DETECTING GLAUCOMA STAGES USING DEEP LEARNING

Abstract

Glaucoma is a major cause of irreversible blindness worldwide, often progressing without noticeable symptoms until significant damage has occurred to the optic nerve. Early detection is crucial to prevent vision loss, yet manual screening of retinal fundus images requires expert ophthalmologists and is not always accessible. This project presents a Glaucoma Identification System, an automated deep learning-based solution designed to detect and classify glaucoma severity from retinal fundus images. The system employs an ensemble of UNet++ segmentation models with multiple encoder backbones, including ResNet34, InceptionV4, and DenseNet121, enhanced with SCSE attention mechanisms. These models collaboratively segment the optic disc and optic cup regions, enabling precise extraction of structural features. The segmented outputs are used to compute the Cup-to-Disc Ratio (CDR), a clinically significant indicator for glaucoma assessment. Based on CDR values, the system classifies cases into four stages: Non-Glaucoma, Mild, Moderate, and Severe. The model is trained on publicly available datasets such as REFUGE and ORIGA, using advanced preprocessing techniques and combined loss functions to improve segmentation accuracy.