Research On Electrical Impedance Reconstruction Algorithm Based On Deep Learning

Electrical Impedance Tomography(EIT)is a non-invasive medical imaging technology with the advantages of non-contact,non-radiation,and real-time performance.But EIT reconstruction is a seriously ill-posed,nonlinear,and ill-conditioned inverse problem,and the computational cost is very high.In order to improve the imaging quality,reduce imaging artifacts as much as possible and better preserve the shape features of graphics,this paper studies the pixel reconstruction and shape reconstruction of EIT,and designs a reconstruction algorithm based on deep learning combined with deep learning neural network.The main contents of the research are as follows:1.Use the finite element method to generate the EIT dynamic imaging dataset.Set target objects of different sizes,shapes,positions,and electrical conductivity,obtain boundary voltage values through the EIT positive problem,and generate data sets about boundary voltage values and conductivity distributions.Secondly,in order to improve the imaging effect of reconstructing the boundary shape of the object,the electric impedance shape reconstruction model of the B-spline curve is used to calculate the boundary voltage value through the EIT positive problem,and generate the boundary voltage value,control point vector and conductivity value of the target object.The EIT dataset.2.Based on EIT dynamic imaging data set,a CNN-LSTM fusion neural network model was designed,and the mapping relationship between voltage and conductivity distribution was established.After continuous training and adjustment,the optimal network model was obtained,which was used to estimate the reference voltage static imaging.3.Based on the EIT data set of B-spline curve,an Autoencoder network(Skip-Autoencoder,Skip-AE)model with two Skip structures is designed to strengthen the learning of the control point information of the graph boundary and establish information about different targets.The mapping relationship between the object boundary voltage value,the vector value of the control point,and the conductivity,after continuous training and multiple parameter adjustments,the optimal network model is obtained for the reconstruction of the current shape.The above two methods are used to perform imaging experiments on simulated voltage data and measured voltage data respectively.The test results show that the CNN-LSTM fusion neural network method designed in this paper can reconstruct the object position more accurately than the reconstruction effect of the NRM and MLP methods.Compared with the NRM and CNN methods,the shape reconstruction effect of the Skip-AE network designed in this paper has a higher imaging quality of the shape of the graph boundary,and the imaging rate has also been improved to a certain extent.