Shape Reconstruction With Multiphase Conductivity for Electrical Impedance Tomography Using Improved Convolutional Neural Network Method

Abstract

Image reconstruction of Electrical Impedance Tomography (EIT) is a highly nonlinear ill-posed inverse problem, which is sensitive to the measurement noise and model errors. An improved Convolutional Neural Network (CNN) method is proposed for the EIT lung imaging. The proposed method is optimized based on the Visual Geometry Group (VGG) model, adding the batch normalization (BN) layer, ELU activation function, global average pooling (GAP) layer, and radial basis function (RBF) neural network. These optimizations help speed up network convergence, and improve reconstruction accuracy and robustness. Nearly 10 thousand EIT simulation models generated from chest CT images of 60 patients are used for the network training. The chest deformation, lung hyperdilation and atelectasis are randomly simulated during the model generation process. The proposed method after training is tested through a series of simulation data and experimental models. The reconstruction quality is quantitatively compared by calculating the root mean square error (RMSE) and image correlation coefficient (ICC). On average, the proposed method achieves 0.082 RMSE and 0.892 ICC through experimental results. The proposed method achieves high-resolution and robust shape reconstructions with multiphase conductivity for EIT lung imaging, especially in the presence of the measurement noise and interference. The proposed method is promising in providing quantitative images for potential clinical applications, such as human thorax imaging.