Research On The Segmentation Method Of Fundus Image Optic Disc And Optic Cup Based On Convolutional Neural Network

The cup-disc ratio of fundus images is an important reference in the diagnosis and screening of glaucoma,the accuracy of its measurements is closely related to the effectiveness of diagnostic screening results,optic disc and optic cup segmentation is an important way to measure the cup-disc ratio.The traditional segmentation method of optic disc and optic cup often adopts artificial design features,which cannot obtain key feature information,have poor adaptability,and often require the processor to have certain medical expertise.All these have limited the further development of traditional optic disc and optic cup segmentation methods.In this thesis,the segmentation of optic disc and optic cup area of fundus images was studied and analyzed in detail.By using the convolutional neural network and combining with the characteristics of fundus images,a new segmentation method of optic disc and optic cup of fundus images,the residual U-Net network and the improved fully convolutional network were proposed.The segmentation experiment of optic disc and optic cup verified that the segmentation effect of these two kinds of segmentation networks was better than that of traditional optic disc and optic cup segmentation method.The main research contents of this thesis include:Firstly,data enhancement is carried out on the dataset of fundus images.On enhanced data and original data,interested in area of fundus image are extracted,after extraction of the polar coordinate transform,the image after polar coordinate transform as the training data,on the training data for U-Net network model of training,has carried on the model test on test data,test results are analyzed according to the evaluation index and evaluation.For U-Net network on the fundus image segmentation result of deficiency,proposed to the residual structure to join U-Net network,the combination of residual U-Net network,improve the learning ability of the U-Net network,after model training and testing,test result is evaluated according to the evaluation index,and through the network and U-Net,comparing the test results verify the residual U-Net network segmentation effect is better.By analyzing the structure of the fully convolutional network and combining with the characteristics of optic disc and optic cup segmentation,it is proposed to add the mean variance normalization operation and residual structure into the fully convolutional network to form an improved fully convolutional network.After model training and testing,the test results are evaluated according to evaluation indexes.By comparing with the test results of U-Net network,it is verified that the segmentation effect of the improved fully convolutional network is better than that of U-Net network.