Electrical Impedance Tomography Image Reconstruction Based on Neural Networks

Abstract

electrical impedance tomography (EIT) is an imaging technique with a promising future. Several methods have been used for EIT image reconstruction, such as Simulated Annealing, Gauss Newton, Kalman filter and D-Bar. Recently, some authors solved this problem using artificial neural network (ANN) through pixel by pixel reconstruction, considering a fixed resolution for the final image. This work proposes a reconstruction based on the EIT forward problem. Two different meshes were considered: a coarse and a refined mesh. The latter was used to produce simulated potentials, which are the inputs for ANN training. The nodes conductivities, which used to create the outputs for training, defined in the coarser mesh. Therefore, the proposed method consists of training the ANN with inputs from a refined mesh and outputs from a coarse mesh. Two ANN architectures are proposed and compared: one based on the LeNet architecture, and another based on the feed-forward fully connected ANN. The obtained image is not dependent on any image resolution. The preliminary results show that the LeNet architecture has better performance.