Deep Learning Architectures for Automated Ocular Disease Recognition

Abstract

Millions of individuals are at risk of preventable vision loss due to optical contingencies such as age-related macular degeneration (AMD), cataracts, diabetic retinopathy, and glaucoma, which pose a major threat to global health. By creating and penetrating deep knowledge models for optic complaint recognition, this study addresses the urgent need for automated, accurate individual tools. We used convolutional neural networks (CNNs) similar to EfficientNet and InceptionResNetV2 to apply transfer knowledge to retinal picture datasets (EyePACS, Messidor, and DRIVE) to categorize various pathologies. To improve model generalizability, our preprocessing channel included normalization, addition, and artifact reduction. The suggested EfficientNet model surpassed birth architectures like ResNet50 and VGG16, achieving 98.2% accuracy and 97.8% F1-score. Key results reveal better performance in identifying diabetic retinopathy stages (AUC 0.99) and early glaucoma (perceptivity 96.5), addressing important individual issues. These findings demonstrate a 12–15% increase over earlier methods that CNN had predicted, making significant progress toward being marks. The study emphasizes the importance of soluble AI for clinical handover while highlighting the transformative potential of deep knowledge in making netting accessible, particularly in low-resource contexts.