Research On Retinal Vessel Segmentation Algorithm Based On Generative Adversarial Network

According to the authoritative medical survey,fundus retinopathy is one of the many eye diseases causing blindness,diabetic retinopathy,macular disease and other fundus diseases greatly harm the health of the public eyes.Early detection and treatment can prevent the deterioration of ocular fundus diseases.The clinician can manually label the collected retinal vessel images and diagnose the fundus diseases by analyzing the morphological changes of retinal vessel.However,due to the imbalance of the number of doctors and patients,the large amount of fundus medical image data,and the color retinal vessel image itself,the contrast of the capillaries at peripheral vessel is low,and the vessel boundary is blurred,causing great difficulties for the task of artificial segmentation of retinal vessels.In order to reduce the pressure of human retinal blood vessels segmentation and unstable factors,at this stage appeared a lot of automation of retinal blood vessels segmentation method,automatic segmentation becomes less human intervention,can accurate and convenient processing large numbers of medical image can be better assisted the doctor’s diagnosis work.But traditional automatic segmentation method in capillary and keeping blood vessels segmentation connectivity aspect also has certain limitation,fundus retinal blood vessels automatic segmentation is still is a hot research topic in medical image processing automation.In this paper,the segmentation of retinal blood vessel images is taken as the foothold and color fundus retinal images are taken as the research object.The automatic segmentation of fundus retinal images is studied based on deep learning algorithm,and the training test and evaluation are conducted on two authoritative fundus image datasets,DRIVE and STARE.The main work of this paper is as follows:(1)Aiming at the the difficulty in segmentation of capillaries and poor blood vessel connectivity in traditional convolutional neural networks in retinal vessel segmentation tasks,a retinal vessel image segmentation model SUD-GAN based on deep convolutional condition generation adversarial network is proposed.The SUD-GAN generator is based on the U-shaped symmetrical convolution structure,which segment the retinal blood vessel image end-to-end.Add residual short connections in each convolution module to avoid the gradient dispersion problem caused by the deepening of the convolutional neural network,and improve the stability of the training of the generative model.The SUD-GAN discriminantor is based on a deep convolutional network.Dense connections are added to the convolution block to form the Densen block module,which builds a deep convolution-intensive strong discriminant network structure,strengthens the propagation of shallow features to the deep network,and enhances the discriminant network pair The discriminative ability of generating samples enables adversarial training to better guide the selection of features.Experimental results show that the SUD-GAN model can pay attention to the target blood vessel area,better weigh the global information and local information of the image,so that the segmented blood vessels maintain better connectivity.(2)Aiming at the problem of the high complexity of the discriminantor in SUD-GAN and the high computational cost,a retinal blood vessel image segmentation model based on image patch discrimination is proposed: SUP-GAN.The generator of SUP-GAN is consistent with SUD-GAN,and adopts a U-shaped symmetric convolution network structure with short residual links to realize end-to-end segmentation of images.SUP-GAN’s discriminantor draws on Patch GAN’s ideas,maps the entire input image into a matrix array,and obtains the output of the discriminant model by averaging all the responses of the entire array,so that the discriminant model pays more attention to the features of local image patch,and integrates the local information to make The final judgment,so as to better supervise the learning of the detailed features of the generator.The overall network has fewer network parameters,which can reduce the burden of computer operation.Comparative experiments show that the SUP-GAN model has better accuracy and sensitivity in the retinal blood vessel segmentation task.While the segmentation performance is close to that of SUD-GAN,the SUP-GAN discriminant model has lower complexity and better real-time performance.