As a noninvasive medical imaging technique, electrical impedance tomography (EIT) reconstructs the internal tissue conductivity distribution by measuring the surface electrical signals. EIT reconstruction is a highly ill-posed, nonlinear problem, and the results of the existing studies suffer from blurred reconstructed images, artifacts, and the presence of adhesion in adjacent tissues. In this paper, we propose DUMD-EIT, a three-stage EIT image reconstruction algorithm with a transposed convolution-U-network-masked diffusion model. The first stage utilizes the transposed convolution to improve the resolution, followed by the second stage of reconstruction using the U-network that introduces the dilated convolution. Finally, the third stage uses the masked diffusion model. The experimental results show that the first two stages of the reconstruction algorithm can effectively improve the blurring and artifacts of the reconstructed image, and the masked diffusion model in the third stage further improves the edge sharpness of the reconstructed image and alleviates the artifacts and other problems. Furthermore, this paper attempts to normalize the difference between the measured voltage and the reference voltage before fusing and splicing them into matrix data to be used as input for the first stage of reconstruction. This approach effectively addresses the issue of the background solution conductivity affecting the accuracy of the reconstruction results. The proposed three-stage Deconvolution-UNet-Masked Diffusion Model for electrical impedance tomography image reconstruction is available on: https://github.com/188898/DUMD-EIT .