Enhancing Multiple-Ocular Disease Using Advance Attention Mechanisms in Deep Learning

Abstract

The accurate classification of multiple ocular diseases remains a crucial challenge in medical imaging, particularly due to overlapping visual features and limited annotated data. This study proposes a deep learning framework that integrates advanced attention mechanisms to enhance the discrimination of ocular disease features from fundus images. Leveraging a hybrid deep neural network architecture consisting of 82 layers, the model introduces a dual-channel attention module that captures both global and local contexts to improve class-specific feature learning. The system was trained and evaluated on a multi-class dataset comprising eight ocular diseases: glaucoma, cataract, diabetic retinopathy, age-related macular degeneration (AMD), retinal vein occlusion, hypertensive retinopathy, optic neuritis, and myopia. Experimental results demonstrate that our model achieves a remarkable classification accuracy of 97%, significantly outperforming baseline CNNs and traditional transfer learning approaches. Furthermore, the model requires only 18.17 minutes of training time on a high-performance GPU environment, indicating its efficiency and suitability for clinical integration. The attention modules were instrumental in boosting sensitivity for minority classes and reducing false positives. The study confirms that advanced attention-driven architectures are critical in elevating the diagnostic capabilities of deep learning models in multi-class ocular disease detection tasks, providing a valuable tool for ophthalmic healthcare.