RA-OCTNET: A RELIABILITY-ORIENTED LIGHTWEIGHT ATTENTION NETWORK FOR ROBUST AND EXPLAINABLE RETINAL OCT ANALYSIS

Abstract

Retinal optical coherence tomography (OCT) is widely used for the diagnosis and monitoring of retinal disease; however, automated OCT classification models require both high discriminative performance and reliability under image-quality variation. This study proposes RA-OCTNet, a robustness-aware attention-enhanced lightweight convolutional neural network for multiclass retinal OCT disease classification. The model incorporates squeeze-and-excitation channel attention, imbalance-sensitive optimisation, corruption-based robustness assessment and Grad-CAM visual explainability. Experiments were conducted on OCTMNIST 64 × 64 images covering choroidal neovascularisation, diabetic macular oedema, drusen and normal retina. RA-OCTNet achieved an accuracy of 0.8500, macro-F1 score of 0.8466 and macro ROC-AUC of 0.9854 using only 1.23 million trainable parameters. Robustness analysis demonstrated comparatively stable performance under contrast variation and occlusion, while Gaussian noise produced the largest degradation. These findings indicate that RA-OCTNet provides a compact, interpretable and reliability-oriented framework for retinal OCT disease classification.