Deep learning for super-resolution ultrasound microvessel imaging

Ultrasound localization microscopy (ULM) is an emerging ultrasound vascular imaging technique that overcomes the resolution-penetration compromise of ultrasound imaging. ULM has demonstrated preclinical and clinical significance for a myriad of applications that benefit from microvascular imaging biomarkers. However, at present, the potential of ULM has not yet been fully realized due to the technical limitations of long data acquisition and post-processing time. The root cause of the problem is that conventional ULM requires low microbubble (MB) concentrations for robust image reconstruction because lower MB concentrations lead to better spatial separation between individual MB signals. This requirement results in a long data acquisition time because a lower MB concentration in the blood stream makes it longer for MB localization signals to fully perfuse the microvascular structure. Therefore, developing ULM reconstruction techniques that can better utilize MB signals under high MB concentrations is essential for facilitating faster and better ULM. In the past few years, deep learning (DL)-based techniques have gained popularity in medical imaging applications. DL has demonstrated superior performance in many aspects of medical imaging than conventional methodologies, thanks to its ability of drawing information from complex features with high levels of abstractions. Additionally, most of the DL techniques do not require explicit modeling of the input-output relationship, which is often challenging for medical imaging processes. Inspired by the recent success of DL in medical imaging, this dissertation research was devoted to the investigation of applying the principles of DL in ULM reconstruction, with the goal of improving MB localization and tracking performance and ultimately the imaging speed of ULM. First, we proposed the study of DL-based implementations of MB localization and tracking in the ULM workflow. The proposed techniques were validated using both in silico and in vivo data sets, showing significantly improved ULM reconstruction performance under high MB concentrations. To further improve the super-resolution microvascular imaging speed, we proposed an alternative DL-based imaging solution that directly provides super-resolved microvessel velocity information using spatial-temporal MB data as the input to the neural network. This method side-steps the root cause of low temporal resolution of ULM and has demonstrated great potential for real-time super-resolution imaging of tissue microvasculature.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Chen, Xi

Contributors dc:contributor

• Song, Pengfei
• Anastasio, Mark
• Oelze, Michael
• Lam, Fan

• Ultrasound Imaging
• Super-resolution
• Deep Learning

Statement dc:rights

• Copyright 2024 Xi Chen

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/124646

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F124646