Artifact Removal Method for Electrical Impedance Tomography Utilizing a Multi-Scale Attention Mechanism

Electrical Impedance Tomography (EIT) image reconstruction frequently encounters significant artifacts due to the inherent high nonlinearity and ill-posed nature of the inverse problem. In this study, we propose an Efficient Multi-Scale Attention Convolutional Neural Network-Transformer (ECNN-Transformer) method specifically designed for EIT lung imaging. Building upon the Visual Geometry Group (VGG) model, we adjust both the depth of the network and the step size of the pooling layer. Additionally, we incorporate an Efficient Multi-Scale Attention Mechanism Module (EMA) along with a Transformer neural network. These enhancements contribute to creating lightweight network models that improve reconstruction accuracy and robustness while effectively mitigating artifacts. The proposed network was trained using 28000 EIT simulation models generated from chest CT images of patients. During model generation, various conditions such as lung tumors, lung hyperexpansion, and lung collapse were randomly simulated. The efficacy of our method was evaluated through a series of simulation data and static experiments. To quantitatively assess reconstruction quality, we calculated metrics including root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and the structural similarity index (SSIM). Experimental results demonstrate that our proposed method achieves an average RMSE of 0.024, a PSNR value of 21.16, and an SSIM exceeding 0.893. This approach facilitates high-resolution reconstruction in EIT lung imaging while accurately restoring conductivity characteristics associated with pathological tissues—particularly in scenarios involving measurement noise. We anticipate that this methodology will yield reliable interference-free images suitable for potential clinical applications such as imaging human lung tumors.