Unified Foundational Modeling for Diffusion-weighted Magnetic Resonance Imaging

Abstract

This dissertation presents a unified foundation for diffusion-weighted magnetic resonance imaging (dMRI) modeling by developing deep learning–based frameworks that improve fiber orientation estimation, free-water partial-volume correction, and cross-dataset generalizability across heterogeneous acquisition schemes. I first introduce two fiber orientation distribution (fODF) modeling strategies—Deep CSD and Deep Multi-CSD—which leverage patch-based CNNs, contrastive learning, and spherical convolutional networks with dynamic heads to enhance scan-rescan reproducibility and enable robust estimation across multi-shell and reduced-direction protocols. I then develop deep learning approaches for free-water estimation in both single-shell and multi-shell settings, including a token-aware transformer (Ts-FWE) that incorporates acquisition-specific configuration tokens to achieve a generalizable “one-for-all” architecture capable of handling unseen protocols. Building on these methods, I propose polyhedral representations and transformer models that reparameterize diffusion signals using icosahedral and hemisphere-aware sampling, culminating in the Unified Polyhedral Representation (UPER), a geometry-aware framework that unifies spatial, angular, and radial diffusion information and supports both supervised and masked-autoencoder self-supervised learning. Evaluated across large multi-site datasets—including HCP-YA, HCP-Aging, MASIVar, MASSIVE, VMAP, ADNI, BLSA, and NACC—these methods consistently improve reconstruction accuracy, microstructural estimation, and intra-subject consistency under diverse acquisition schemes. Collectively, this work establishes a scalable, acquisition-agnostic foundation for next-generation dMRI modeling, enabling more reliable, reproducible, and harmonization-free diffusion MRI analysis.