A Multidomain Weight-Sharing Neural Structure With Dyadic Propagator for Nonlinear Impedance Imaging

Objective: Data-driven methods for nonlinear impedance imaging have severe limitations, including low generalization ability, low reliability, and high computational consumption in the data preprocessing stage. In this work, a physics-inspired multidomain weight-sharing (MDWS) method is proposed to overcome the above difficulties and accelerate the reconstruction process. Methods: We construct a multidomain neural structure by connecting two networks in the current domain and electrical field domain with a dyadic propagator, that is, a Green’s function physical layer. The neural structure is further assembled into a weight-sharing architecture based on the intrinsic nature of nonlinear impedance imaging. Full physical constraints including current, electrical field, and contrast conductivity of the electrical impedance tomography (EIT) model are used in the loss function to ensure the physical consistency of the results. Besides, data consistency layers containing data mismatch and corresponding feedback are included to further increase the accuracy and reliability. Conclusion: Tests in both synthetic and experimental validations show that the MDWS outperforms the benchmarked learning method like the dominant current learning scheme (DCS) with similar trainable parameters and traditional optimization methods like SOM. Significance: First, the strong generalization ability of the MDWS helps to reconstruct objects unseen in the training data, which is extremely important in real applications because of the lack of priori information. Second, the data mismatch is another important feature of reconstructed results and is often been ignored by previous learning-based methods. Feedback from data mitigates the mismatch stage by stage and reconstructs results with higher reliability. Third, since no iteration is needed in the data preprocessing stage of the MDWS, the reconstruction time of the proposed method is been reduced to less than 0.2 s, which is a huge leap considering the time for one iteration is about 0.8 s and is important in real-world applications like real-time medical lung imaging.