Evaluating ResNet for Automated Multi-Class Ocular Disease Detection in Retinal Imaging

Ocular diseases can significantly impact quality of life and are especially prevalent in underserved communities where access to qualified ophthalmologists is limited. Since early detection has proven crucial, communities with insufficient healthcare professionals or resources could potentially benefit from a deep learning-based diagnostic system that enables faster, and therefore, more accessible care. This study investigates the potential of implementing a ResNet-based deep learning (DL) model as a scalable solution for ocular disease diagnosis as opposed to simpler machine learning models. Each model was trained using over 5,000 retinal fundus images of six common ocular disorders (age-related macular degeneration, cataracts, glaucoma, diabetic retinopathy, pathological myopia, and hypertensive retinopathy). While existing studies have explored traditional machine learning models such as K-Nearest Neighbors (KNN) and Random Forest models for automated diagnosis based on tabular data, their performance on complex retinal images remains limited. Compared to the KNN and Random Forest, which achieved accuracies of 80% and 68%, respectively, the ResNet model attained an accuracy of 89%, demonstrating improved specificity across all disease categories. This indicates superior capability in handling the inherently complex patterns present in medical imaging. Therefore, implementing ResNet-based screening tools can enhance timely detection and treatment in rural settings where specialist availability is scarce, and better patient outcomes for equitable healthcare access. Future work should explore the viability of real-world implementation through integration with portable diagnostic platforms.