HyGAT-Retina: A hybrid GNN-CNN with attention for retinal vessel segmentation and classification; development and applications in emergency medicine

Introduction Accurate segmentation and classification of retinal blood vessels in fundus images is of the greatest significance to offer early staging and emergency treatment of many ophthalmic and systemic diseases. Quick and efficient analysis of vessels is important in emergency clinical environments to minimize diagnostic uncertainty and time-to-diagnosis. The purpose of the present study is to come up with a hybrid deep learning architecture that enhances not only segmentation precision but also artery-vein classification and emphasizes clinically significant areas. Methods This study suggests HyGAT-Retina, a combined architecture, which combines convolutional neural networks (CNNs), graph neural networks (GNNs), and attention mechanisms. The CNN takes low-level hierarchical features of fundus images. All these characteristics are converted into vessel-linked graphs and trained with a GNN to simulate topology. There is an attention module which points out important vascular areas. A decoder network produces vessel segmentation maps then a classification module is used to identify arteries-veins. The DRIVE, STARE and CHASE DB1 datasets were used for the experimented. Results HyGAT-Retina had an average Dice coefficient of 0.9274, sensitivity of 0.8932, and specificity of 0.9815 of vessel segmentation, which is 2.3, 1.8 and 0.7 times higher than the current state-of-the-art methods. To classify vessels, the model had a total accuracy of 94.6, which outperforms state-of-art methods that have the highest performance by 3.2. Conclusions The suggested HyGAT-Retina model is able to reasonably represent structural and contextual vascular data, which results in high-quality segmentation and classification. Its high rates of generalization to various datasets prove its ability to make credible clinical decision support when screening and diagnosing retinal diseases.