Research On GAN Network Data Generation Method For Fundus Image Structure Segmentation

Retinal segmentation of fundus image is an important application of computer vision in the medical field.Through the segmentation of the fundus image,it is possible to effectively judge various diseases,such as diabetic retinopathy,arteriosclerosis,hypertension,glaucoma and other fundus-related diseases.In recent years,there are many deep learning network frameworks.Deep learning method instead of the ophthalmologist artificial observation method to improve the detection efficiency and reduce the artificial false detection rate has become a focus in current research.The task of retinal segmentation of fundus images is the main research content of this thesis.In view of the problem that the amount of fundus vascular image data is very deficient and difficult to label,this thesis proposes a method of automatically generating fundus vascular images using GAN network.The generator uses two parts,an encoder and a decoder.The encoder brings richer semantic information for feature extraction.In the decoder stage,deconvolution is used to recover the size of the feature map,and the feature map in the encoder is fused to repair the lost feature information.In order to realize the segmentation of fundus image structure,this thesis proposes a segmentation method based on deep learning,which is the fundus of the fundus image,such as optic disc,macular,hemorrhage and exudate.A multi-scale convolutional neural network is designed for the problem that the fundus image structure is complex and the structural details are not clearly segmented.The data of the network is input in a multi-scale manner,which increases the learning ability of the network for the details of the segmentation structure.The data of each channel is extracted by the VGG-19 convolution layer,and then the feature vector is transmitted to the multilayer perceptron.Finally outputs the segmented fundus image through the fully connected layer which has extracted feature vectors extracted from all channels.GAN greatly expands the data set and supplements the training samples,thereby improving the four-channel segmentation network model.Experiment results show that our method solves the segmentation problem of fundus image.Different models for fundus structures show good generalization ability and high accuracy rate in verification sets.The proposed framework detects blood vessels with an accuracy of 0.927,cup disc with an accuracy of 0.973,exudate with an accuracy of 0.939 and hemorrhage accuracy of 0.904,and finally a good retinal segmentation is achieved.