D-bar reconstructions with nonsmooth learned spatial priors in 2D electrical impedance tomography

The use of 2D Electrical Impedance Tomography (EIT) for imaging in clinical settings has gained increasing attention from the medical community in recent years. This is due in part to state-of-the-art reconstruction algorithms which have led to enhanced EIT image quality. Advances in direct D-bar reconstruction methods, for example, have allowed the inclusion of spatial priors which provide improved image sharpness and robustness. As a first step, these techniques require polygonal estimates of boundaries of regions of interest in the 2D spatial domain. In the literature, the methodology for choosing such boundaries has involved extracting this spatial information from previous medical scans, which may not exist in practice, or from an anatomical atlas, which may not be representative of individual patient physiology and pathology. Manual extraction from previous scans also leads to labor-intensive procedures and the introduction of human bias. Furthermore, in previous works, some of the sharpness provided by the introduction of priors was lost due to a mathematical need for smoothing of the a priori conductivity distribution, which also introduced computational overhead. In this work, we address these problems via (1) a method for the automated selection of boundaries via trained convolutional neural networks, and (2) use of an alternative mathematical formulation which eliminates the need for smoothing of the conductivity prior. We present a scenario where the network is trained and validated using simulated thoracic phantoms on circular domains.