Hybrid ResNet-Vision Transformer Ensemble for Early Glaucoma Detection From Fundus Images

Abstract

Glaucoma is a silent and progressive eye disease that can cause irreversible vision loss if left undiagnosed. Early detection and screening are, therefore, critical, particularly in underserved regions with limited ophthalmologic resources. This article presents a deep learning (DL)-based hybrid ensemble framework for early glaucoma detection from retinal fundus images. The proposed model integrates ResNet50 and vision transformer (ViT) architectures, enabling one to capture local spatial features and the other to learn global contextual representations. Preprocessing involves extracting the green channel and applying contrast-limited adaptive histogram equalization (CLAHE) to enhance retinal vessel visibility and texture detail. To improve generalization and mitigate class imbalance, data augmentation is applied before model training. The ensemble fuses softmax outputs from both models to produce the final classification. Experiments conducted on the Standardized Multi-Channel Dataset for Glaucoma (SMDG-19) dataset comprising 17 242 labeled fundus images demonstrate that the ensemble achieves 99.8% accuracy, F1-score = 1.0, and MCC = 0.994, confirming strong robustness to contrast variations, noise, and illumination changes.

Keywords: Contrast-limited adaptive histogram equalization (CLAHE) preprocessing, data augmentation, deep learning (DL), glaucoma detection, hybrid ensemble, Matthews correlation coefficient (MCC), ResNet50, retinal fundus images, robustness analysis, vision transformer (ViT)