Resolution improvement and algorithmic dependence of machine learning for post-processing respiratory EIT images

Electrical impedance tomography (EIT) is an imaging modality in which electric fields arising from currents applied on electrodes are used to form dynamic images of physiological processes, such as respiration, by plotting the reconstructed conductivity distribution. This work investigates the effectiveness of using machine learning to post-process EIT images of respiration, validated against a CT scan taken immediately after the EIT data collection, and the dependence of the results on the reconstruction algorithm used to compute the pre-processed image. Here, a training set for post-processing is computed from a set of CT scans, and a deep learning neural network is used to post-process EIT images from patients with cystic fibrosis reconstructed using (a) the D-bar method and (b) one step of a Gauss-Newton method. The images are compared using the structural similarity index measure (SSIM) to 'ground truth' EIT images derived from the CT scans. Results show that while the deep learning post-processing method effectively sharpens edges and reduces blurring, the spatial accuracy depends on the original algorithm used to compute the pre-processed image. The D-bar images result in a higher SSIM than the Gauss-Newton images, and while both sets of images are able to detect a large region of air trapping, they differ when estimating the extent of pathology.