Generative Adversarial Networks Methods for Electrical Impedance Tomography Reconstruction and Post-processing

Abstract

Electrical Impedance Tomography (EIT) is a versatile method applied for imaging in fields such as phase flow analysis, medical diagnostics, and sensing materials imaging. It provides real-time, non-invasive cross-sectional imaging, aiding in pipeline optimization, energy reduction, and environmental impact mitigation. For medical purposes, EIT has been integrated with conductive fabrics for artificial skin, enabling simultaneous monitoring of electrical conductivity and mechanical properties. This thesis proposes Generative Adversarial Network (GAN) models for EIT reconstructions related to the mentioned fields. These neural networks based models effectively handle complex conductivity distributions, learn intricate features directly from data, and support learning by simulation cases. Their adaptability to various scenarios, data-driven nature, and ability to incorporate regularization techniques make them promising tools for EIT applications. The results demonstrate that the designed GAN-based models successfully reconstruct target positions and a more uniform background. The performance is also contrasted to alternative neural network approaches, including neural network and deep neural network models. The evaluation metrics indicate superior performance of the GAN models in comparison.