A Convolutional Image Reconstruction Network with Spatial Pyramid Pooling for EIT

Introduction: Electrical Impedance Tomography (EIT) offers a valuable approach for visualizing the distribution of multiple media in multiphase flows. However, image reconstruction remains challenging due to its inherent nonlinearity and ill-posedness. Methods: To address these challenges, we propose a deep learning-based convolutional reconstruction network incorporating Spatial Pyramid Pooling (SPP). This architecture utilizes a fully connected block to model the complex nonlinear relationship between boundary voltage measurements (from the EIT system) and the pixel distribution within the imaging domain. Convolutional blocks and an SPP block are integrated to enhance feature dimensionality and diversity, thereby mitigating the ill-posedness of the inverse problem. Results: Simulation and experimental results demonstrate that the proposed method significantly improves reconstruction accuracy compared to established techniques, including Tikhonov Regularization (TR), Total Variation (TV), and V-Net. Discussion: The nonlinear relationship between the measurement information obtained from the electrical impedance tomography system and the pixel information of the measured space is implemented by a fully connected block. The convolution block and the spatial pyramid pooling block increase the dimension and diversity of features, thereby improving the ill-posedness of image reconstruction. Conclusion: The proposed convolutional network with spatial pyramid pooling provides a robust and accurate solution for EIT image reconstruction in multiphase flow visualization, outperforming conventional and deep learning-based benchmarks. This approach holds promise for enhancing EIT applications requiring high-fidelity imaging.