Automated Glaucoma Diagnosis from Fundus ImagesUsing Deep Learning

Abstract

Glaucoma is a leading cause of irreversible blindness worldwide, often progressing without early symptoms and resulting in significant vision loss if left undiagnosed. Early detection is therefore critical, but current manual screening methods are time-consuming and require specialized expertise, limiting accessibility in many regions. This thesis presents a fully automated framework for glaucoma diagnosis using deep learning on retinal fundus images. The proposed method integrates three main phases: image enhancement with Contrast-Limited Adaptive Histogram Equalization (CLAHE), hierarchical segmentation of the optic disc and cup using a fine-tuned Segment Anything Model 2 (SAM2), and glaucoma classification based on neuroretinal rim analysis with a Vision Transformer (ViT). Experimental results on the ORIGA and REFUGE datasets demonstrate high segmentation accuracy, with Dice scores of 95.84% and 90.63% for the optic disc, and over 90% for the optic cup. For classification, the ViT model achieves an F1-score of 95.02% and a precision of 97.39% on the REFUGE dataset, outperforming several traditional convolutional neural networks. The framework is efficient, interpretable, and suitable for real-time clinical deployment. While specificity remains a challenge due to class imbalance and lack of data augmentation, the results highlight the potential of combining advanced segmentation and transformer-based classification for robust, scalable glaucoma screening. This work lays a strong foundation for future improvements and broader clinical application in automated eye disease diagnossi.