Deep Learning-Based Multi-Class Classification Approach for Diabetic Eye Diseases

Diabetes-related eye diseases are the leading cause of blindness amongst working-age individuals and represent a serious global health concern. Early detection is essential to prevent vision loss. However, current screening techniques face limitations in scalability and accessibility. A deep learning (DL) method for automatically classifying retinal images into four groups—diabetic retinopathy, normal, glaucoma and cataract—is presented in this work. We developed and assessed a convolutional neural network (CNN) model using transfer learning (TL) with EfficientNetB3 as the foundational architecture, utilising a diverse dataset of over 4,000 retinal images collected from various sources. Our model achieved an overall accuracy of 91.3%, with class-specific sensitivities ranging from 87.6% to 94.2% and specificities from 92.1% to 95.8%. The proposed approach focuses on clinically relevant regions for decision-making. This automated approach could significantly enhance early detection capabilities, especially in resource-limited settings and complement existing ophthalmological services to reduce the burden of preventable blindness.