Disc-Cup Aware Ensemble Network (DCE-Net) for Glaucoma Detection

Abstract

Glaucoma is that serious eye disease which causes serious impairment of visual acuity and it causes permanent blindness in the advanced stages of this disease by damaging the optic nerve head. Prevention of such permanent blindness depends totally on proper detection and timely treatment of glaucoma at its early stages. To diagnose this retinal diseases, retinal fundus imaging has been used in recent years. As the careful analysis of such images is challenging, identifying the region of interest in fundus images like optic cup (OC) and optic disc (OD) is difficult because of anatomy and vascular patterns. Among all the many structural markers, the most useful in the diagnosis of glaucoma is the ratio of the OC to the OD known as the CDR. Manual testing of the fundus images is time-consuming and relies on the skill and interpretation ability of the clinician. This paper proposed an automated ensemble model DCE-Net for glaucoma detection through deep learning techniques YOLOv10 and modified DeepLabV3. The YOLOv10 object detection model was utilized for the detection of the OD and the OC, whereas the modified DeepLabV3 used for the pixel-level segmentation. The final segmented masks produced by the modified DeepLabV3 were applied to the computation of the CDR as a glaucoma indicator. This model evaluates performance on 3703 images achieved from thirteen publicly available fundus image data and obtains 98.67% accuracy and 98.84% F1 score.