Self-Supervised Learning EIT Based on Physics Cycle Consistency with Unpaired Data

Electrical impedance tomography (EIT) is a promising imaging modality for the detection of time-varying conductivity distribution in industrial and biomedical fields. However, current deep learning-based image reconstruction methods require large-scale and diverse paired voltage-conductivity data for training. This paper presents a physics-informed cycle generator adversarial network for self-supervised image reconstruction of EIT. The network takes advantages of physical forward model and fully convolutional neural network to address the forward and inverse problems respectively. By devising the physics cycle consistency losses, the proposed method is trained without the paired conductivity distribution data. Subsequently, the fully convolutional neural network reconstructs the conductivity distribution from voltage measurement by its nonlinear mapping ability. The comparative experiments are conducted on simulated data and experimental data to demonstrate the performance of the proposed method. Both simulations and experiments demonstrate the proposed method outperforms other numerical methods.