Research On Segmentation Of Pulmonary Fissure Image Method Based On Generative Adversarial Network

Before the test and diagnosis of most pulmonary diseases,doctors need information about pulmonary fissures obtained through medical imaging technology for pulmonary function assessment.Manual segmentation of pulmonary fissure CT images is time consuming and laborious.Therefore,an aut omatic segmentation method is needed clinically.Most conventional medical image segmentation algorithms need to manually design the features of pulmonary fissure based on the prior knowledge of anatomy.Due to different pulmonary fissure structures,there are many complex shapes such as deformation,fracture and attached fissure in CT images,and easy adhesion with lung organs,blood vessels and other tissues,these phenomena greatly increase the difficulty of pulmonary fissure segmentation.The flexible frame structure of convolutional neural network can not only automatically distinguish the pulmonary fissure and non-pulmonary fissure regions,learn various linear information in the CT images of pulmonary fissure,but also automatically optimize the parameters in the network model,avoiding the limitations of manual design features.This article takes the CT image of pulmonary fissure as the research object,with the purpose of reducing the false positive rate,false negative rate and improving the segmentation accuracy as the basis,studies the segmentation of pulmonary fissure in the Glucold database based on U-Net and the generation of adversarial networks,and three different segmentation networks were realized.The main work of this paper is summarized as follows:(1)Data preprocessing.Firstly,the original data format was transformed to extract the pulmonary slice data.Enrich the training data through data enhancement to improve the generalization ability of the model.(2)Network structure design.In order to optimize the U-Net network performance and calculation speed,the Res-U-Net model is designed.The residual module introduced is more conducive to the extraction of deep features,and the Dropout layer is added to solve the problem of overfitting.(3)A pulmonary fission segmentation model based on generative adversarial network is proposed.Based on the Res-U-Net network and combined with the advantages of generative adversarial network in image generation task,a structure of pulmonary fissure segmentation network based on generative adversarial network is designed to make the network extract deep features better.In this network,Res--U-Net is used as the segmentation network,and the discriminator network is introduced to optimize the parameters of the s egmentation network.The two networks are trained alternately against each other to improve the segmentation performance.(4)Loss function and discriminator network are improved.In order to reduce the blur level of image and the loss of semantic information,the loss function in the original GAN network is combined with the L1 distance to improve the loss function.In addition,in order to achieve more accurate segmentation,this paper designs two discriminator networks,image and Markov discriminator(Patch GAN).The comparison between the segmentation results of Mevislab 3D visualization and evaluation index shows that patch GAN model has better performance.