DELTA: Delving Into High-Quality Reconstruction for Electrical Impedance Tomography

Abstract

Owing to the success of neural network (NN)-based methods in solving inverse problems, recent works study their applicability in electrical impedance tomography (EIT). Deep NNs, especially convolutional NNs (CNNs), are the main modules for expressing nonlinear features compared to traditional model-based methods. However, the convolution with fixed receptions typically expresses only local features, which limits the performance of learning global features as well as multiscale information. In this article, we propose a learning-based reconstruction framework, named DELTA, that is composed of convolution operators and Transformers, which offers both high spatial resolution as well as impedance resolution in multiconductivity tasks. The core of our design is the U-shaped backbone, which is improving with the Swin Transformer (SwinT) and multidilation convolution sparse attention (MDSA). Furthermore, a parallel bottleneck composed of channel-wise attention (SE-ViT) and dynamic large kernel (DLK) is embedded to increase feature propagation. To utilize the multiscale features, dynamic feature fusion (DFF) is proposed to hybrid various information from encoder, bottleneck, and decoder. Our extensive evaluations on two benchmarks, multiphase shape reconstruction, and lung-phantom reconstruction, revealing the effectiveness of our contributions. Our DELTA achieves a structural similarity index (SSIM) of 0.9646 (multiphase shape reconstruction) and 0.9406 (lung-phantom reconstruction), as well as an improvement in computational complexity compared to state-of-the-art Transformer-based methods. The DELTA method is expected to provide a high performance in structural and functional imaging.