An Efficient Glaucoma Classification for Fundus Images by Pre-trained Deep CNNs and Accuracy-Weighted Voting Classifier Fusion

Glaucoma, a chronic eye disease characterized by progressive optic nerve damage, remains a significant public health concern worldwide. Early and accurate diagnosis is crucial for timely intervention and preventing irreversible vision loss. In this paper, developed an approach for glaucoma classification utilizing pre-trained convolutional neural networks (CNNs), specifically ResNet-101, ResNet-18, and DarkNet-19. Also employed accuracy-based weighted voting as a classifier fusion technique. Leveraging the capabilities of these CNN architectures, extract high-level features from retinal fundus images and apply transfer learning to adapt the networks for glaucoma classification. Leveraging the capabilities of these CNN architectures, extract high-level features from retinal fundus images and apply transfer learning to adapt the networks for glaucoma classification. To enhance classification performance, a fusion strategy based on accuracy-weighted voting is introduced, where the contribution of each classifier is weighted by its individual accuracy on the validation set. Experimental evaluations conducted on three glaucoma datasets—ACRIMA, RIMONE-v2, and Drishti-GS—using an 8:2 ratio for training and testing yielded promising results: 99.3% accuracy for the ACRIMA dataset, 95.6% accuracy for the RIMONE-v2 dataset, and 90% accuracy for the Drishti-GS dataset. These results confirm the efficiency of the approach proposed in the current study, particularly with regard to the accuracy of glaucoma diagnosis across different datasets. By enabling intervention after early detection, this approach could have a significant impact on glaucoma management, which in turn improves patient outcomes.