V-Shaped Dense Denoising Convolutional Neural Network for Electrical Impedance Tomography

Electrical impedance tomography (EIT) is a state-of-the-art noninvasive imaging method, which is sensitive to the dynamic changes of the lung impedance. However, the nonlinear ill-posed mathematical problem limits its spatial resolution. In order to generate reliable reconstruction image of EIT, a supervised deep convolutional neural network (CNN)—V-shaped dense denoising net (VDD-Net) is proposed in this article, which consists of four blocks sequentially, i.e., preimaging block, feature extraction block, image reconstruction block, and image denoising block. The conjugate gradient algorithm is utilized in preimaging block to map the boundary voltage to the conductivity distribution, and the postprocess block based on deep CNN is used to filter the artifacts of the initial reconstructions and obtain the sharp features in the region of interest. The images reconstructed by VDD-Net are also compared with these of Tikhonov regularization, convolutional neural network, and V-Net methods. The resolution of the reconstruction images can reach 256 $\times256$ . The reconstruction quality is quantitatively measured by relative error (RE), structural similarity indices (SSIM), distortion (DT), and widening (WD). On average, the VDD-Net achieves 0.140 RE, 0.961 SSIM, 0.167 DT, and 0.985 WD through experimental reconstructions. These results demonstrate that VDD-Net is promising in EIT for expressing the nonlinear mapping between the measurements and the parameters in the observation domain.