Computationally Efficient SVD Filtering for Ultrasound Flow Imaging and Real-Time Application to Ultrafast Doppler

Over the past decade, ultrasound microvasculature imaging has seen the rise of highly sensitive techniques, such as ultrafast power Doppler (UPD) and ultrasound localization microscopy (ULM). The cornerstone of these techniques is the acquisition of a large number of frames based on unfocused wave transmission, enabling the use of singular value decomposition (SVD) as a powerful clutter filter to separate microvessels from surrounding tissue. Unfortunately, SVD is computationally expensive, hampering its use in real-time UPD imaging and weighing down the ULM processing chain, with evident impact in a clinical context. To solve this problem, we propose a new approach to implement SVD filtering, based on simplified and elementary operations that can be optimally parallelized on GPU (GPU sSVD), unlike standard SVD algorithms that are mainly serial. First, we show that GPU sSVD filters UPD and ULM data with high computational efficiency compared to standard SVD implementations, and without losing image quality. Second, we demonstrate that the proposed method is suitable for real-time operation. GPU sSVD was embedded in a research scanner, along with the spatial similarity matrix (SSM), a well-known efficient approach to automate the selection of SVD blood components. High real-time throughput of GPU sSVD is demonstrated when using large packets of frames, with and without SSM. For example, more than 15000 frames/s were filtered with 512 packet size on a 128 × 64 samples beamforming grid. Finally, GPU sSVD was used to perform, for the first time, UPD imaging with real-time and adaptive SVD filtering on healthy volunteers.