"Electrical Impedance Tomography"

"Electrical impedance tomography (EIT) still faces significant challenges in complex imaging scenarios involving multiple targets, diverse shapes, and varying conductivity distributions, including limited reconstruction accuracy, loss of boundary details, and the presence of severe artifacts, which greatly affect the imaging quality and stability in practical applications. To address these limitations, this paper proposes a novel imaging approach that deeply integrates physical modeling with data-driven learning, leveraging the inherent characteristics of the EIT physical model in combination with the powerful representation capability of neural networks. Specifically, a physics-informed deep learning model is developed by embedding the Jacobian matrix of the EIT model into an encoder-decoder neural network architecture, and a total loss function is formulated by combining data fidelity and physical consistency terms to enhance the network's capacity for modeling complex conductivity distributions. The proposed method is systematically evaluated on simulated datasets containing diverse shapes and conductivity values, and further validated through physical experiments. Experimental results demonstrate that the proposed method significantly outperforms conventional data-driven models in terms of reconstruction accuracy, noise robustness, and generalization capability, effectively overcoming the accuracy bottlenecks and generalization limitations associated with multi-shape and multi-conductivity EIT reconstruction, while substantially mitigating the ill-posed nature of the EIT inverse problem."