Date of Award
Spring 2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Electrical Engineering (ENAS)
First Advisor
Duncan, James
Abstract
Magnetic Resonance Spectroscopic Imaging (MRSI) is a non-invasive imaging technique for studying metabolism and has become a crucial tool for understanding neurodegenerative diseases, cancers and diabetes. High spatial resolution MRSI is needed to characterize small lesions and intra-lesion heterogeneity, but in practice MRSI is acquired at low resolution due to time and sensitivity restrictions caused by the low metabolite concentrations. Therefore, there is an imperative need for a post-processing approach to generate high-resolution MRSI from low-resolution data that can be acquired fast and with high sensitivity. In computer vision, enhancing image resolution is known as image super-resolution, which is a challenging ill-posed problem as a single low-resolution image can correspond to multiple high-resolution images. Traditional super-resolution MRSI methods relied on model-driven approaches with manually crafted regularizations, which may not accurately capture the characteristics of high-resolution MRSI. Our research proposes a shift towards data-driven methods, particularly deep learning, to achieve super-resolution in MRSI. A significant challenge in developing deep learning-based super-resolution MRSI is the lack of public MRSI datasets suitable for training, due to the difficulty in acquiring a large number of high-resolution ground truth images. To address this, we develop a unique high-resolution in vivo 1H-MRSI dataset acquired from 25 patients with high-grade glioma, designed specifically for training and evaluating deep learning neural networks for super-resolution MRSI. Based on this dataset, we develop multiple deep learning models to upscale low-resolution MRSI metabolic maps to a higher resolution. We first explore how the incorporation of multi-parametric MRI can improve super-resolution MRSI. We develop a multi-encoder architecture with an attention mechanism to extract useful anatomical prior information from each MRI modality. The model is trained using an adversarial loss based on Generative Adversarial Networks (GAN), which helps to restore high-resolution details effectively. Later, we improve this model by introducing network conditioning strategies to modulate network parameters with multiple hyper-parameters. These conditionings allow multi-scale super-resolution, metabolite-awareness and adjustable sharpness within a single network, significantly reducing training time and network size. We subsequently explore the normalizing flow-based models, which are more stable, interpretable and performant than GAN-based models. The performance of our flow-based model is enhanced with novel designs, including learnable base distribution, guide loss and data-consistency loss. The flow-based model inherently allows sharpness adjustment and uncertainty estimation. Expanding on these ideas and aiming to achieve even better performance, we explore diffusion models, which have recently demonstrated superior learning capability than other generative models in a range of tasks. Combining the excellent learning capability of diffusion models and the sampling efficiency of normalizing flow models, we deliver our most performant framework, the Flow-based Truncated Denoising Diffusion Model (FTDDM) for super-resolution MRSI. To address the limitation of conventional diffusion models that require iterating through a large number of diffusion steps, FTDDM shortens the diffusion process by truncating the diffusion chain, and the truncated steps are estimated using a normalizing flow-based network. Experimental results demonstrate that FTDDM outperforms other generative models while speeding up the sampling process by over 9-fold compared to the baseline diffusion model. We conduct extensive evaluations, including neuroradiologists’ assessment from clinical perspectives, which confirm the clinical advantages of FTDDM. In summary, our model is able to transform lower resolution and faster scans into high-resolution metabolic maps. This reduction in scan time elevates MRSI’s potential to be integrated in a routine clinical neuroimaging protocol. Finally, this dissertation also presents preliminary results from two complementary studies aimed at enhancing the resolution of MRSI-based extracellular pH imaging and deuterium metabolic imaging, showcasing the potential of deep learning in advancing other MRSI technologies.
Recommended Citation
Dong, Siyuan, "Data-driven Methods for Super-resolution Magnetic Resonance Spectroscopic Imaging" (2024). Yale Graduate School of Arts and Sciences Dissertations. 1323.
https://elischolar.library.yale.edu/gsas_dissertations/1323
