Dissertations, Theses, and Capstone Projects

Date of Degree

6-2026

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy

Program

Computer Science

Advisor

Chia-Ling Tsai

Committee Members

Mikael Vejdemo-Johansson

Jonathan Gryak

Chao Chen

Subject Categories

Artificial Intelligence and Robotics | Biomedical Informatics | Data Science | Diagnosis

Keywords

medical imaging, image synthesis, weaker supervision, image segmentation

Abstract

Comprehensive visualization of the retina is essential for diagnosing and monitoring blinding diseases such as Diabetic Retinopathy and Retinopathy of Prematurity (ROP), where pathological changes often extend beyond a single field of view. Despite significant advances in automated retinal image analysis, clinical deployment remains limited by two fundamental data gaps: a structural learning gap, arising from scarce expert annotations and poor generalization across imaging domains, and a spatial coverage gap, caused by the difficulty of acquiring multi-view retinal images in fragile populations. Although these challenges are often addressed independently, this dissertation argues that they are tightly coupled: accurate, topology-preserving vascular structure substantially improves the anatomical validity and patient specificity of synthesized retinal views, which is critical for enabling reliable automated analysis in low-resource clinical settings.

This dissertation proposes a unified framework that progresses from robust structure learning to patient-specific retinal view synthesis. To support consistent validation across all studies, we introduce ROP-500, an expert validated benchmark from preterm infant retinal imaging that serves as a common evaluation testbed rather than a primary contribution. ROP-500 enables standardized assessment of vessel segmentation, disease staging, and topology-aware analysis while highlighting the challenge of generalization beyond densely annotated domains.

To address the structural learning gap, the dissertation develops scalable, label-efficient methods for vascular structure extraction. We first present an adversarial semi-supervised segmentation framework that substantially reduces reliance on expert annotations while preserving fine-grained vascular detail. Expanding the scope to domain shifts, we introduce a topology-aware unsupervised domain adaptation approach that enforces vascular connectivity across heterogeneous imaging domains using principles from topological data analysis.

Leveraging these topological principles, the dissertation then addresses the spatial coverage gap through topology-aware synthesis modeling. We present a topology-aware synthesis framework that evolves from preventing vascular hallucination in single-view settings to patient-specific retinal extrapolation. We first demonstrate that persistence-based losses effectively suppress vascular hallucinations in a single-view setting. Building directly on this foundation, we extend the framework to a Dual-Topology and Geometry-Aware architecture. By integrating multi-image conditioning with a predictive geometry encoder, this capstone model anchors synthesis to patient-specific spatial relationships. Collectively, this dissertation demonstrates that topology-aware learning bridges both data gaps, advancing the feasibility of automated whole retina analysis.

Download

Share

COinS