A generative approach to Electrical Impedance Tomography image reconstruction using prior information

A core challenge in Electrical Impedance Tomography (EIT) is the solution of the inverse problem. This relates to reconstruction of conductivity images from the associated voltage measurements, when a current injection pattern is applied. The success of traditional reconstruction approaches is limited due to the ill-posed nature of the problem, leading to poor performance, when it comes to fine image details. This research focuses on the use of EIT in tactile sensing. It proposes a generative adversarial network (GAN), trained using geometric shapes, to leverage prior information for improved image reconstruction. The GAN discriminator is used to provide a prior loss term for training the reconstruction network. The loss function of the reconstruction network consists of two terms, i.e., the mean squared error and the prior loss from the GAN Discriminator, respectively. Experimental results demonstrate that our approach outperforms state-of-the-art deep learning methods, achieving a mean squared error of 0.0574 and a structural similarity index of 0.2177.