SEC-UNet: Squeeze-and-Excitation Vision Transformer and Cross Attention Enhanced UNet for Electrical Impedance Tomography

Electrical Impedance Tomography (EIT), recognized for its visualization and non-invasive characteristics, holds broad application prospects in industrial and biomedical fields. However, its inherent ill-posedness and nonlinearity in the inverse problem led to low spatial resolution, boundary distortion, and inaccurate conductivity representation in current imaging methods. This paper introduces an enhanced U-shaped deep imaging approach, referred to SEC-UNet, which optimizes skip connections using SE attention and Vision Transformer (ViT) to build global information features. Moreover, a cross-attention module is introduced for fusing multi-scale information between encoder and decoder. The experiments show that SEC-UNet achieves a Root Mean Square Error (RMSE) of 2.5067 and a Structural Similarity Index (SSIM) of 0.9282, improving RMSE by 31.01% and SSIM by 1.25% over state-of-the-art Transformer-based and ConvNets-based methods. The visualizations demonstrate better boundary preservation and consistency with actual distributions. The results confirm the enhanced robustness and generalizability, indicating the superiority of the Transformer framework in soft-field imaging over ConvNets.