Multilayer Graph Convolutional Feature Fusion Neural Encoding and Decoding Network for Fine Segmentation of Retinal Vessels

Fundus retinal vessels segmentation can assist doctors in the diagnosis of ophthalmic diseases and cardiovascular and cerebrovascular diseases. However, due to the complex topological structure of blood vessels and unclear boundaries, it greatly increases the difficulty of segmentation. A graph convolutional feature fusion network is proposed based on the U-shaped structure to address these issues. This network uses a graph convolution module to model the global contextual information between pixels in encoder features, making up for the lack of global modeling ability in ordinary convolutions. Then, a multi-scale feature fusion module is used to fuse the encoder features and decoder features to reduce the impact of noise information in the feature layer on the segmentation results. Finally, a multi-level feature fusion module is used to fuse and output the features of each layer of the decoder, reducing the loss of spatial information and the reuse of deep features during the downsampling process. Verified on the public datasets DRIVE, CHASEDB1, and START, the F1 values and the AUC values are better than the other two methods.