Research On EIT Imaging Method Of Lung Lesions Based On Improved CGAN Networ

In modern society,the occurrence of lung cancer,pleural effusion,emphysema and other lung injury diseases has seriously affected People’s Daily life.Current methods of detecting lung disease,such as X-ray fluoroscopy and chest radiography,radiate patients,cannot be continuously monitored,and it is difficult to detect the initial changes of tissue lesions.Electrical Impedance Tomography(EIT)is a new direction in medical imaging technology.It has the advantages of non-invasive,non-radiation and visualization.It has great potential in early detection of lung diseases and dynamic monitoring of lung function and has the potential to be a long-term medical monitoring tool.Due to the nonlinear and ill-posed inverse problems of EIT,traditional lung algorithms are prone to fall into local minimum solutions in the process of solving,resulting in low resolution of reconstructed images and lack of detailed features,which cannot accurately diagnose lung diseases.In recent years,neural network algorithms have been widely used in electrical imaging field because of its powerful autonomous learning ability.By training network models with large amounts of data,the quality of EIT reconstructed images is significantly improved.Although they have been greatly improved compared with traditional EIT algorithms,they still have room for improvement,and their computational amount,feature extraction ability,generalization ability and so on also need to be improved.In view of their shortcomings,the main work of this paper is as follows:(1)According to the characteristics of EIT imaging,the traditional conditional generative adversarial network(CGAN)structure is improved.An improved selfattentional residual condition generation adversarial network(SAR-CGAN)was proposed to improve the quality of image reconstruction.The self-attention mechanism of the network is used to fully extract image features and improve the training speed of the network.The residual block structure in the network is used to establish a close connection between each transmission layer,which guarantees the transmission of information to the maximum extent and makes the network training effect more stable.(2)This article made four lung injury models based on actual training samples.Firstly,lung contours with different shapes were extracted from a large number of CT images.Then,lung injury models were established in the finite element simulation software,and lesions of different positions and sizes were added to simulate various lung diseases.Finally,input images and label images of the network were obtained through image reconstruction algorithm.(3)The network proposed in this paper is compared with the traditional CGAN and convolutional neural network(CNN).In order to meet the needs of simulation sample training model used in the experiment,the simulation parameters are set according to the experimental system.The anti-noise ability,generalization ability and effectiveness of the network in this paper are proved by using three different simulation samples,which are noisy sample,robust sample and experimental sample.(4)Finally,the method proposed in this paper is applied to the actual chest model to optimize the image reconstruction quality of the lung object and improve the feasibility of the method in the real-time monitoring of actual lung diseases.