Automatic Optic Disc and Cup Segmentation for Glaucoma Screening

Abstract

There is a high rate of irreversible blindness due to glaucoma, making it imperative to use accurate and automated retinal image analysis to diagnose glaucoma at an early stage. In this study, a novel hybrid deep learning system for automatic segmentation of the optic disc (OD) and optic cup (OC) in retinal fundus images to facilitate computer-aided glaucoma screening is proposed. The proposed framework combines image enhancement, region-of-interest (ROI) localization, multi-scale feature extraction, residual learning, attention-guided encoder–decoder network, skip connections and adaptive feature fusion to ensure accurate segmentation in various illumination changes, occlusion of vessels and retinal anatomical structure variations. Dice Score, Intersection over Union (IoU), Precision, Recall and Specificity were used to validate the proposed model. Results of the experiments show an overall average satisfactory segmentation performance with the average Dice Score of 97.48% and Specificity of 99.10% for the segmentation of the optic disc and optic cup, respectively. Additionally, an ablation study verifies the effectiveness of the proposed architecture that further enhances the Dice Score from 94.91% of the baseline encoder–decoder system to 97.12% after incorporating multi-scale feature extraction, attention fusion, and image enhancement. The reliable estimation of the cup-to-disc ratio is achieved by accurate segmentation, which could enable early glaucoma diagnosis and offers a powerful and efficient, clinically viable and automated solution for ophthalmic screening.