SDEIT: Semantic-Driven Electrical Impedance Tomography

TL;DR

SDEIT couples an implicit neural representation for EIT with semantic priors from Stable Diffusion 3.5 to address ill-posed reconstruction without paired data. The framework comprises a physics-driven EIT Block for data fidelity and TV regularization, and a diffusion-based SD Block that enforces semantic and structural priors via an SSIM-based loss. By mapping image-domain guidance into the INR optimization loop through a plug-and-play regressor, SDEIT achieves sharper, more faithful conductivity reconstructions and shows robustness to noise in simulations and experiments. The approach highlights a new pathway for multimodal priors in ill-posed inverse problems and discusses avenues for domain adaptation through lightweight fine-tuning of generative models.

Abstract

Regularization methods using prior knowledge are essential in solving ill-posed inverse problems such as Electrical Impedance Tomography (EIT). However, designing effective regularization and integrating prior information into EIT remains challenging due to the complexity and variability of anatomical structures. In this work, we introduce SDEIT, a novel semantic-driven framework that integrates Stable Diffusion 3.5 into EIT, marking the first use of large-scale text-to-image generation models in EIT. SDEIT employs natural language prompts as semantic priors to guide the reconstruction process. By coupling an implicit neural representation (INR) network with a plug-and-play optimization scheme that leverages SD-generated images as generative priors, SDEIT improves structural consistency and recovers fine details. Importantly, this method does not rely on paired training datasets, increasing its adaptability to varied EIT scenarios. Extensive experiments on both simulated and experimental data demonstrate that SDEIT outperforms state-of-the-art techniques, offering superior accuracy and robustness. This work opens a new pathway for integrating multimodal priors into ill-posed inverse problems like EIT.

Figures (10)

• Figure 1: Architecture of the proposed SDEIT framework. The coordinates of FE nodes and the grid are mapped through positional encoding and fed into a four-layer MLP to estimate the conductivity: $\sigma_\text{meas}$ in the measurement domain and $\sigma_\text{grid}$ in the image domain. $\sigma_\text{meas}$ is processed in the EIT Block to compute the data loss $\mathcal{L}_{data}$ and TV loss $\mathcal{L}_{TV}$, while $\sigma_\text{grid}$ is refined in the SD Block to derive the semantic-based regularization loss $\mathcal{L}_{ssim}$. The model parameters are iteratively updated based on the total loss $\mathcal{L}$ until convergence.
• Figure 2: Simulation results comparing EIT reconstructions using TV, INR+TV, and the proposed SDEIT method.
• Figure 3: Robustness analysis of TV, INR+TV, SDEIT using a simulated thorax phantom under varying noise levels.
• Figure 4: Experimental phantom reconstructions using TV, INR+TV, and the proposed SDEIT method across four test cases. Each row corresponds to a different experimental setup, showing the ground truth (left) and reconstruction results (right three columns).
• Figure 5: Loss curve (left) and log loss (right) plot of the proposed SDEIT approach against iteration steps for experimental test cases 1-4.