Research Of Retinal Fundus Image Segmentation Based On Deep Learning Technology

There are a wide variety of fundus diseases,such as Diabetic Retinopathy(DR)with a high incidence and Neovascular Glaucoma(NVG),etc.Once the disease develops to a certain stage,it will cause irreversible damage to vision and even blindness.At prese nt,the diagnosis of fundus lesions is based on direct observation of fundus physiological structure or lesions by professional doctors.This process is time-consuming and laborious,at the same time,the fundus image disease is complex,the degree of various lesions is greatly inconsistent,the professional level of the doctor is higher,resulting in the fundus early screening work can not achieve comprehensive and efficient.In recent years,computer image processing technology has been used to help doctors achieve auxiliary diagnosis,greatly improving the efficiency of doctors’ work diagnosis,and at the same time improving the popularity and accuracy of fundus lesions screening.This article in view of retinal blood vessels,optic disc and cup problem whether the physiological structure of pathological changes,combined with the deep learning algorithm,using convolution Neural Network(Convolutional Neural Network,CNN)and generate against Network(Generative Adversarial Network,GAN)of retinal blood vessels,optic disc and cup segmentation,can be done in view of the fundus disease effectively early screening,to help the patient timely detection and treatment fundus disease.The main work contents and contributions are as follows:(1)Automatic segmentation of fundus retinal vessels is considered to be an important link in the early diagnosis and prevention of diseases such as diabetes and retinopathy.In Chapter3,this paper builds a new vascular segmentation network based on the U-NET model.Firstly,the deep separable convolution is used to replace the standard 3*3 convolution in the encode-decoding process,which ensures the segmentation accuracy and greatly reduces the number of parameters in the model.Secondly,the attention module is added to the traditional hop connection structure to improve the useful information in the encoder,and a feature fusion module is designed,which can maximize the long-distance dependence of the network and fuse more layers of features.Experimental results show that the proposed algorithm of this chapter is excellent,and has the potential to be applied in the early screening and diagnosis of ocular fundus vascular diseases,and provides assistance for the development of automatic vascular segmentation of lightweight fundus acquisition equipment.The results show that the algorithm in this paper can extract more detailed vascular information,and the ACC and SE values on DRIVE and STARE data sets are 0.9627 and 0.8215,and 0.9713 and 0.8477,respectively.(2)Due to the complexity of fundus images,optic disc segmentation is easily affected by blood vessels and lesions,and traditional methods cannot accurately segment optic disc.For solve the issue in effectively in the fourth chapter,we proposed a segmentation method based on eye cup deep learning network,on the basis of GAN of generator is improved to the network,the generator for added features fusion module U-Net,increase the receptive field of the model,not only increased the network with long-distance dependencies,and feel all the characteristic information of wild area can be completely covered.According to the experimental results,it can be clearly seen that the network with feature fusion module can process the edg e information better without any holes and edge missing.Secondly,a weighted loss function was designed to improve the segmentation accuracy of optic disc,and the overall GAN-based segmentation network realized automatic end-to-end segmentation.Through the quantitative analysis of the experiment,the quantitative indexes of DICE,MIOU and accuracy are 0.8210,0.9576 and 0.9845(optic disc),and 0.8157,0.9086 and 0.9873(optic cup)respectively,which realizes the end-to-end high-precision automatic segmentation based on GAN.