DiffusionEIT: Diffusion Model for Electrical Impedance Tomography

Electrical impedance tomography (EIT) is considered to be an imaging modality that can accomplish noninvasive continuous monitoring due to its low-cost and low-injury properties. Recently, various machine learning algorithms have been proposed for EIT. However, artifacts that affect the smoothness of the reconstructed image persist in many of these algorithms due to their neural network structure which directly maps voltage to conductivity. We propose a diffusion model based EIT (DiffusionEIT) method for reconstructing smooth high-resolution images in this article. DiffusionEIT iteratively denoises the initial input Gaussian noise to reconstruct an ordered conductivity distribution image, and innovatively uses a cross-modal mechanism based on the transformer structure to fuse the voltage data into the image generation process. The efficacy of DiffusionEIT's image generation capability and cross-modal mechanism is substantiated by simulated data. The algorithm's resistance to noise is tested by voltage data with different noise levels, and finally, experiments are designed to evaluate the algorithm's usability in real-world situations. The results show that DiffusionEIT can obtain a high level of image reconstruction capability, possesses excellent noise immunity and is a new idea for EIT by generative models.