TL-GWO: Fine-tuned transfer learning with grey wolf optimizer for accurate fundus image-based eye disease classification

Abstract

This study introduces an automated diagnostic framework that detects diabetic retinopathy, glaucoma, and healthy retinas in color fundus images. It leverages several transfer learning (TL) backbones—DenseNet121, ResNet50, ResNet101V2, InceptionResNetV2, and Xception—augmented with additional dense layers, whose architecture and key training hyperparameters are optimized by the Grey Wolf Optimizer (GWO). To enhance image quality and improve feature visibility, Contrast Limited Adaptive Histogram Equalization (CLAHE) is applied during preprocessing, followed by data augmentation techniques such as rotations, shifts, and flips to reduce overfitting. The proposed framework systematically searches the hyperparameter space for optimal configurations, including layer depth, neuron count, activation functions, learning rate, and optimizers, eliminating the limitations of manual tuning. Experimental evaluations on the Eye Disease Image Dataset reveal that integrating GWO-driven optimization with TL significantly improves model generalization and robustness. ResNet101V2 achieved the highest performance with 89.32 % accuracy and an F1-score of 89.37 %, outperforming all other architectures across every evaluation metric. These findings demonstrate the potential of combining advanced TL strategies with metaheuristic optimization to develop reliable and scalable computer-aided diagnostic systems for ophthalmic disease detection.