A comparison of techniques to improve pulmonary EIT image resolution using a database of simulated EIT images

Abstract

Inherently low spatial resolution is a well-known challenge in electrical impedance tomography image reconstruction. Various approaches such as the use of spatial priors and post-processing techniques have been proposed to improve the resolution, but in the literature, comparisons using a common dataset representative of clinical images have not been considered. Here, we consider a database of 81,710 simulated EIT datasets constructed from pulmonary CT scans of 89 infants. Four techniques for improved image resolution and several combinations thereof are proposed and compared quantitatively on 16,341 known test cases reserved from the database. The techniques include an end-to-end deep learning reconstruction approach, post-processing of real-time one-step Gauss–Newton (GN) reconstructions using machine learning, post-processing using the Schur complement method, the use of an initial guess for the one-step GN method derived from the image database, and a method that makes use of the eigenfunctions of the principal component analysis of image vectors in the database. All methods resulted in improved metrics of error measurement compared to the Newton one-step error reconstruction method used as the basis for comparison.