Web-Based Retinal Disease Detection Using Deep Learning and Fundus Imaging

Abstract

Retinal diseases are a leading cause of preventable vision loss, especially in areas lacking specialized eye care. Early diagnosis is essential for effective treatment and preventing permanent damage. This thesis develops an automated deep learning system for detecting and classifying retinal diseases from fundus images. We utilized two available datasets in this study: APTOS 2019 for diabetic retinopathy severity grading and ODIR 2019 for multi-disease classification, in- cluding glaucoma, age-related macular degeneration (AMD), cataract, pathologic myopia, retinal vein occlusion (RVO), and other conditions. To enhance training efficiency and improve accuracy, we applied several prepro- cessing techniques, including Contrast Limited Adaptive Histogram Equalization (CLAHE) and Ben Graham’s enhancement method. Various deep learning models including ResNet152, DeiT, Swin-V2, and MaxViT were trained and evaluated un- der different configurations. Among these models, Swin-V2 demonstrated superior performance in terms of both accuracy and generalization capabilities. We implemented a two-stage diagnostic pipeline: the first model detects the presence of retinal abnormalities, while a second model, activated upon detection of diabetic retinopathy, estimates the severity level. To ensure practical applicability, we developed a web-based platform that integrates these models into a user-friendly diagnostic tool. The results demonstrate the potential of combining deep learning approaches with fundus imaging to provide scalable and accurate retinal disease screening. This work establishes a solid foundation for future clinical validation and integration into telemedicine platforms.