Model-Based Deep Unrolling Framework for Electrical Impedance Tomography Image Reconstruction

Due to the "soft-field" characteristics of the sensitivity field in electrical impedance tomography (EIT), the inverse problem is nonlinear and ill-posed, resulting in poor reconstructions for the boundaries objects' shape and boundary. This article proposes a two-stage learning-based method for solving inverse imaging problems (IIPs) named FISCR-Net, comprising a pre-reconstruction module and a deep unrolling imaging module, respectively. The pre-reconstruction with deconvolution and feature pyramid block transforms boundary voltage measurements into an electrical distribution matrix. The deep unrolling module is embedded, in a faster iterative shrinkage/thresholding algorithm (Faster ISTA) which is proposed and unrolled into a learning-based framework. Specifically, the iterative calculations are modeled utilizing two subnets, one is a two-branch attention block for global- and local-feature extraction, and the other is a multiscale convolution block and residual connections for feature aggregation. Additionally, the hyperparameters of the proposed method such as shrinkable thresholds and step sizes are independent of manual tuning. The experimental results in multiphase reconstruction of average RMSE, PSNR, and SSIM are 3.0976, 39.7720, and 0.9867 (compared to the FISTA-Net method, the SSIM and PSNR are increased by 2.84% and 7.72%, respectively, while the RMSE is decreased by 20.33%), respectively, which demonstrate that FISCR-Net has a significant improvement compared to the state-of-the-art methods, and exhibits strong robustness and satisfactory generalizability in complicated imaging tasks. This method promises widespread applications in industrial and medical applications.