Deep learning-based super-resolution ultrasound microvascular imaging

Abstract

Super-resolution ultrasound microvascular imaging (SR-UMI) is a fast-growing field that leverages contrast microbubbles (MB) to achieve super-resolution and track blood flow. Conventional SR-UMI is primarily based on MB localization and tracking, which is a challenging task in vivo with complex MB responses subject to various tissue characteristics and imaging settings. Conventional MB localization and tracking approaches based on experimentally calibrated or engineered MB templates typically fall short of capturing the widespread distribution of MB signals in vivo, leading to suboptimal SR-UMI performance. In contrast, deep learning (DL) provides many powerful tools to better represent the complex landscape of MB and blood flow signals that lead to enhanced SR-UMI. In this talk, I will first introduce several DL-based MB localization techniques with discussion on challenges and solutions involving DL training within the context of SR-UMI. I will then introduce several other novel DL-based SR-UMI techniques that do not rely on MB localization or tracking to achieve super-resolution and infer blood velocity. Some of the techniques are contrast-free. In vivo imaging examples from chorioallantoic membrane as well as mouse brain will be presented as validations of the DL-based SR-UMI techniques. In vivo applications in cancer and Alzheimer’s disease will also be presented and discussed in this talk.