A Physics-Embedded Dual-Learning Imaging Framework for Electrical Impedance Tomography

TL;DR

This paper introduces a dual-learning framework combining CNN and PINN models to improve electrical impedance tomography, enabling more robust and efficient conductivity reconstruction from sparse, noisy boundary data.

Contribution

It proposes a decoupled dual-learning approach that reduces computational complexity and enhances robustness in EIT imaging under realistic measurement conditions.

Findings

Reduces the number of forward networks from K to 1.

Improves robustness of conductivity reconstruction.

Enhances efficiency under noisy, sparse data conditions.

Abstract

Electrical Impedance Tomography (EIT) is a promising noninvasive imaging technique that reconstructs the spatial conductivity distribution from boundary voltage measurements. However, it poses a highly nonlinear and ill-posed inverse problem. Traditional regularization-based methods are sensitive to noise and often produce significant artifacts. Physics-Embedded learning frameworks, particularly Physics-Informed Neural Networks (PINNs), have shown success in solving such inverse problems under ideal conditions with abundant internal data. Yet in practical EIT applications, only sparse and noisy boundary measurements are available. Moreover, changing boundary excitations require the simultaneous training of multiple forward networks and one inverse network, which significantly increases computational complexity and hampers convergence. To overcome these limitations, we propose a Physics-Embedded Dual-Learning Imaging Framework for EIT. The dual-learning strategy is composed of a supervised CNN-based forward network, which learns to predict a discrete internal potential distribution under fixed Neumann-to-Dirichlet boundary conditions, and an unsupervised PINN-based inverse network, which reconstructs the conductivity by enforcing the governing PDE through discrete numerical differentiation of the predicted potentials. This decoupled architecture removes the need for smooth conductivity assumptions, reduces the number of forward networks required from $K$ to 1, and improves reconstruction robustness and efficiency under realistic measurement constraints.(https://github.com/XuanxuanYang/CNN-PINNframework.git)