DeepcomplexEIT: Exploring the Image Reconstruction of Complex-Valued EIT

Complex-valued electrical impedance tomography (Cv-EIT) has insights that visualize the electrical properties (conductivities and permittivity) of various healthy and injured organizations/tissues, which is a promising technique in industrial and medical imaging. Nevertheless, most of the current research has mainly focused on the conductivity parameter, ignoring the influence of the impact on the permittivity. To address the above challenges, a novel learning-based Cv-EIT image reconstruction method is proposed, referred to as DeepcomplexEIT, which could reconstruct the distributions of conductivity and permittivity simultaneously with the multiphysics information interactions. The DeepcomplexEIT is designed to obtain high-quality complex-valued admittivity distributions by leveraging the advantages of both convolutional neural networks and Transformers. In detail: 1) the U-shaped architecture is modified using depth-separable convolution and pooling in the complex-valued domain; 2) the 2-D filter with learnable cutoff frequency is proposed for featuring the multifrequency information in the spatial and spectral domains; and 3) a novel complex-valued Vision Transformer (Cv-ViT) and cross-domain attention are designed for featuring the local–global multiscale information with the multiphysics interactions and complementation. Our extended experiments demonstrate that DeepcomplexEIT outperforms state-of-the-art (SOTA) complex-valued models in terms of the complicated shape features and multiphase distributions with respect to the admittivity. The performances are evaluated using the tank phantoms with a 16-electrode EIT system and about 67-dB signal-to-noise ratio (SNR), where the average quantitative metrics (conductivity/permittivity) are root-mean-square error (RMSE) of 1.982/0.946 and structural similarity index metric (SSIM) of 0.992/0.994 with multiphase inclusions, as well as RMSE of 2.593/2.506 and SSIM of 0.989/0.992 with lung-shaped inclusions, respectively. Overall, the DeepComplexEIT is expected to further promote the multiparameter visualization in practical applications.