approach to improve these properties and correspondingly improve tensor completion.