Publication

Abstract : Automated retinal disease diagnosis has emerged as a pivotal innovation in modern ophthalmology, promising to enhance clinical workflows and patient management. While existing diagnostic systems show potential, they often lack the precision, generalizability, and computational efficiency needed for real-world deployment. This study aims to overcome these limitations by developing an advanced deep learning framework for robust and efficient retinal disease classification. The proposed methodology integrates three key components: a Transformer-based VesselNet (T-VesselNet) for enhanced feature extraction, ChannelNet for optimized information flow, and a Graph-Enhanced Multi-Scale Fusion Network (GEMFNet) to capture hierarchical retinal patterns. Furthermore, a novel Neuroweb Optimizer dynamically adjusts hyperparameters—including filter size, kernel dimensions, and layer-wise neuron allocation—to maximize model performance. When evaluated on four benchmark datasets (DRIVE, STARE, CHASE_DB1, and HRF), the model achieves state-of-the-art performance, with accuracy (98.7 %), sensitivity (96.2 %), specificity (99.1 %), and Dice coefficient (97.4 %). The system’s real-time segmentation capability (0.23 s per image) demonstrates strong clinical applicability. These findings establish the framework as a transformative solution that bridges the gap between algorithmic innovation and clinical implementation. By enabling earlier, more reliable disease detection, this research not only improves diagnostic outcomes but also sets a new benchmark for AI-assisted ophthalmology practice.