Harnessing the power of deep learning for ophthalmic disease diagnosis : a study on glaucoma and exudative age-related macular degeneration

Abstract

This thesis aims to bring about a paradigm shift in managing and classifying medical diseases, mainly focusing on two prevalent eye conditions: glaucoma and age-related macular degeneration (AMD). Our research integrates advanced imaging technologies and artificial intelligence (AI) to transition from a static approach, which classifies diseases based on a single time point, to a dynamic one, which monitors disease progression over time. The first part of the thesis focuses on improving the diagnosis of normal tension glaucoma (NTG) using a U-net to segment large central vessels. By calculating radial peripapillary vessel density in a dataset of 3051 OCTA images, this approach discriminates between normal subjects, those with suspected glaucoma, and NTG subjects. The methodology includes layer and clock-hour analysis for every subject, enabling dynamic diagnosis and monitoring of vessel density abnormalities over time. The developed classification model accurately distinguishes between glaucomatous and normal subjects, with an AUC of the ROC and PRC curve of 0.97 and 0.96, respectively. Furthermore, the exact clock hour and layers contributing to the decision-making can be localised, enhancing the model's interpretability. The model's performance on a suspect glaucoma dataset introduces uncertainty, reflecting the model's capacity to learn and interpret ambiguous cases implicitly. This capability could significantly improve early intervention in clinical practice by accurately identifying asymptomatic or suspect cases. The second part of the thesis presents a novel machine-learning model designed to monitor the progression of AMD over time. Using a 'time-differential superimposition' methodology, we create artificial longitudinal datasets for training a model to manage exudative AMD using Macular crosssectional OCT images. This model discerns multiple stages of AMD progression within the same image, marking a significant advancement in the field. The decision-making process is interpreted using Grad- CAM and EigenCAM visualisations, illuminating the areas the AI model deems necessary for AMD progression. The model's ability to focus on areas of clinical relevance and provide near-perfect prognostic classifications on a per-slice basis illustrates its potential utility in a clinical setting, offering actionable insights for clinicians. The model displayed substantial agreement when tested against three independent clinicians on an external validation set. Fleiss's Kappa scores revealed strong agreement among all experts (0.876), between the AI model and the human experts (0.887), and almost perfect agreement between the consensus of experts and the AI model (0.96). These results underscore the interpretability of the AI model, providing clinicians with intuitive and actionable insights. Efforts are underway to translate the AI model into commercially available software, demonstrating the practical and marketable application of the research.This thesis paves the way for future research, advocating a shift from static to dynamic disease classification and management while enhancing the interpretability and reliability of AI models. The findings of this study have far-reaching implications for ophthalmology and beyond, highlighting the potential for seamlessly integrating the developed AI tools into clinical practice. These tools enhance clinicians' capabilities, contributing to more effective and dynamic disease management. Including uncertainty quantification and model interpretability underscores the importance and feasibility of transparent AI applications in healthcare.