Unsupervised Sensitivity Prior Generation With Diffusion Model for EIT Image Reconstruction

The supervised deep learning methods for electrical impedance tomography (EIT) remains limited by hand-crafted and domain-specific instructions during training. The existing unsupervised solvers for EIT trade attention to the image domain with ignorance to physical natures in the electrical field. In this article, we propose an unsupervised conditional diffusion model, referred to as SPfusion, for inhomogeneous sensitivity prior generation, which is utilized as the system matrix to solve EIT inverse problem. The conditional diffusion model first undergoes coarse training to ensure stable generation, followed by a dedicated fine training to sharpen detailed fidelity. In SPfusion, we propose a novel dual-domain denoising network, named dual-domain sensitivity prior network, that introduces the dual-domain Cartesian-based and polar-based coordinates dual-domain Transformer to efficiently capture multiscale latent representations. In addition, the decoder withU-conditional cross-attention is proposed to fully preserve dual-domain long-range dependencies. SPfusion achieves accurate noise estimation, high-fidelity detail generation of the inhomogeneous sensitivity prior, as well as high-quality image reconstructions based on model-driven methods. Experiments conducted on simulation data, real-world measurements, and publicly available benchmark demonstrate that SPfusion outperforms supervised/unsupervised and previous diffusion-based methods in terms of reconstruction accuracy and stability. The project and further information are available online.