Automated Glaucoma Detection and Classification from Large-Scale Fundus Image Dataset Using YOLOv8 and CNN

Abstract

Glaucoma is a major eye condition that slowly damages the optic nerve and remains one of the top causes of permanent blindness around the world. This study presents an automated framework for early detection and classification of glaucoma using artificial intelligence techniques applied to large-scale retinal fundus image dataset of over 17,000 images. The optic disc (OD) and optic cup (OC) were localized using YOLOv8. Following this, we conducted Region of Interest (ROI) extraction and contour masking to isolate the OD and highlight critical regions for further examination. We extracted essential features, such as the Cup-to-Disc Ratio (CDR), Vertical CDR (VCDR), neuroretinal rim (NRR) thinning, and compliance with the ISNT (Inferior > Superior > Nasal > Temporal) rule, resulting in a detailed tabular dataset. For classification, we applied ML and DL models. YOLOv8 demonstrated superior detection precision and CNN led the classification models with 87.13% accuracy. The proposed method offers a reliable, automated solution that can support large-scale glaucoma screening in clinical settings. This framework has the potential to assist ophthalmologists by improving the speed and accuracy of early glaucoma diagnosis, reducing the risk of vision impairment in affected patients.