High-dimensional MR spectroscopic imaging integrating physics-based modeling and machine learning

Magnetic resonance spectroscopic imaging (MRSI) is a powerful modality that allows noninvasive mapping and quantification of a number of endogenous molecules, providing a unique molecule-specific window into the human body, and has shown significant impact in many basic science and translational studies. Nevertheless, in vivo applications of MRSI are still hindered by several long-standing technical challenges including low sensitivity, poor resolution, slow imaging speed, and contamination from nuisance signals. The fundamental reasons for these limitations are the inherently low abundance of the molecules of interest and the high dimensionality of the underlying imaging problem due to the need to encode and recover the high-dimensional spatiospectral/spatiotemporal image function. These problems pose tremendous challenges to conventional imaging hardware and software paradigms. Recent advancements in advanced instrumentation, computing, and machine learning (ML) technologies have brought unparalleled opportunities to tackle these challenges and innovate a new generation of imaging systems and workflows. This dissertation focuses on developing novel imaging approaches for robust, high-resolution, and high signal-to-noise-ratio (SNR) multidimensional MRSI that synergize advanced MRI systems, novel data acquisition strategies, physics-based modeling, and data-driven machine learning. Specifically, we have developed (1) A robust deep learning framework that effectively utilizes spectral features from quantum mechanical simulations and experimental parameter estimations to learn accurate nonlinear low-dimensional models of the high-dimensional spectroscopic signals; (2) Novel formulations that effectively integrated the learned representations, spatiospectral encoding model and other complementary prior information for solving different long-standing technical problems in MRSI, i.e., SNR-enhancing reconstruction and signal separation; (3) Efficient algorithms that leverage the advantages of both the traditional iterative algorithms and learning based reconstruction, with in-depth complexity and convergence analysis. We have conducted thorough evaluations of our methodologies using carefully designed simulations and in vivo experiments on both healthy and patient populations. The results demonstrated the effectiveness of our methods in improving MRSI reconstruction and molecular signal quantification over state-of-the-art methods, under practical scenarios. We expect the methods described in this thesis to provide a set of powerful MRSI technologies for in vivo metabolic studies. The proposed imaging framework also presents new possibilities for providing high-quality information for detecting and quantifying physiological and pathological biochemical variations in both healthy subjects and patients and may lay a foundation for future clinical translations.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Bioengineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Li, Yahang

Contributors dc:contributor

• Lam, Fan
• Sutton, Brad
• Anastasio, Mark
• Insana, Michael

• Mr Spectroscopic Imaging
• Spectroscopy
• Low-dimensional Models
• Neural Network
• Manifold Learning
• Spatiospectral Constraint
• Complex Convolutional Neural Network
• Deep Autoencoder
• Deep Learning
• Denoising
• Low-dimensional Modeling
• Multi-te 1h-mrsi
• Regularized Reconstruction
• Short Te 1h-mrsi
• Signal Separation
• Learning Low-dimensional Manifold Projection.

Statement dc:rights

• Copyright 2023 Yahang Li

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/122092

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F122092