GViT-RSNet: A retinal vessel segmentation network using graph convolutional attention and multi-scale vision transformer

Addressing the challenge of segmenting fine blood vessels in retinal images, this paper proposes a Retinal vessel Segmentation Network based on Graph convolutional attention and multi-scale hierarchical Vision Transformer (GViT-RSNet). The network comprises an encoder, a decoder, and an aggregation module. Within GViT-RSNet, an efficient Up-Convolution Module (UCM) upscales the feature maps, a Graph Convolutional Attention (GCA) robustly enhances the feature maps, and a Gated Fusion Module (GFM) generates the refined segmentation outputs. We conducted fair comparison experiments on three publicly available retinal segmentation datasets, and our proposed model outperforms several other state-of-the-art (SOTA) methods. GViT-RSNet consistently achieves superior performance for the mAcc metric across all three datasets, outperforming other methods by at least 0.37%, 0.07%, and 0.43%, respectively. A series of experiments demonstrated the efficiency, robustness, and applicability of GViT-RSNet. The architecture and components of our proposed model are versatile and can be applied to other downstream medical image analysis and semantic segmentation tasks.