Electrical Impedance Tomography Image Reconstruction With Attention-Based Deep Convolutional Neural Network

Abstract

Electrical impedance tomography (EIT) is a promising functional and structural imaging method in process tomography. However, due to the ’soft-field’ nature and the high dependence on the prior information, it often suffers serious artifacts in quantitative analysis. Most recently, EIT image quality has improved significantly because of the state-of-the-art deep learning-based models in the aspect of solving the inverse problem, especially fully convolutional networks (FCN) and V-Net variants. Despite their success, these deep convolutional networks (CNNs) have two limitations: (1) The long-range information transition is frequently lost and the reverse gradient often disappears in deep CNNs; (2) Some novel skip connections, such as residual and dense connections, often occupy substantial computational resources. To overcome these two limitations, we propose V <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> A-Net, a new neural architecture based on redesigned feature transited connections by the terms of (1) A pre-reconstructor based on the iterative Newton-Raphson method, which maps the nonlinear function between the measurements and the initial images, (2) Dual cascaded V-Net are combined, which play the role of an encoder and a decoder, respectively, (3) A new parallel attention mechanism via channel attention and coordinate attention to learn the conductivity distributions and boundary-shaped feature separately, and (4) the light-weight skip connections reduce the computational resources (or accelerate the inference speed) of EIT imaging. The V <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> A-Net is evaluated by using the multi-phase flow industrial applications, and the results demonstrate that (1) V <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> A-Net has better performance in shape reconstruction with sharp ’corner’, (2) V <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> A-Net could reconstruct the model accurately where it has some low-contrast conductivity distributions, (3) V <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> A-Net enhances the quality of interfaces with the stratified flow, and (4) the pruned V <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> A-Net achieves significant speedup compared with the VDD-Net or V <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> DNet. The analyses show that the average relative error is 0.05, the average correlation coefficient is 0.92, the average structural similarity is 0.92 on the testing datasets. In addition, the average relative cover ratio is 0.97 and the average relative contrast ratio is 0.98 on the testing datasets.