HIHU-Net: A Hyper-Information Hybrid U-Net for Image Reconstruction with Electrical Impedance Tomography

Electrical impedance tomography is widely used for visualization of multiphase medium distributions. Due to the 'soft field' characteristic and the nonlinear nature, the reconstructions often suffer from the serious artifacts. It is a core to design a framework carefully, which could recover the sharp boundary and accurate conductivity distributions. In this research, we proposed a novel hyper-information hybrid U-Net (HIHU-Net) with 'end-to-end' operation for image reconstruction of observation domain. In this method, the strategy of multi-scale spatial feature fusion and the dense connections for multi-channel information propagation is adopted to effectively optimize the U-Net architecture and enhance the feature representation to extract more valuable information.This method could not only improve the parameter efficiency but also reduce the consumption of computing resource. In addition, the dense connection makes it easier for the forward information and the reverse gradient to propagate during the training of HIHU-Net. In order to acquire higher quality image reconstructions, a stack of small kernels with the size of 3 x 3 are utilized on the shape recovery, while a few large convolutional kernels with the size of 9 x 9 are applied to the parameter reconstruction. The training of HIHU-Net is purely implemented by simulated data. HIHU-Net was tested on the simulated data and the experimental data. The results show that the proposed method performs better than numerical algorithms and the other deep learning-based method.