Dissertation Defense: Jipeng Zhang
"Retinal AI for Macular Degeneration: Prediction, Image Quality, and Privacy Protection", Department of Biostatistics and Health Data Science, School of Public Health.
Advisors and Committee Chairs: Wei Chen and Qiong Wu
Abstract:
Retinal imaging is fundamental to the diagnosis and monitoring of age-related macular degeneration (AMD), a leading cause of irreversible vision loss in older adults. Because AMD progression is gradual, heterogeneous, and asymptomatic in its early stages, clinical assessment alone is insufficient, creating demand for quantitative methods that leverage retinal imaging data for objective diagnosis and early risk prediction. This dissertation advances retinal AI for AMD across three areas: longitudinal AMD prediction, annotation-efficient image quality assessment, and privacy-preserving image transformation.
The first part addresses longitudinal prediction of AMD progression. We proposed LONGL-Net, which predicts future AMD risk by first synthesizing future fundus appearances through a U-Net-based GAN, then grading disease severity from both observed and predicted images. Building on this, we developed TV-LSTM to integrate longitudinal image embeddings with genetic variants while explicitly modeling irregular follow-up intervals inherent to real-world clinical data. Together, these frameworks advance the prediction of AMD progression through longitudinal imaging, multimodal learning, and time-dependent modeling to support earlier identification of high-risk individuals.
The second part focuses on assessing image quality to improve the reliability of retinal AI pipelines. We proposed FM-PUCon, a weakly supervised framework combining retinal foundation-model embeddings, positive-unlabeled pseudo-label mining, and supervised contrastive learning. This approach enables quality classification without exhaustive manual labeling of poor-quality images, substantially reducing annotation burden. Extensive evaluations demonstrated that FM-PUCon transfers successfully across fundus photography and optical coherence tomography (OCT), improving the robustness of downstream retinal AI applications.
The third part explores privacy-preserving retinal image generation to support secure sharing of retinal image data. Fundus photographs implicitly embed biometric information that allows AI models to predict sensitive attributes, such as biological age and genetic signals, posing privacy risks when images are shared or repurposed beyond their intended clinical use. We investigated whether image-to-image generative models could transform fundus photographs to suppress this information while preserving image quality and disease-relevant features. We found that suppressing age-related information also degraded AMD prediction, revealing that age-related and AMD-related features are closely entangled in fundus images. These findings characterize a fundamental privacy–utility trade-off for age-related retinal diseases.
Contribution to public health
This dissertation contributes to public health by supporting earlier detection of AMD risk, strengthening the reliability of automated retinal screening pipelines, and informing strategies for privacy-preserving sharing of ophthalmic imaging data. More broadly, the methods developed in this dissertation are applicable to other retinal diseases and medical imaging domains, contributing to the development of trustworthy AI in healthcare.