A Sensitivity-Guided Unsupervised Learning Method for Image Reconstruction of Electrical Impedance Tomography

Abstract

Electrical impedance tomography (EIT) detects time-varying conductivity distribution and has grown to be a promising imaging modality in industrial and biomedical fields. However, current deep learning-based image reconstruction methods require a large number of voltage-conductivity samples for training. This article proposes a sensitivity-guided unsupervised learning method for EIT (SULEIT) image reconstruction. First, the voltage measurements are projected into voltage feature maps and a fully convolutional network (FCN) is designed to nonlinearly reconstruct the conductivity distribution images. Subsequently, the reconstructed images are converted to the measurement domain through the EIT forward modeling. Moreover, the loss function consisting of the mean absolute error and an $L_{1}$ regularization term (RT) is devised to evaluate the disparity between the measured and converted voltage measurements. By combining data-driven techniques with physical constraints, the neural network is enforced to learn the inherently nonlinear mapping from the voltage measurements to conductivity images. The proposed method enables the training of the neural network without the prior knowledge of the true conductivity distributions. Experiments show that the proposed SULEIT method obtains higher correlation coefficient (CC) values and lower root-mean-square error (RMSE) values, which demonstrate its superior imaging quality to the alternative numerical and unsupervised learning methods.