Abstract

Glaucoma is a leading cause of irreversible blindness, often diagnosed too late due to subtle symptoms and the reliance on manual evaluation of retinal images. Early and accurate detection is essential for preventing vision loss, yet conventional deep learning methods face challenges in feature generalization and spatial attention. Existing convolutional neural network (CNN)-based and standard GAN approaches often underperform in preserving subtle pathological features and attention mechanisms required for robust glaucoma detection. Moreover, the lack of residual attention integration in multihead architectures limits diagnostic precision. This study proposes a novel deep learning model termed Residual Multihead Multilayer Attention GANs (RMMLA-GANs) that combines the strengths of Generative Adversarial Networks (GANs), residual learning, and multihead attention mechanisms. The generator incorporates multi-layer residual attention blocks and self-attention heads to enhance critical feature localization. A contrastive discriminator improves inter-class feature separability. The model was trained and validated using the RIM-ONE and DRISHTI-GS1 datasets. Our RMMLA-GANs model achieved superior performance over four existing hybrid approaches: Attention U-Net, Dense-GAN, ResNet- GAN, and VGG-GAN. It achieved an accuracy of 96.7%, sensitivity of 97.1%, specificity of 95.4%, AUC of 0.982, and F1-score of 96.3%, outperforming the best existing method by 3.2% in AUC and 2.8% in sensitivity.

Authors

P. Neethu Prabhakaran¹, Dini Davis², P.P. Priya³, Mini Mohan⁴, P.A. Shemitha⁵
IES College of Engineering, India^1,2,3,4, Zayed University, Dubai Campus, United Arab Emirates⁵

Keywords

Glaucoma Diagnosis, Deep Learning, Attention GANs, Residual Learning, Retinal Imaging