SDEIT: Semantic-Driven Electrical Impedance Tomography.

Area of Science:

• Medical Imaging
• Computational Science

Background:

• Solving ill-posed inverse problems like Electrical Impedance Tomography (EIT) requires effective regularization with prior knowledge.
• Integrating anatomical prior information into EIT is challenging due to complex and variable structures.

Purpose of the Study:

• Introduce SDEIT, a novel semantic-driven framework integrating Stable Diffusion 3.5 into EIT.
• Utilize natural language prompts as semantic priors for guiding EIT reconstruction.
• Demonstrate improved structural consistency and fine detail recovery in EIT.

Main Methods:

• Coupling an implicit neural representation (INR) network with a plug-and-play optimization scheme.
• Leveraging Stable Diffusion (SD)-generated images as generative priors.
• Employing natural language prompts for semantic guidance, avoiding paired training datasets.

Main Results:

• SDEIT significantly outperforms state-of-the-art techniques in both simulated and experimental EIT data.
• The framework demonstrates superior accuracy and robustness in image reconstruction.
• Achieved enhanced structural consistency and recovery of fine details.

Conclusions:

• SDEIT represents the first application of large-scale text-to-image generation models in EIT.
• The method offers increased adaptability to diverse EIT scenarios by not requiring paired datasets.
• Opens new avenues for integrating multimodal priors into ill-posed inverse problems like EIT.