Machine learning-assisted multiparametric and dynamic mapping of multiple molecules with MR spectroscopic signals

Li, Yudu

Permalink

https://hdl.handle.net/2142/116041 Copy

Description

Title

Machine learning-assisted multiparametric and dynamic mapping of multiple molecules with MR spectroscopic signals

Author(s)

Li, Yudu

Issue Date

2022-07-06

Director of Research (if dissertation) or Advisor (if thesis)

Liang, Zhi-Pei

Doctoral Committee Chair(s)

Liang, Zhi-Pei

Committee Member(s)

• Anastasio, Mark
• Bresler, Yoram
• Sutton, Brad

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Magnetic resonance imaging
• Magnetic resonance spectroscopic imaging
• Machine learning
• Multiparametric imaging
• Molecular imaging

Language

eng

Abstract

Magnetic resonance (MR) signals naturally carry information about a wide spectrum of biologically important molecules and their biophysical parameters at steady and/or dynamic state. This rich set of information provides comprehensive insights into tissue properties, which are not only useful for understanding the structural, functional, and metabolic organization of the organ being studied, but also helpful for the diagnosis and treatment of a broad range of diseases. However, current MR methods can only detect a few number of molecules and parameters with very limited imaging performance in terms of speed, signal-to-noise ratio (SNR), and spatiotemporal resolution. This dissertation focuses on developing novel imaging approaches, in both data acquisition and data processing, to enable multi-parametric and dynamic imaging of multiple molecules in high spatiotemporal resolution. For multi-parametric imaging of multiple molecules, we have made technical innovations and advances in both data acquisition and processing. In data acquisition, a novel pulse sequence is designed which synergistically integrates MR imaging, spectroscopy, and parametric mapping to achieve the desired multi-modal imaging capability. Our data acquisition scheme also achieves optimal acquisition efficiency by leveraging a small flip angle with short repetition time, advanced multi-dimensional trajectories, and sparse sampling. Data processing is a key component of the proposed method which consists of image reconstruction and parameter quantification. For image reconstruction, a novel machine learning-based framework is proposed to address image reconstruction from highly noisy and sparse imaging data. Our method uses a novel tensor image model for efficient representation of the desired signal and effective absorption of prior information learned from both spin physics (e.g., resonance structure of each molecule) and training data (e.g., distribution of physiological and experimental variations). The proposed signal model and learned image priors are incorporated using the Bayesian estimation framework for image reconstruction. SNR and resolution analyses demonstrate that our method achieves substantial SNR improvement and moderate resolution degradation. For parameter quantification, novel machine learning-based algorithms are developed for magnetic resonance spectroscopic imaging (MRSI) spectral quantification and myelin water fraction (MWF) estimation. Our proposed methods are featured by the use of improved signal models and significant priors learned from both physics and training data. Uncertainty and perturbation analyses reveal that our quantification methods achieve significant reduced estimation uncertainty and enhanced robustness. For dynamic imaging of multiple molecules, we extend the proposed image reconstruction framework by further leveraging deep learning to enable high-resolution dynamic MRSI. Our method synergistically integrates physics-based subspace modelling and data-driven deep learning for effective sensitivity enhancement, making high-resolution dynamic metabolic imaging possible. Specifically, a novel subspace model was used to represent the dynamic MRSI signals; deep neural networks were trained to capture the low-dimensional manifolds of the spectral and temporal distributions of practical dynamic MRSI data. The learned subspace and manifold structures were integrated via a regularization formulation to effectively enhance the imaging sensitivity. The feasibility, reproducibility, and accuracy of the proposed imaging approach are evaluated through computer simulations, phantom studies, and in vivo experiments.

Graduation Semester

2022-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/116041

Copyright and License Information

Copyright 2022 Yudu Li

Owning Collections