An Artificial Intelligence Technique for Glaucoma Segmentation and Detection from Retinal Fundus Image

Glaucoma is a painless and persistent neurodegenerative disorder that occurs when intraocular pressure becomes too high in the eye. It may lead to permanent and irretrievable vision loss. The existing diagnostic techniques are generally reliable, but due to highly dependency on trained specialists, the treatment process is delayed. To address these challenges, this study offers an offline deep learning–based framework for glaucoma detection using retinal fundus images that can operate quickly, reliably, and with minimal dependence on highly specialized clinical expertise. The system begins by identifying the Region of Interest (ROI) through a brightest-spot algorithm and validates input quality using a VGG-16 model. Optic disc and optic cup structures are then segmented with a multiclass U-Net architecture, and classification are performed using Support Vector Machine, Random Forest, K-Nearest Neighbors, and a Convolutional Neural Network. Our experimental results using VGG-16 achieved 95% accuracy for input validation, and the brightest-spot method attained 92% accuracy in ROI extraction. The U-Net model shows Dice coefficients of 0.93 and 0.87 for optic disc and optic cup segmentation, respectively, and with corresponding Jaccard indices of 0.85 and 0.78. Whereas the proposed CNN outperformed among all the classifiers, attaining an outstanding accuracy of 99% and an F1-score of 99% with the automatically learned high-level features, while the other classifiers also deliver competitive results, producing accuracies around 91% and F1-scores between 90-91%. These findings ensure the robustness of the proposed framework as a reliable and efficient tool for automated glaucoma detection.