From Promise to Practical Reality: Transforming Diffusion MRI Analysis with Deep Learning Enhancement

Abstract

Magnetic Resonance Imaging (MRI) is a non-invasive imaging modality widely used for highresolution anatomical and physiological imaging. Diffusion MRI (dMRI) leverages water molecule diffusion to probe tissue microstructure, offering insights into white matter architecture. However, dMRI data often suffer from low angular/spatial resolution and acquisition artifacts, limiting accurate Fiber Orientation Distribution (FOD) estimation and downstream analyses. Deep learning methods have emerged to mitigate these limitations, but enhancing angular resolution within clinical scanner constraints remains challenging. While approaches like FOD-Net improve angular resolution from single-shell acquisitions, low spatial resolution still introduces blurring, degrading FOD quality. Most existing methods have been validated mainly on healthy subjects, with limited clinical evaluation, and no unified framework addresses both angular and spatial limitations together. This thesis advances deep learning-based dMRI enhancement toward clinical translation by addressing computational inefficiency, limited clinical validation, acquisition protocol dependency, explainability gaps, and fragmented artifact handling. Chapter 2 presents FOD-Net 2.0 for angular resolution enhancement and acquisition protocol optimization. Chapter 3 develops FastFOD-Net, an efficient framework with the most comprehensive clinical evaluation to date, showing improved disease differentiation, connectome interpretability, and reduced sample size needs. Chapter 4 explores uncertainty estimation to assess enhancement reliability. Chapter 5 introduces UFREE, a unified framework tackling angular and spatial resolution jointly via a two-stage learning strategy. Overall, this work delivers robust, scalable, clinically validated solutions and a comprehensive evaluation platform, bridging the gap between deep learning promise and real-world dMRI application.