A DEEP LEARNING ENSEMBLE FOR REFERABLE GLAUCOMA DETECTION: AN INDEPENDENT ANALYSIS OF THE JUSTRAIGS CHALLENGE DATASET

Abstract

Fundus imaging is a non-invasive method for screening retina and is widely used for diagnosis of ophthalmological diseases. For early detection of glaucoma, clinically relevant structural changes in the optic nerve head, neuroretinal rim, and peripapillary region can be visualized in fundus images. In this study, we investigate automated referable glaucoma detection using deep learning ensemble model trained on the publicly available JustRAIGS Challenge dataset. We have proposed an independent ensemble framework combining convolutional and transformer-based architectures, specifically EfficientNet-B3 and Swin Transformer Tiny, to leverage complementary feature representations. To address class imbalance between referable and non-referable cases, generative adversarial network–based data augmentation was employed. The proposed ensemble achieved competitive performance across clinically relevant evaluation metrics, demonstrating its potential for robust and scalable glaucoma screening.

Keywords:

Glaucoma screening; Fundus imaging; Deep learning ensemble; Referable glaucoma; JustRAIGS dataset; EfficientNet; Swin Transformer; Generative adversarial networks (GANs)