GPU-based Blockwise Non-Local Means Filtering with Microbubble Separation for 3D Ultrasound Localization Microscopy

Three-dimensional (3D) ultrasound localization microscopy (ULM) using a 2D matrix probe and microbubbles (MBs) has been introduced to visualize microvasculature in three dimensions, surpassing the ultrasound diffraction limit. Numerous studies have shown that 3D ULM offers not only can achieve high resolution, but also solve the issue of out-of-plane motion. However, when the microbubble (MB) signal-to-noise ratio (SNR) is low, distinguishing MB signals from background noise can be challenging. While 2D blockwise spatiotemporal nonlocal means (NLM) filtering effectively reduces random noise by leveraging repetitive track patterns, the wide intensity range of MBs can result in the loss of weak MB signals in small vessels, especially when they differ significantly from stronger signals in nearby larger vessels. To address this, the MB separation method was used to separate slow-flow MBs from fast-flow MBs. We then extended the blockwise NLM from 2D to 3D, improving the detection of low-intensity, slow-flow MB signals amidst noise. Although this 3D NLM approach offers superior noise suppression compared to its 2D counterpart, it requires more computational power. To mitigate this, we developed a GPU-based 3D NLM architecture, which was evaluated for both processing time and noise suppression. The proposed method improved microbubble signal contrast and demonstrated fast denoising, making 3D ULM more suitable for clinical applications.