A Conditional Diffusion Model for Electrical Impedance Tomography Image Reconstruction

Shuaikai Shi, Ruiyuan Kang, Panos Liatsis

Abstract: Electrical impedance tomography (EIT) is a promising noninvasive imaging technique with applications spanning from lung imaging to industrial process monitoring and tactile sensing. Despite its low cost and suitability for real-time imaging, EIT image reconstruction remains a challenge due to its ill-posed nature. State-of-the-art techniques achieved advances by leveraging spatial regularizers and the paradigm of image regression, however, they struggle with weak generalization and noise susceptibility. This research introduces CDEIT, a novel conditional diffusion model for EIT reconstruction. Unlike previous diffusion-based works, CDEIT directly conditions the probability distribution of conductivity images on the boundary voltages in an end-to-end manner. This supports implicitly learning the prior of the conductivity images from the data, thus improving reconstruction accuracy and robustness. By iteratively refining the conductivity images, CDEIT captures intricate spatial details, while effectively mitigating noise. The model backbone is a transformer-based U-net, which incorporates multiscale and windowed attention mechanisms for superior feature extraction and information fusion. Furthermore, we introduce a generalization framework based on voltage and current normalization, which supports the application of EIT image reconstruction models trained on simulated data to real-world scenarios, with varying background conductivities and excitation currents. This advancement eliminates the need for retraining and significantly enhances the practical application of EIT reconstruction models. Experiments conducted on a synthetic dataset and two real datasets demonstrate that CDEIT outperforms state-of-the-art methods. On the simulated dataset, the proposed approach improves the peak signal-to-noise ratio (PSNR) by 1.52 and 0.81 dB, compared to the state-of-the-art, two-branch networks, structure-aware hybrid-fusion learning (SA-HFL) and dual-branch U-net (DHU-Net), respectively. Moreover, it improves the structural similarity index measure (SSIM) and correlation coefficient (CC) metrics by 0.004 and 0.007, respectively, compared to DHU-net. The CDEIT software is available as open-source (https://github.com/shuaikaishi/CDEIT) for reproducibility purposes.