A hybrid of iterative Gauss–Newton and one-dimensional convolutional neural network for high-resolution electrical impedance tomography

Keiya Minakawa,Keigo Ohta,Hiroaki Komatsu,Tomoko Fukuyama,Takashi Ikuno

DOI: https://doi.org/10.1063/5.0185371

IF: 1.697

2024-01-01

AIP Advances

Abstract:We developed a processing method using benefits of both iterative Gauss–Newton (IGN) and a one-dimensional convolutional neural network (1D-CNN) for high-resolution electrical impedance tomography. The proposed method logically combines conductivity images reconstructed by different methods. The accuracies of the mathematical IGN method, 1D-CNN method, and the proposed method were compared. Utilizing the ideal potential data obtained through simulations, along with the experimental potential data derived from cement samples, we reconstruct the conductivity distribution. When utilizing the simulation data, the IGN method produces larger errors in the reconstructed images as the size of the foreign object decreases. The proposed method reconstructs the position and size more accurately than the IGN and 1D-CNN methods. When utilizing the experimental data, 1D-CNN and proposed methods were more accurate in terms of the position and size than the IGN method.

materials science, multidisciplinary,physics, applied,nanoscience & nanotechnology

What problem does this paper attempt to address?

The paper aims to address the high-resolution imaging problem in Electrical Impedance Tomography (EIT). Specifically, the researchers developed a new post-processing method that combines the Iterative Gauss-Newton (IGN) method and a 1D Convolutional Neural Network (1D-CNN) to improve the spatial resolution of EIT images.

In EIT, due to the inherently ill-posed nature of its inverse problem, it is challenging to obtain accurate impedance distributions through mathematical methods. To solve this issue, the researchers proposed a logical multiplication method (AND method), which combines the conductivity images generated by IGN and 1D-CNN and performs logical operations on these images to obtain more accurate results. Through comparative analysis of simulation data and experimental data from actual cement samples, it was found that the AND method outperforms traditional IGN and 1D-CNN methods in reconstructing the position and size of small-sized foreign objects. Additionally, the study defined a standardized evaluation metric to objectively assess the accuracy of different methods in reconstructing the position and size of images.