Application Of Stacked Residual Networks Combined With Multiscale Feature Fusion For Retinal Vessel Segmentation

Characterization changes of retinal blood vessels are closely related to early diagnosis of many comprehensive diseases.The shape analysis of blood vessels is the basis of disease treatment,and fundus vascular segmentation is the most important step in the diagnosis of these diseases.Therefore,accurate segmentation of retinal blood vessels in the fundus can help doctors to better diagnose the related diseases,which is important for giving appropriate treatment options in the later stage.At present,retinal blood vessel segmentation algorithm based on convolution neural network has been widely used in the field of blood vessel segmentation,but there are still the following problems: 1.The context information of multi-level semantic features of the network is not trained specifically,which can easily lead to reduced segmentation accuracy;2.The skipping connection structure of the network results in the loss of some characteristic map information,which easily leads to the phenomenon of undivided peripheral blood vessels.3.The convolution has a smaller field of actual perception,and with the deepening of the network,it is easy to lose some necessary low-level details.For the above problems,the research in this paper has the following aspects:Firstly,to solve the problem that context information of multi-level semantic features is not trained specifically,this paper presents a method of retinal vascular segmentation based on MU-Net(Multi-Scale U-Net).MU-Net uses four dilated convolutions with different expansion rates to learn the feature maps under different receptive fields,and combines them with multiscale operations for feature fusion.In this paper,four open datasets are compared,and the overall segmentation accuracy of MU-Net is 0.9581,0.9673,0.9661 and 0.9651,respectively,which is better than that of U-Net.Secondly,to solve the problem that network skip connection structure is easy to cause the loss of feature information,this paper presents a residual skip connection structure based on residual convolution to replace the original jump connection structure of MU-Net.Residual skip junction structure enhances the learning ability of features through residual structure,reduces the loss of detail information,and consequently retains more information about the characteristics of small vessels.Finally,in order to reduce the loss of low-level detail information during feature propagation,an improved method based on MU-Net is presented for retinal vascular segmentation based on MRNet(Mutli-Scale Res Net Network).MRNet replaces ordinary convolution and skip connection structures in MU-Net with stacked residual convolution and residual skip connection structures,respectively,to reduce the loss of information.In ablation experiments on four open datasets,the accuracy of MRNet was0.9717,0.9792,0.9796 and 0.9848,respectively,which was higher than that of UNet and MU-Net.Comparing with other advanced methods in the experiment,the result of MRNet segmentation is the best as a whole.