Automatic Glaucoma Classification and Feature Detection for the Justraigs Challenge

Glaucoma is a vision-threatening eye disease requiring early detection. Fundus imaging is a non-invasive and cost-effective tool for glaucoma diagnosis, but manual analysis is expensive and subjective. This paper proposes a deep learning approach for automated glaucoma detection and feature analysis. We leverage a large and diverse dataset from the JustRAIGS challenge to train two Convolutional Neural Network (CNN) models: a recommendation model for classifying glaucoma presence and a justification model for identifying specific glaucomatous features. To address class imbalance, we employ data sampling and domain adaptation techniques. The justification model further benefits from fine-tuning a pre-trained referral model. Our results show that our recommendation model achieves comparable results to state-of-the-art binary glaucoma classifiers and is superior to human classification done by the final 3 graders of the challenge dataset, which averaged 86% accuracy and 92% specificity compared to the referral model’s 96,80% and 97,09% respectively. The justification model also performed well compared to other development phase submissions of the challenge.