Rank-Assisted Deep Residual Reconstruction Network for Non-Contrast Ultrasound Imaging of Blood Microvessels

Singular value decomposition of extended ensembles of non-contrast ultrasound echoes can enable imaging of deep-seated microvasculature and underlying microflow. This process, however, is computationally expensive and requires ad hoc tuning of parameters, e.g. tissue clutter rank, based on some empirical criteria not generalizing to different real-world scenarios. Here, we present a novel model-based deep learning approach that accelerates tissue removal via simplified clutter suppression followed by a refinement reconstruction network for high resolution imaging of microvasculature. Additionally we show a realistic simulation model to create extended dataset for training a deep residual reconstruction network with access to true ground truth. We present results in terms of computation speed-up, resilience to improper clutter removal parameters, and contrast improvement using our simplified rank-assisted deep residual reconstruction network (RA-DR2Net) on both simulated and real in vivo data.