Identifying glaucoma with deep learning by utilizing the VGG16 model for retinal image analysis
Vaibhav C. Gandhi, Priyesh P. Gandhi, Ankit D. Oza, Mohammed Kh. Al-Nussairi, Ahmed Adnan Hadi, Ahmed A. Al‐Amiery, Amanuel Zewdie
Abstract
Glaucoma is a leading cause of irreversible vision loss, and its timely detection remains a major clinical challenge due to subtle variations in optic nerve structure and inconsistent image quality across datasets. This study presents a hybrid glaucoma detection framework that combines the deep feature extraction capacity of the VGG16 network with the decision-making strength of machine learning classifiers, including Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting (GB). Unlike prior CNN-only or handcrafted ML pipelines, this work introduces a compact hybrid architecture that reduces trainable parameters by ∼40 %, integrates clinically interpretable attention maps, and validates generalisation across three independent datasets. This dual emphasis on computational efficiency and clinical interpretability distinguishes the framework from earlier CNN–ML combinations that primarily focused only on accuracy. Experimental results show that the VGG16–SVM model achieves 97.7 % accuracy, 95.7 % sensitivity, 98.4 % specificity, and an AUC of 0.98, outperforming several state-of-the-art CNN baselines. These findings underline the potential of the framework as an efficient, interpretable, and scalable solution for large-scale glaucoma screening, particularly in resource-limited clinical environments. • Deep learning–guided glaucoma detection using VGG16 and ML classifiers. • Cross-dataset fundus analysis ensures robust generalization performance. • Attention-map visualizations enhance clinical interpretability. • ∼40 % fewer parameters versus ResNet, improving computational efficiency. • Open-source code release supports reproducibility and deployment.