Retinal Fundus Image Segmentation Based On U-Net And Attention Mechanism

The morphological changes of microvessels in the human retina are closely related to ocular diseases such as hypertension and cardiovascular disease,and are also closely related to retinal vessel.Retinal fundus vessels images are an important basis for doctors to diagnose various ophthalmic diseases and other related diseases.However,due to the complex morphology of retinal blood vessels and the lack of clarity of fine blood vessels,the efficiency of manual diagnosis is relatively low and prone to subjectivity.Therefore,automated segmentation methods for retinal fundus blood vessel images have broad application prospects.However,there are still shortcomings in the current mainstream segmentation methods,such as missing edge microvessels in retinal images,blood vessel breaks,and the need to improve segmentation accuracy.This paper proposes an attentionbased and side-output semantic supervision convolutional neural network for retinal vessel segmentation,which effectively segments continuous capillaries compared to mainstream methods.Moreover,the proposed network is considered as a deep high-level semantic network,and a shallow low-level semantic network and a deep-shallow semantic network fusion method are designed,which outperforms many U-Net-based methods in retinal vessel segmentation.The main research work and innovations carried out in this paper are as follows:(1)In the attention-based and side-output semantic supervision convolutional neural network for retinal vessel segmentation,we propose a novel convolution method aimed at the encoding and decoding stages,which combines residual modules to fully embed the new convolution into the feature extraction layer.This improves the ability to extract basic-level vessel features and better segment the complex structure of the vessel tree.(2)In the attention-based and side-output semantic supervision convolutional neural network for retinal vessel segmentation,we embed attention mechanisms at the bottom and skip connections of the network,allowing the spatial and channel information of the retinal microvessels to be fully captured and adjusted the weight ratio of the side outputs to assist in supervising vessel segmentation.(3)We design a deep-shallow triple-fusion supervised semantic retinal vessel segmentation network based on(1)and(2).In the side output layer of the deep-shallow dualnetwork,we embed dual multi-level semantic fusion methods to assist in supervising vessel feature extraction of the deep-shallow dual-network by fusing the weight outputs of the deep and shallow decoders respectively.We use a deep high-level semantic network to extract high-level vessel features and a shallow low-level semantic network to extract low-level vessel features.Finally,the third network optimizes the feature fusion of the deep-shallow dual-network,ensuring the full extraction of vessel features.According to the experimental results,the new retinal fundus image segmentation network proposed in this paper performs well on the DRIVE and STARE datasets.The highest accuracy,sensitivity,and F1 scores of the network on the first dataset are 97.10%,83.30%,and 83.34%,respectively.97.38%,88.04% and 83.11% on the STARE dataset,respectively.Compared with other methods,the proposed method achieved better results in some segmentation performance indicators,and the probability maps of segmented vessels had fewer fractures and better vessel continuity.