Research on Reference Voltage Prediction for Electrical Impedance Tomography Based on Fully Connected Neural Network

Electrical impedance tomography (EIT) is a visualization techniquetoreconstruct conductivity distribution variations that reflect pathological changes in human tissues based on the boundary voltage measurement. Difference imaging is commonly used in the reconstruction to reduce modeling errors. Cerebral hemorrhage or ischemia can cause concentration changes of the intracranial ions, affecting the conductivity distribution. Consequently, the reference voltage obtained at a specific instant is inaccurate in the difference imaging. This paper proposesa reference voltage prediction method for brain EIT by fully connecting a neural network (FCNN). The reference voltage can be accurately predicted by establishing a nonlinear mapping between the measured and reference voltages. Firstly, a three-layer brain model is constructed, including the scalp, skull, and brain tissue layers. The measured boundary voltage is used to construct the input matrix, and the true reference voltage is applied to construct the output matrix in the network. Anumber of training datasets are established to train the network. During the back-propagation of the loss function, an adaptive moment estimation algorithm is employed to update the parameters of FCNN. Then, the nonlinear relationship between the boundary measurement and the true reference voltage can be acquired, and the reference voltage can be predicted. Simulation and experiments validate the proposed method. Compared with the true reference voltage, simulation results show that the voltage relative error ranges from 0% to 0.10% under the noise-free condition and 0% to 0.15% under the noisy condition. The reference voltage predicted by the proposed method well approaches the true reference voltage. Image reconstruction is performed based on the predicted reference voltage. The results show that the simulated stroke in the brain tissue layer can be reconstructed. The average blur radius of the reconstructed image increases, and the average correlation coefficient decreases gradually when the signal-to-noise ratio decreases. The feasibility of the proposed method is also tested when the conductivity of the scalp layer, skull layer, and brain tissue layer changes. It is found that the reconstructed image is very similar to the true conductivity distribution. The phantom experiment also validates the excellent performance of the proposed method. The following conclusions can be drawn. (1) Due to the powerful mapping ability of FCNN, the proposed method can establish the nonlinear relationship between the measured boundary voltage and the true reference voltage in the brain EIT. (2) The difference between the predicted and true reference voltage is minor. The conductivity distribution of different models can be well reconstructed using the predicted reference voltage in the image reconstruction. (3) The proposed method only requires boundary measurement to obtain the information of reference voltage, avoiding the reference voltage calibration problem. © 2024 China Machine Press. All rights reserved.