Rizzo, Rudy (2023). Multiparametric MR Spectroscopy: evaluation of quantitative frameworks based on modeling and deep learning. (Thesis). Universität Bern, Bern

Preview

Text 23rizzo_r.pdf - Thesis Available under License Creative Commons: Attribution-Noncommercial (CC-BY-NC 4.0). Download (129MB) | Preview

Abstract

Magnetic Resonance Spectroscopy (MRS) is a superb technique for the diagnosis and treatment monitoring of many diseases. It relies on the non-invasive acquisition and evaluation of the metabolic content in a selected body area. In fact, levels of metabolites’ concentration can provide cell-type-specific insights into the function and pathophysiology of various organs. Furthermore, the chemistry of the human body can be explored through sensitization of the Nuclear Magnetic Resonance (NMR) signal to a variety of properties beyond plain metabolite concentrations. This is the realm of multiparametric MRS. T2 relaxation rates contain essential information on the cellular microenvironment, acting as potential biomarkers for abnormalities and carrying a crucial role in absolute quantification. In contrast, metabolic diffusion is driven by particle features and the geometry of intracellular compartments. It may provide inference on the chemical exchange and bonding, untangle interactions between metabolites and other compounds, or discern behaviors across different tissue compartments. Conventional techniques for measuring these quantities, such as Diffusion-Weighted MRS (DW-MRS) or Multiple Echo Times (MTE) acquisitions, are time-consuming, inefficient, and rarely used within clinical timeframes. Instead, to solve the issues for absolute quantification, clinical MRS relies on regionally T2 tabulated values, which do not account for inter-subject variability and are often lacking, particularly in pathology. Besides ignoring a relevant MR quantity, such assumptions introduce quantification errors and biases. To speed up the quantification tasks, recent developments in machine learning algorithms, explicitly concerning deep learning (DL) architectures, have found an increasing interest within the scientific community. DL toolboxes come as pre-trained models where run-time metabolic concentrations are provided on-the-fly. However, despite the uprising research throughput, a low acceptance rate for such tools is currently found in the clinics. On the one hand, that is due to the high complexity of such architectures, which translates into a low level of interpretability still conveyed to black-box assumptions. But on the other hand, DL does not yet provide reliable and well-established uncertainty measures of its predictions, which leaves the average user unaware of potential intrinsic errors. The current work starts with an introduction given in Chapter 1, where a brief overview of the relevant properties of NMR, from classic and quantum physic perspectives, is treated. It follows, in Chapter 2, a quick dive into a practical MRS experiment where the fundamentals of signal excitation and recording, as well as the crucial properties of MRS signals, are discussed in a synthetic yet realistic context. Chapter 3 wants to disclose the MR properties and challenges of two human organs (brain and prostate) that drove the focus of the various investigations encountered during this work. It follows an overview of the methods, clinical interests, and potentials in the context of multiparametric MRS, mainly focusing on estimating metabolite-specific T2 rates and diffusion properties, which is disclosed in Chapter 4. The background part is concluded in Chapter 5 with an overview of the principles of DL, exploring in more detail the current challenges on interpretability and uncertainty measures as well as the state-of-the-art designs deployed for MRS quantification. The first part of the main contribution of this work is given in Chapter 6. DL is deployed across many architectural designs and is twinned to tailored MRS processing aiming to enhance features in the data that are more or less prone to DL computation. Results are further analyzed concerning dataset biases and possible training strategies to overcome such limitations, namely, ensemble of models and data augmentation. A window into interpretability and uncertainty measures is also explored, offering a first method to integrate MRS predictions of concentrations with their reliability. An analysis that compares these measures to traditional Cramer Rao Lower Bounds (CRLBs) in fitting follows. The second part of the main contributions of this work is explored in Chapter 7 with a focus on different aspects of multiparametric MRS. First, the urge of speed in multiparametric MRS to simultaneously and accurately produce metabolite concentrations and T2 rates is explored, introducing a novel acquisition method that combines bi-dimensional fitting and truncated multi-echo acquisitions. The benefits and limitations of the methodology are disclosed both in a single-voxel experimental fashion and in a 2D MRSI setup targeting the human brain. Second, diffusion-weighted MRS is deployed for the first time in the human prostate to untangle and explore MRS properties. The results offer an alternative viewpoint to the complex chemical bonding of proteins with some of the main prostatic metabolites. Eventually, an initial investigation of DW-MRS for pathological cases is outlined. Although severely limited by a small cohort of patients, it promises exciting potential currently interpreted based solely on the prostate microstructure.

Item Type:	Thesis
Dissertation Type:	Cumulative
Date of Defense:	22 May 2023
Subjects:	500 Science > 570 Life sciences; biology 600 Technology > 610 Medicine & health
Institute / Center:	04 Faculty of Medicine > Department of Radiology, Neuroradiology and Nuclear Medicine (DRNN) > Institute of Diagnostic and Interventional Neuroradiology
Depositing User:	Hammer Igor
Date Deposited:	11 Feb 2025 07:43
Last Modified:	11 Feb 2025 23:25
URI:	https://boristheses.unibe.ch/id/eprint/5812

Actions (login required)

View Item