Enhanced EfficientNetB3 Framework for Accurate Multi-Class Eye Disease Classification Using Deep Learning

The purpose of this study is to develop an advanced deep learning model for accurate multi-class eye disease classification, focusing on Cataract, Glaucoma, Diabetic Retinopathy, and Normal conditions. The model uses transfer learning for feature extraction and classification, while leveraging an enhanced EfficientNetB3-based architecture. Utilized a Kaggle dataset comprising 4,217 images, allocating 80% for training and dividing the remaining 20% evenly, with 10% designated for validation and 10% for testing. Data augmentation with random rotations, flips and brightness adjustment was performed to ensure model generalization and reduce model overfitting. The model had an overall accuracy of 95%, and high precision, recall and F1 scores for all classes. The training and validation processes effectively revealed trends in accuracy. Analysis of the confusion matrix also showed minimal misclassifications, and in particular, no misclassifications between overlapping classes, showing the model is reliable. The model’s balanced performance was reinforced by the classification report with macro and weighted averages of 0.94-0.95. Combining transfer learning, data augmentation, and optimized architectures are shown to be important in advancing eye disease detection in this research. The proposed model is a promising model for integration into real time diagnostic tools and mobile applications to detect early and improve clinical decision-making which will ultimately contribute to preventive healthcare strategies.