Advancing Glaucoma Diagnosis Through Multi‐Scale Feature Extraction and Cross‐Attention Mechanisms in Optical Coherence Tomography Images

ABSTRACT Glaucoma is a major cause of irreversible vision loss, resulting from damage to the optic nerve. Hence, early diagnosis of this disease is crucial. This study utilizes optical coherence tomography (OCT) images from the “Shahroud Eye Cohort Study” dataset which has an unbalanced nature, to diagnose this disease. To address this imbalance, a novel approach is proposed, combining weighted bagging ensemble learning with deep learning models and data augmentation. Specifically, the glaucoma data is expanded sixfold using data augmentation techniques, and the normal data is stratified into five groups. Glaucoma samples were subsequently merged into each group, and independent training was performed. In addition to data balancing, the proposed method incorporates key architectural innovations, including multi‐scale feature extraction, a cross‐attention mechanism, and a Channel and Spatial Attention Module (CSAM), to improve feature extraction and focus on critical image regions. The suggested approach achieves an impressive accuracy of 98.90% with a 95% confidence interval of (96.76%, 100%) for glaucoma detection. In comparison to the earlier leading methods ConvNeXtLarge model, our method exhibits a 2.2% improvement in accuracy while using fewer parameters. These results have the potential to significantly aid ophthalmologists in early glaucoma detection, leading to more effective treatment interventions.