Optical Flow Assisted Super-Resolution Ultrasound Localization Microscopy using Deep Learning

Abstract

Ultrasound localization microscopy provides resolution enhanced ultrasound images and demonstrates clinical potential in myocardial infarction and diabetes. The conventional model-driven methods localize the microbubble by tracing the peak of the point spread function. Such numerical schemes demonstrate weakness in identifying superimposed microbubbles, indicating the limitations for super-resolution (SR) images. Recently, learning-based approaches have been studied for precise localization of densely distributed microbubbles. However, prior arts reconstruct the SR images from static B-mode images, which results in inconsistent localization of microbubbles across sequential frames. In this paper, we propose a temporal relational ultrasound microscopy network (TRUM-Net). The TRUM-Net adopts optical flow estimation of consecutive frames and a feedback loop for detailed super-resolution imaging. The proposed scheme enhances the accuracy of microbubble localization by 25.8% and the structural similarity up to 54.9%.