Statistic Shape Constrained Image Reconstruction Method For Electrical Impedance Tomography

Electrical Impedance Tomography(EIT)reconstructs the 2D/3D conductivity distribution inside the domain through the electrical measurements at the boundary of the observed domain.EIT has advantages of high temporal resolution,non-invasion,no-ionizing radiation,and simple structure,and is promising in multiphase flow measurement and medical monitoring.However,due to its "soft field",the reconstruction problem is severe non-linear and ill-posed,resulting in low resolution of reconstructed images,which limits its further promotion.Therefore,it is the focus and difficulty of EIT research to develop high precision and high robust reconstruction algorithms.In this study,the statistic shape prior is extracted and characterized by large-scale image dataset,which is used to constrain the EIT reconstruction procedure.The ill-posedness of the EIT inverse problem is reduced and the imaging precision is improved.The main work completed is as flow:(1)Taking lung EIT as an example,the human chest CT data is collected.The original images are preprocessed by segmentation,alignment,downsampling,and filling.The principal components of the data are extracted by an unsupervised learning method.The statistical shape information represented by the principal components constructs the statistical shape space of the anatomical structure of the human lung.(2)Based on the extracted statistical shape prior,a Statistic Shape Constrained Reconstruct(SSCR)framework is proposed.For the case of high real-time requirements and insignificant pulmonary structural lesions,a one-step non-iterative SCCR method(OSSCR)is proposed;For the case of low real-time requirements and significant pulmonary structural lesions,an iterative SSCR method(ISSCR)is proposed.It is proved the proposed methods have high imaging precision and are not sensitive to measurement noise and model error through simulation,phantom experiments,and vivo experiments.(3)Based on the human chest CT dataset,a statistical shape reconstruction model based on supervised learning is designed.A numerical simulation dataset is constructed for model training.A convolutional neural network is constructed for EIT image reconstruction,including a pre-reconstructor mapping measurement signal to EIT image and a post-process network based on deep residual network.The model training is completed and numerical simulation is carried out.The results prove that the proposed supervised learning method can significantly improve the quality of EIT shape reconstruction and is well robust against model errors.The phantom experiment results prove that the trained model using simulation data is successfullymigrated to experiment data and has good generalization ability.