A physics-embedded dual-learning imaging framework for electrical impedance tomography.

Area of Science:

• Medical Imaging
• Computational Electromagnetics
• Applied Physics

Background:

• Electrical Impedance Tomography (EIT) is a noninvasive imaging method for conductivity distribution reconstruction.
• EIT inverse problems are nonlinear and ill-posed, challenging traditional regularization techniques.
• Existing Physics-Informed Neural Networks (PINNs) struggle with sparse, noisy boundary data common in practical EIT.

Purpose of the Study:

• To develop a robust and efficient imaging framework for EIT overcoming limitations of current methods.
• To address the challenges posed by sparse, noisy boundary measurements and computational complexity in EIT.
• To propose a novel Physics-Embedded Dual-Learning Imaging Framework for enhanced EIT reconstruction.

Main Methods:

• A supervised Convolutional Neural Network (CNN) predicts internal potential distributions.
• An unsupervised Physics-Informed Neural Network (PINN) reconstructs conductivity by enforcing the governing partial differential equation (PDE).
• A decoupled dual-learning architecture is employed, reducing the need for multiple forward networks.

Main Results:

• The proposed framework improves reconstruction robustness and efficiency under realistic EIT measurement constraints.
• The decoupled architecture eliminates assumptions of smooth conductivity.
• Computational complexity is significantly reduced by requiring only one forward network.

Conclusions:

• The Physics-Embedded Dual-Learning Imaging Framework offers a promising solution for practical EIT applications.
• This novel approach enhances the performance of EIT imaging with sparse and noisy data.
• The framework demonstrates improved efficiency and robustness compared to traditional methods.