Comparative Evaluation of Transfer Learning Architectures for Glaucoma Classification

Abstract

Glaucoma is a leading cause of irreversible blindness worldwide, primarily resulting from elevated intraocular pressure (IOP) that leads to progressive optic nerve damage. Early detection is essential to prevent permanent vision loss; however, it remains challenging due to the disease’s asymptomatic nature in its initial stages. This study explores the application of deep learning in glaucoma detection, utilizing the Xception architecture to classify glaucoma from retinal fundus images. A total of 6,540 retinal images were obtained from the publicly available EyePACS-AIROGS-light-V2 dataset variant designed for glaucoma detection and systematically divided into training, validation, and testing subsets. Data augmentation techniques such as image rotation, flipping, and brightness adjustments were applied to enhance model robustness and reduce the risk of overfitting. The performance of the Xception model was benchmarked against other popular convolutional neural network architectures, including EfficientNet B7 and ResNet-50, using key evaluation metrics: accuracy, precision, sensitivity, and specificity. The Xception model outperformed the others, achieving an accuracy of 93%, a precision of 95%, a sensitivity of 91%, and a specificity of 94%. These results underscore the model’s effectiveness in identifying glaucomatous patterns within retinal images. The comparative analysis reveals that Xception achieves the best balance between accuracy and computational efficiency, owing to its depthwise separable convolutional design. Beyond quantitative results, this study provides insight into how architectural choices affect model performance in ophthalmic imaging, offering guidance for developing scalable AI-assisted screening systems.