A Compressive Learning-Based Scheme for Nonlinear Reconstructions in Electrical Impedance Tomography

Recently, the application of deep learning techniques has provided significant advances in solving nonlinear electrical impedance tomography (EIT) problems. However, when state-of-the-art performance is pushed further, these models become bigger and more difficult to use in computation-limited scenarios. In order to reduce the computational overhead and memory footprint, an EIT-specific compressive learning-based scheme (CLS) is proposed. The CLS is implemented in a two-stage strategy. First, a noniterative inverse operator, named the dominant current-based method, is introduced to map the EIT measurement data to approximate conductivity images. Second, a wavelet-based compressive convolutional neural network (CNN) with separable convolution is learned by rough image and target image pairs, where a custom loss function with mixed structural similarity (SSIM) metrics is proposed for training. Our proposed CLS has only 0.26% trainable parameters and 0.61% network size of benchmark methods, including the dominant current deep learning scheme (DC-DLS) and the deep D-bar method. Nevertheless, extensive numerical and practical experiments demonstrate that it achieves a comparable level of reconstruction quality when compared with these benchmark methods. As a learning-based approach, it also exhibits advantages over traditional iterative methods, such as the subspace optimization method (SOM). It is anticipated that the suggested CLS would encourage further compression studies and applications in EIT.