Research On Retinal Vessel Segmentation Technology Based On Cascaded U-shaped Network

Retinal vessels are one of the important parts of vascular microcirculation seriously affected by cardiovascular diseases,and their structural changes can directly reflect the influence of cardiovascular diseases on the morphology and structure of vascular network.In clinic,retinal vascular network is of great significance in the diagnosis of arteriosclerosis,hypertension,diabetes,nephritis,diabetic retinopathy,glaucoma and other ocular diseases.There are a large number of small vessels in the retinal vascular network,and the connections between the vessels are close,so the structure of the retinal vascular tree is complex.In addition,the contrast between the blood vessel and the background is small,and the fundus image is easily affected by uneven illumination and pathological noise.These objective factors make retinal vessel segmentation a challenging task.This topic carries out related research work on retinal vessel segmentation methods.Based on the analysis of the characteristics of retinal images,the topic focuses on the retinal vessel segmentation technology based on cascade Unetwork,and puts forward the corresponding solutions.The main research content and specific work of this paper are reflected in the following aspects:(1)Aiming at the problems that small blood vessels in retinal images are difficult to segment and lesions are missegmented,a retinal blood vessel segmentation method based on Efficient Net-B5 cascade U-network is proposed.The Efficient Net-B5 encoder is used to encode the image features to extract deeper features of retinal vessel images.Secondly,a double convolutional attention mechanism was proposed to effectively obtain the information of cross-channel interaction.Experimental results on the CHASE_DB1 dataset show that the proposed algorithm can segment retinal vessels accurately and has good anti-interference performance at the same time.(2)In view of the existence of artifacts in the FOV edge of the retinal blood vessel segmentation image,a morphological post-processing algorithm is used to shrink the FOV edge of the retinal mask image inward,so as to optimize the false segmentation of the FOV edge of the retinal image,and finally achieve effective segmentation of the retinal blood vessel.(3)In order to further improve the effect of retinal blood vessel segmentation,ASPP_CBAM module is introduced to improve the Ladder Net network,and an improved Ladder Net retinal blood vessel segmentation algorithm is proposed.The improved network makes full use of the feature information of the multi-scale receptive field,while preserving the important features through the attention mechanism,so as to achieve a more ideal segmentation effect in retinal blood vessel segmentation.(4)Compared with FC_RCF,U-Net and Res-UNet,the experimental results show that on the CHASE_DB1 dataset,The retinal vessel segmentation algorithm based on Efficient Net-B5 cascade U-network has the best F1-score,specificity,accuracy and AUC,reaching 79.57%,98.21%,96.30% and 97.51%,respectively.The F1-score on the DRIVE dataset reaches 81.74%.Compared with FC_RCF,U-Net,Res-UNet and Ladder Net,the improved Ladder Net algorithm has the highest F1-score,AC and AUC values on the DRIVE and CHASE_DB1 datasets.On the CHASE_DB1 dataset,compared with the original Ladder Net network,the F1-score is improved by 1.17%,and the AUC is increased by 0.17%.The F1-score on the DRIVE dataset reaches 82.2%,and the AUC value reaches 98%.Therefore,it is verified that the retinal vessel segmentation network proposed in this paper has high accuracy and robustness.