Research On Retinal Vessel Segmentation Method Based On Improved U-shaped Network

In the color retinal image,the vascular vein structure shows a radial extension from the center of the optic disc to the border of the visual field,and the category of vessels does not change after passing through the bifurcation and intersection points,but the width of the vessel endings will become narrower,the brightness will become darker,and the central light reflection will become darker or even disappear,especially in the area close to the contour of the visual field,the difference of the above characteristics will gradually become smaller and difficult to trace.For the above characteristics of the retinal vessels this paper proposes the research of retinal blood vessel segmentation method based on improved U-shaped network,designs two different network models,and carries out a large number of experiments on the public data sets DRIVE and CHASE.The experimental results show that the two network models effectively complete the segmentation task and achieve good segmentation results.The main work of this paper includes:(1)Introduced the theoretical knowledge of retinal blood vessel segmentation,including data sets,image preprocessing methods,and evaluation criteria.The principles and methods of classical models in deep learning algorithms are described in detail,and the advantages and disadvantages of the semantic level segmentation model U-Net network are emphatically analyzed.(2)An improved U-Net based retinal blood vessel segmentation method,RRDU-Net model was proposed.In view of the problem that the receptive field range of the convolution core is insufficient during the convolution operation,which leads to the inability to fully extract the vascular features,the original convolution layer is converted into a deformable convolution module,which combines branches of more scale shapes to enhance the recognition efficiency of the feature map,increase the receptive field range of the convolution core,and then improve the effect of vascular feature extraction;Aiming at the gradient vanishing problem caused by the sampling operation,the cyclic residual convolution module is introduced in the process of sampling on the network,which helps to train the deep network architecture,solve the gradient vanishing problem and avoid the influence of redundant features.The experiment is carried out on the DRIVE dataset,and the data results obtained are95.59% accuracy,97.92% specificity,and 79.63% sensitivity.The experimental data show that RRDU-Net model can better extract the features of the fundus image and achieve better segmentation performance.(3)A retinal blood vessel segmentation method based on attention mechanism is proposed.U-Net network has the problem of less information flow transmission path and fewer coders and decoders.In this paper,two U-Net networks are connected in series,and a CSU-Net network model is proposed.The channel attention module and the selective convolution core module are introduced to enhance the channel information connection between blood vessels,generate dynamic convolution cores,redistribute the weight of shared parameters and convolution cores,and achieve the purpose of enhancing blood vessel segmentation.The accuracy,sensitivity and specificity were 95.61%,78.55% and 98.10% respectively after training on the DRIVE dataset;The accuracy,sensitivity and specificity obtained on CHASE dataset are96.56%,79.67% and 98.06%,respectively.The experimental results show that this algorithm can effectively segment retinal vessels.