Deep Generative Models for Shape Analysis, Realistic Brain Image Generation, and Longitudinal Analysis 16 views

Understanding how brain anatomy varies across individuals and changes with aging and neurodegenerative disease requires methods that capture subtle anatomical differences, generate realistic brain images, and model longitudinal changes. However, existing approaches often rely on global measurements or sacrifice fine anatomical details for computational efficiency. This dissertation develops computational methods involving statistical shape analysis, deep generative modeling, and diffeomorphic registration to address these challenges. First, a point-based shape analysis framework quantifies localized shape asymmetry, revealing morphological differences overlooked by traditional volumetric measurements. Second, a correspondence-preserving diffusion model generates smooth and diverse anatomical shapes while maintaining point correspondences necessary for population-level analysis. Third, RealDeal, an image-to-image diffusion model, enhances the realism and anatomical fidelity of synthetic brain MRIs by restoring fine structural details lost during latent compression. Finally, a predictive registration framework models continuous, subject-specific anatomical trajectories to predict future brain changes. Together, these contributions advance the measurement, synthesis, and longitudinal characterization of brain anatomy, providing computational tools for studying healthy aging and neurodegenerative diseases.

PHD (Doctor of Philosophy)

Generative Models; Brain MRI; Shape Analysis; Hippocampal Asymmetry; Point-Based Models

English

All rights reserved by the author (no additional license for public reuse)

2026-10-29