RCRC: A Deep Neural Network for Dynamic Image Reconstruction of Electrical Impedance Tomography

A deep neural network is proposed for solving the dynamic image reconstruction problems in electrical impedance tomography (EIT), which can realize the filtering, smoothing, and prediction of the dynamic conductivity reconstruction. This framework includes a reconstruction network, convolutional neural network (CNN) encoder, recurrent neural network (RNN) model, and CNN decoder, thus is termed by RCRC. The RCRC can automatically learn prior spatial–temporal information from the voltage-to-conductivity training dataset and utilize it to enhance the conductivity reconstruction accuracy. Circular acceleration and pendulum systems are simulated with a water tank model. Stochastic data interpolation and dynamic data synthesis methods were proposed to generate large-scale dynamic dataset from a small-scale static dataset. The experimental results show that RCRC can accurately recover dynamic conductivity images from EIT noisy voltage sequence. Long-term conductivity prediction was also achieved by using the proposed network.

I. Introduction

Electrical impedance tomography (EIT) is a noninvasive imaging technique that uses electrical measurements at the boundary of an observation domain to calculate the conductivity distribution. Because of its advantage of high speed and low cost, EIT has been widely used in the fields of multiphase flow monitoring [1], [2], geophysical subsurface imaging [3], and physiological/anatomical measurement and monitoring [4], [5]. Unlike computed tomography (CT) or ultrasound tomography (UT), the sensitivity field of EIT strongly depends on the target conductivity distribution [6]. This kind of soft field characteristic makes EIT suffer from ill-posedness and nonlinearity and limits its spatial resolution.