Image Reconstruction Using Supervised Learning in Wearable Electrical Impedance Tomography of the Thorax

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• Electrical impedance tomography (EIT) is a non-invasive imaging method.
• Contactless CCEIT offers potential for wearable devices.
• Machine learning shows promise for EIT image reconstruction.

Purpose of the Study:

• Investigate deep neural networks for CCEIT image reconstruction with 32 electrodes.
• Compare fully connected deep neural networks and conditional generative adversarial networks (cGANs).
• Evaluate reconstruction quality against algebraic methods.

Main Methods:

• Generated training data using numerical simulation of a thorax phantom with modeled lung illnesses (pneumothorax, pleural effusion, hydropneumothorax).
• Employed a 32-electrode sensor and the ECTsim toolbox for forward problem and measurement simulation.
• Trained two supervised artificial neural networks (fully connected and cGAN) for image reconstruction.

Main Results:

• Both deep learning networks outperformed algebraic reconstruction methods in pixel-to-pixel metrics.
• cGANs demonstrated significantly better performance than fully connected networks, particularly with noisy data.
• ROC AUC analysis indicated that algebraic methods can still yield satisfactory diagnostic value.

Conclusions:

• Supervised deep learning, especially cGANs, enables effective CCEIT image reconstruction for regional lung function assessment.
• CCEIT with advanced machine learning holds potential for diagnosing lung conditions like pneumothorax and pleural effusion.
• Further research into deep learning for higher electrode count EIT is warranted.