Envisioning Automated Glaucoma Screening: Domain Generalization for Deep Learning-Based Glaucoma Classification

Abstract

Rapid advancements in deep learning algorithms for computer vision tasks have produced powerful models that can accurately classify diseases from medical images across a variety of specialties. In the field of ophthalmology, these systems have the potential to enable large-scale eye disease screening programs in areas that lack access to vision care or trained specialists. However, deep learning models must be able to generalize to unseen domains, such as retinal images from different healthcare institutions, machines, and patient populations, before they can be deployed for population-wide clinical screening. In this thesis, we analyze domain shift across four public retinal image datasets and investigate its effect on baseline glaucoma classification model performance. We find that domain shift severely degrades classification ability, and, specifically, image features extracted during model training do not generalize to out-of-domain images. We further test three state-of-the-art domain generalization methods and find that one method marginally improves model generalization, though not to an adequate level for clinical use. In the final chapter, we connect the findings of our research to broader themes in global health and health policy. Overall, this thesis begins to fill a noticeable gap in domain generalization research for deep learning-based glaucoma classification.