Electrical Impedance Tomography was widely used in medical imaging, two-phase flow industrial inspection and special material inspection due to its non-invasive measurement characteristics, intuitive results visualization and convenient measurement methods. However, the inverse process of image reconstruction is inherently under-determined and ill-conditioned, resulting in some deviations in the results. An improved deep neural network based on Resnet34 is designed to solve the inverse problem of electrical impedance tomography for it. The training and test data sets were established, by setting the pixel point as the center in the field, the random radius and resistivity distribution change intensity in a small range, and the boundary voltage at each electrode in the case of 32 electrodes was simulated forwardly. The method was proved to be able to converge quickly, and can obtain better judgment performance compared with Gauss-Newton iteration method, total variation method and Tikhonov regularization algorithm after parameter adjustment and training.