An Extensive Comparative Study on Early Glaucoma Classification Using Fundus Images

Abstract

Glaucoma is a chronic, progressive optic neuropathy that is among the leading causes of irreversible blindness worldwide. Because it remains largely asymptomatic until an advanced stage, a large proportion of cases are detected only after substantial and permanent vision loss. Conventional diagnostic procedures are time-consuming, resource-intensive and dependent on specialist expertise, which limits large-scale screening. Deep learning and machine learning applied to retinal fundus images have therefore emerged as promising tools for automated, low-cost and early glaucoma detection. This paper presents an extensive comparative study of recent artificial-intelligence-based glaucoma classification methods. It reviews the deep-learning architectures, classifier models, benchmark fundus datasets and performance metrics that underpin this field, and organises the surveyed literature into five methodological categories: convolutional and transfer-learning networks, vision-transformer and attention-based models, hybrid deep-learning–machine-learning and feature-fusion pipelines, classical machine learning with handcrafted features, and segmentation-assisted and multimodal frameworks. Reported accuracies range from about 76% for early single-network baselines to over 99% for ensemble, hybrid and segmentation-guided pipelines. The study consolidates the current state of the art, highlights recurring limitations such as class imbalance, limited generalisation and low sensitivity, and identifies open research directions including explainable and privacy-preserving distributed learning.