Machine Learning-Based Image Reconstruction in Wearable CC-EIT of the Thorax: Robustness to Electrode Displacement

TL;DR

This study shows that training a neural network with data including electrode displacement improves image quality in wearable CC-EIT of the thorax.

Contribution

A new algorithm for simulating electrode displacement and training a neural network to improve CC-EIT robustness.

Findings

Electrode displacement significantly reduces image quality when not included in training data.

Training with displaced electrode samples improves reconstruction quality and robustness.

Pixel-to-pixel metrics confirmed the effectiveness of the displacement-aware training approach.

Abstract

This study investigates the influence of variable electrode positions on image reconstruction in capacitively coupled electrical impedance tomography (CC-EIT) of the human thorax. Images were reconstructed by an adversarial neural network trained on a synthetic dataset generated using a tomographic model that included a wearable elastic band with 32 electrodes attached. Dataset generation was conducted using a previously developed numerical phantom of the thorax, combined with a newly developed algorithm for random selection of electrode positions based on physical limitations resulting from the elasticity of the band and possible position inaccuracies while putting the band on the patient’s chest. The thorax phantom included the heart, lungs, aorta, and spine. Four training and four testing datasets were generated using four different levels of electrode displacement. Reconstruction was conducted using four versions of neural networks trained on the datasets, with random ellipses included and noise added to achieve an SNR of 30 dB. The quality was assessed using pixel-to-pixel metrics such as the root-mean-square error, structural similarity index, 2D correlation coefficient, and peak signal-to-noise ratio. The results showed a strong negative influence of electrode displacement on reconstruction quality when no samples with displaced electrodes were present in the training dataset. Training the network on the dataset containing samples with electrode displacement allowed us to significantly improve the quality of the reconstructed images. Introducing samples with misplaced electrodes increased neural network robustness to electrode displacement while testing.

Figures