Supporting Data for "An Efficient and Comprehensive Framework for Ultrasound Localization Microscopy"

The human body is an extraordinary machine where the circulatory system delivers oxygen and nutrients everywhere via the vascular network. Small blood vessels like capillaries predominantly perform this exhaustive work and often reveal early signs prior to significant pathological changes, like cancer, diabetes, and atherosclerosis. However, a medical imaging modality that could resolve blood vessels at the microscale, deep within tissues, was unavailable until the advent of ultrasound localization microscopy (ULM).Through the precise localization and tracking of microbubbles (MBs), ULM is capable of reconstructing microvascular network and flow velocity maps with sub-wavelength resolution. Yet, ULM faces several challenges, including the long data acquisition time to accumulate enough MBs to fully reconstruct the vasculature, user-dependent parameter selection, a lack of ground truth for validation, and so on. Particularly, the in vivo ULM results require rigorous evaluation that remains challenging due to the absence of a definitive ground truth, despite the development of numerous algorithms designed to optimize the ULM pipeline.This thesis introduces an efficient and comprehensive ULM framework that aims to tackle the aforementioned challenges and facilitates its clinical translation. Specifically, the state-of-the-art ULM workflow requires exhaustive empirical selection of parameters across different systems, impeding ULM from its clinical translation and adoption. Hence, the first objective is to establish a robust, efficient, and general ULM pipeline for ULM without manual intervention as presented in Chapter 3. This is achieved by adapting a vision transformer (ViT) deep learning model in the ULM workflow for speckle reduction and MB signal extraction. Ample synthetic datasets were constructed to train the ViT-based model, which was demonstrated to excel in accuracy and robustness on public datasets, including in silico flow with additive electronic noise and four diverse in vivo tissues. The adapted ViT seamlessly integrates with state-of-the-art downstream steps, toward parameter- and expertfree ULM for potential clinical applications. However, the evaluation on the in vivo datasets is not direct and remains limited due to the lack of ground truth. Chapter 4 subsequently presents the first of its kind onchip vasculature protocol, allowing for customization of versatile vascular patterns on agarose-based material. This protocol provides a validation phantom with high optical and acoustical transparency, showing significant potential for the development and optimization of ultrasound microvascular imaging techniques, beyond ULM. Chapter 5 further advanced the protocol and constructed functional chips for comprehensive evaluation of ULM. We systematically revisited the resolution limit of ULM, revealing its ability to resolve adjacent vessels as the key resolution metric. The velocity validation of ULM was further employed with the help of our proposed functional chips, providing credible evidence that the velocity information derived from ULM was reliable to demonstrate the in vivo hemodynamic world.The thesis work tackled major challenges in ULM from the perspective of both algorithm development and methodology innovation, providing novel insights and contributions into the state-of-the-art ULM workflow, paving the way for robust clinical translation of ULM.