TransADMM: Transformer enhanced unrolling alternating direction method of multipliers framework for electrical impedance tomography

Abstract

Electrical impedance tomography (EIT) provides an imaging modality to visualize structural and functional information simultaneously. However, the spatial and impedance resolution of reconstructions by optimization-based algorithms cannot meet the on-site application requirements due to the nonlinear and ill-posed nature of the EIT inverse problem. Moreover, the generalization for various real-world applications is also challenging based on the ‘post-processing’ ideas with convolutional neural networks (CNNs). In pursuit of an efficient and generable approach, we present TransADMM for solving the EIT inverse problem, a novel model-based deep unrolling framework that draws inspiration from the well-known alternating direction multiplier method (ADMM) improved with regularization by denoising. Specifically, each iteration step in TransADMM corresponds to a computing update of the RED-ADMM. Furthermore, a U-shaped architecture based on hybrid Transformer is proposed for implicit solving the data consistent term. Moreover, a learnable RED is designed for adaptively adjusting the penalty pattern to fit different reconstruction tasks. As a result, TransADMM is designed to learn all parameters end-to-end without manual tuning, such as regularization weights, denoising functions, iteration steps, etc. The extensive experiments are verified utilizing various tasks, and the outcomes show that TransADMM has considerable advantages over existing state-of-the-art learning-based imaging methods in terms of quantitative metrics and visual performance. It can be concluded that the TransADMM has good generalization and perturbation robustness, which promotes the EIT application in industry and medicine fields.