Background:

Glaucoma is a leading cause of irreversible blindness globally, primarily characterized by progressive damage to the optic nerve, often associated with elevated intraocular pressure. Early detection is critical to preventing vision loss; however, traditional diagnostic methods are constrained by their dependency on specialized equipment and skilled personnel. To address these limitations, this study evaluates and compares Convontional Neural Networks (CNNs) for glaucoma detection, utilizing both multi-class and binary classification approaches. Specifically, it investigates the effectiveness of ResNet-50 and DenseNet-201 architectures in classifying retinal images. Additionally, the study assesses the interpretability of these models through Gradient-weighted Class Activation Mapping (Grad-CAM ) visualizations, providing insights into how each architecture identifies key features associated with glaucoma. By integrating advanced CNN architectures and interpretability techniques, this research aims to enhance early glaucoma detection and contribute to more accessible diagnostic methods.

Results: For binary classification, the utilized combined ResNet-50 and DenseNet-201 models achieved a precision of 1, recall of 0.92,Specificity of 1, F1 score of 0.958, and accuracy of 0.961. For multi-class classification, the models yielded a precision of 0.8889, recall of 0.8421, Specificity of 0.935, F1 score of 0.8649, and accuracy of 0.9074. Grad-CAM visualizations provided insights into the models' focus areas and decision rationale.

Conclusions: The binary classification approach demonstrated superior performance compared to the multi-class approach, indicating its potential for practical application in glaucoma detection. The use of Grad-CAM enhanced model interpretability, supporting the clinical applicability of AI-driven diagnostic tools.