A Sensitivity Prior Generation With Conditional Diffusion Model for EIT Reconstruction

Unsupervised diffusion models excel in image generation and can solve EIT inverse problem, while current methods ignore nonlinear priors including inhomogeneous distributions and require numerous iterations demanded long imaging time. To address this, we propose DiffEIT, a condition-guided diffusion model combining DDPM in sensitivity-domain with three-branch diffusion frameworks in homogeneous, inhomogeneous, and differential sensitive priors. The measurement-condition guided strategy enhances the sensitivity matrix with different measurement states. Moreover, an improved U-Net with dense connections, pyramid features and cross-attention are introduced to improve the vanilla U-Net in traditional DDPM. The reconstructions are obtained utilizing FISTA-Net, where the results excel the existing EIT imaging methods. Additionally, our DiffEIT achieves accurate EIT reconstruction in only 25 steps.