Research On Deep Learning Electrical Impedance Tomography Method Based On Error Constraint

As a functional imaging technology in medical imaging,electrical impedance tomography(EIT)acquires the conductivity distribution of human tissues and organs by injecting excitation current into the body surface by setting electrodes,collecting boundary voltage,and processing the collected boundary voltage with appropriate algorithm.Due to the advantages of noninvasive,real time,and visualization,EIT has broad application prospects in the fields of clinical monitoring.The inverse problem of EIT is highly ill-posed and nonlinear.Since the nonlinear fitting ability of traditional algorithms is poor,reconstruction images are blurry and detailed features are lacked.In recent years,deep learning algorithms have made major breakthroughs in solving imaging problems.Compared with analytical methods,applying deep learning networks to image reconstruction of EIT,higher resolution images can be obtained.Although a lot of data and time are needed to build a network model,this process can be completed before formal reconstruction.However,the neural network is hard to interpret,and the robustness and generalization ability of the network model cannot be guaranteed.Therefore,a deep learning scheme based on error-constraints is proposed in this paper to improve the robustness and generalization ability of EIT image reconstruction.The main works are as follows:(1)An Error-constrained Network(Ec-Net)is designed in this paper for EIT image reconstruction,the robustness of the network is improved by establishing the residual mapping between the initial conductivity image and the reconstructed image error.In order to undo image error,Ec-Net utilizes feature fusion module,dilated convolution blocks and the structure of residual in residual to extract the multi-scale features of the conductivity images for image reconstruction,and EIT images with better detail and boundary preservation can be obtained.(2)Human lung conductivity distribution is taken as the research object in this paper,and a simulation database is made for training Ec-Net.The database is composed of numerical simulation samples and chest simulation samples.The numerical samples are made to simulate lung tissue,square and triangle samples are added to improve the generalization ability and boundary preservation of the network on the basis of the traditional round samples.In order to meet the needs for the reconstruction of lung conductivity distribution,multiple conductivity distribution samples were added.Considering the specificity of human chest geometry,chest simulation samples are constructed for the imaging of lung ventilation and lung injury.(3)Both simulation and experiment are conducted to verify the effectiveness of Ec-Net.The reconstruction results show that Ec-Net can accurately reconstruct irregular inclusion boundaries and conductivity differences between inclusions while showing good anti-noise ability for measurement noise higher than 30 d B.Even effective reconstruction can be achieved for the new conductivity distribution(inclusions with size/shape variations),that means Ec-Net has good generalization ability.