Scalable Hybrid Deep Learning-Based Architecture for Glaucomatous and Healthy Eye Classification in Retinal Fundus Images

Abstract

Glaucoma remains one of the leading preventable causes of irreversible blindness worldwide, with early detection being essential for preserving vision. This study presents a hybrid deep learning framework that integrates VGG16 and ResNet-50 architectures to improve the classification of glaucoma severity using retinal fundus images. A balanced dataset of 2,081 images was utilized, with data augmentation and the Synthetic Minority Over-sampling Technique (SMOTE) applied to address class imbalance and enhance model generalization. All images were normalized and resized to 224 × 224 pixels, and training was conducted for 50 epochs with a batch size of 32, resulting in approximately 14.7 million trainable parameters. The proposed hybrid model achieved an average accuracy of 83%, surpassing standalone VGG16 (80%), ResNet-50 (70%), and EfficientNet-B0 (51%), underscoring the benefit of combining hierarchical and residual feature extraction. In addition, it achieved a precision of 82%, recall of 81%, and an F1-score of 82%, with a final loss value of 0.36. Quantization-aware training was employed to optimize computational efficiency, reducing the average prediction time to 95 milliseconds per image and enabling near real-time deployment in low-resource clinical environments. While some misclassifications were observed due to close visual similarities among glaucoma stages, the proposed approach demonstrates strong potential as a scalable and efficient solution for automated glaucoma screening and early detection.