UC Irvine

The rapid proliferation of advanced medical imaging modalities, such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT), has generated vast volumes of three-dimensional data that are central to modern clinical practice. While these datasets offer unprecedented opportunities for diagnosis, treatment planning, and disease monitoring, they also pose major computational challenges: robust methods are required to capture anatomical variability, ensure topological plausibility, and generalize across patient populations. Traditional approaches, relying on hand-crafted features and iterative optimization, are often limited in efficiency and scalability. More recently, deep learning has transformed medical image analysis by enabling end-to-end frameworks that automatically extract hierarchical features and model complex anatomical structures.This dissertation advances the state of 3D medical image analysis across three key research areas. First, in image registration, it introduces neural velocity field representations for accurate diffeomorphic transformations, an on-the-fly guidance framework that generates pseudo ground truth during training, and an adaptive attention-based model for coarse-to-fine correspondence discovery. Together, these contributions seek to overcome limitations of both traditional and early learning-based methods. Second, in 3D shape representation, it presents a hybrid neural diffeomorphic flow that learns a common anatomical template and generates new, biologically plausible instances while preserving topological consistency. Third, in data generation, it develops advanced generative models to address dataset scarcity: a cross-dimensional supervision framework for OCT-to-OCTA translation that reconstructs fine vascular details, and a paired 3D image–mask generation model that synthesizes realistic volumetric data for downstream tasks.Collectively, this work demonstrates how deep learning can deliver accurate, efficient, and generalizable solutions for 3D medical imaging. By advancing registration, representation, and data generation, the dissertation lays the foundation for intelligent computational tools that enhance clinical diagnosis and biomedical research.