Abstract: To solve the problem of poor segmentation effect, data over-fitting, and imbalance of positive and negative samples in fundus blood vessel segmentation, a retinal blood vessel segmentation algorithm based on transformer architecture (Transformer) and multilayer perceptron (MLP) is proposed. First, data augmentation is used on training images to prevent over-fitting. Then,several transformers fused with convolution modules are used as a robust encoder to gain multi-scale feature information. Finally, a decoder consisting of MLP is adopted to complete pixel-level classification on a feature map. In addition, the combination of Tversky loss and binary cross-entropy loss is applied to solve the sample imbalance problem. Experiential results on various datasets indicate that the proposed algorithm has a good performance,which is better than other existing algorithms.

Key words: deep learning , multi-headed self-attention , multilayer perceptron , image segmentation , retinal blood vessel