A Deep-nonlinear Based Image Reconstruction Method For Electrical Impedance Tomography

Electrical impedance tomography(EIT)is a novel imaging technique that developed gradually in recent years.Considering the promising advantages,such as the low consumption,non-invasion and portability,EIT has been a broad application prospect in fields of biomedical engineering,multi-phase flow analysis,structure non-destructive testing,etc.However,due to the‘soft-field’nature of electrical field of measurement domain,the image reconstruction of EIT suffers from the serious nonlinearity and ill-posedness.On the other hand,the number of sensors(or electrodes)attached on the boundary is less than that of the pixels of image domain.Because of the characteristics of nonlinear,ill-posed and ill-conditional,EIT images often suffer from the low resolution and serious artifacts.It is a core challenge to design some algorithms with strong robustness and high-resolution.Recently,with the wide application of deep learning,deep neural networks(DNNs)has been succeeded in most inverse problems.Inspired by these cases,DNNs and deep convolutional neural networks(DCNNs)are introduced as a basic idea which utilize to reconstruct the EIT images.The main works are listed as follows:(1)Based on the DNNs model,the multi-layer hieratical stack auto-encoder(MHAE)is proposed as a data-driven image reconstruction method.MHAE utilizes a number of stacked autoencoders with hieratical structure,and the self-supervised and the supervised mode are also used to train the MHAE network.These two strategies could realize the high-level feature reused and increase the efficiency of the parameter-optimizing.The numerical results are carried out in the simulated test data and experimental data based on digital acquisition system(DAS),MHAE has strong anti-noise ability and better generalization performance.The reconstructions could show the medium distribution with accurate parameter and clear boundary.(2)Based on the DCNNs,the dual V-shaped dense connected convolutional neural network(V~2D-Net)is proposed to realize a shape-reconstruction scheme to reconstruct EIT image.There are two stage in V~2D-Net.Firstly,the nonlinear function is mapped between the measurement voltages and initial electrical parameters in pre-reconstructor.After that,the deep CNNs with dense connection is utilized to post-process the noise and artifacts in the initial reconstructions and the EIT images with accurate conductivity and sharp boundary are obtained.In addition,the dense connections are added in the deep CNNs,which increase the forward information flow and reverse gradient flow to alleviate the gradient-vanishing problems and make the deep CNNs training easier.The reconstructions demonstrate the stronger noise-robustness,better generalization and high-resolution performance.(3)The conception of voltage feature matrix is proposed in the research.It transfers the measurement sequence to a matrix with spatial information.By this data expressions,the prior information with spatial information of observation domain,the distribution of sensors(electrodes)as well as the data collection mode are introduced indirectly.The voltage feature matrix has some observations as follows:(a)supplement the lacked information of sensitivity matrix which calculated by the linear approximated,(b)reducing the training difficulty in the‘end-to-end’manner,(c)avoiding the over-fitting problem in the deep CNNs.Moreover,the voltage feature matrix expresses the characteristics both measurement value and data distribution.It reduces the ineffective learning when the measured perturbations superimpose the signals.What issues we should notice is the V~2D-Net could be applied the experimental data directly which it is purely trained by the simulated training samples.The successfully migration proves the good generalization ability of it.