Quantitative Image Analysis for Objective Classification of Retinal Diseases

As a neurovascular network, the retina can be a target of various systemic conditions and diseases. For example, diabetic retinopathy (DR), the microvascular complication of diabetes mellitus (DM), is a progressive sight-threatening disease that is asymptomatic in the early stages. As a transparent structure, the retina can be visualized noninvasively by various imaging modalities such as optical coherence tomography (OCT), and OCT-angiography (OCTA). Both modalities provide high-resolution cross-sectional imaging, with OCT enabling the visualization of multiple retinal layers and OCTA enabling visualization of the individual retinal vascular layers. In the early stages of eye diseases, digital quantitative biomarkers may be able to detect subtle neurovascular abnormalities. The development of quantitative OCT and OCTA characteristics may also be used in machine learning models for treatment assessment, disease detection and progression. Therefore, in this thesis we explore two aims. The first aim is the development of quantitative features in retinal images. In this aim, we quantified neurovascular abnormalities in the retina using OCT and OCTA features in retinopathies such as DR. The second aim is objective classification of retinopathy using retinal images. In this aim, we demonstrate deep learning for automated DR classification and artery-vein classification in OCTA. Additionally, we demonstrate deep learning OCTA construction using single-scan volumetric OCT. The development of quantitative features and application of AI in routine clinical use may help to increase efficiency of healthcare professionals with rapid and automated screening of retinal diseases.