Deep learning for fast denoising filtering in ultrasound localization microscopy.

Super-resolution ultrasound (SR-US) microvascular imaging has important implications in the diagnosis of tumor growth and metastasis. Addition of a denoising filter step in ultrasound localization microscopy (ULM) imaging has been shown to effectively reduce the error localizations of microbubbles (MBs) and improve the resolution of SR-US imaging. However, previous image-denoising methods (e.g., block-matching 3-D, BM3D), required long data acquisition and processing times，which made ULM only able to be processed offline. To overcome this limitation, we proposed deep learning (DL) denoising based on contrastive semi-supervised network (CS-Net). For in vitro flow phantom experiment, the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of single microbubble image were 26.9051 dB and 4.0109 dB. For in vivo experiment of New Zealand rabbit tumor, the SNR and CNR were 12.2921 dB and 6.0554 dB. In addition, single microvessel of 25 μm and two microvessels separated by 44.5 μm could be clearly displayed. In terms of denoising speed, the DL denoising speeds for in vitro and in vivo experiments were 0.041 s/frame and 0.062 s/frame, respectively. Experimental results show that DL denoising can improve the resolution of SR-US while reducing denoising time, thereby making further contributions to the clinical real-time imaging of ULM.