Deep learning based reconstruction enables high-resolution electrical impedance tomography for lung function...

Area of Science:

• Medical Imaging
• Electrical Impedance Tomography (EIT)
• Deep Learning

Background:

• Traditional regularized least square methods for lung time difference electrical impedance tomography (tdEIT) reconstruction suffer from ill-posedness and low spatial resolution.
• Existing deep learning validation often relies on simulated data and focuses on image quality, not clinical indicator accuracy.
• There is a need for robust validation of deep learning EIT reconstruction using in vivo human data and clinical metrics.

Purpose of the Study:

• To evaluate a deep learning-based tdEIT reconstruction method using in vivo human chest data.
• To benchmark the accuracy of deep learning EIT reconstructions against spirometry measurements.
• To assess the potential of deep learning tdEIT for predicting clinical lung function indicators.

Main Methods:

• A variational autoencoder was trained on high-resolution human chest simulations.
• The trained model was applied to an EIT dataset from 22 healthy subjects during various breathing paradigms.
• Reconstructed EIT data was benchmarked against simultaneous spirometry measurements.

Main Results:

• Deep learning reconstructed global conductivity showed a significant correlation (> 0.9) with measured volume-time curves.
• EIT-derived lung function indicators were highly correlated (> 0.75) with standard spirometry indicators.
• The method generated high-resolution EIT images.

Conclusions:

• Deep learning reconstruction of lung tdEIT can accurately predict lung volume and spirometry indicators.
• The developed method demonstrates potential as a competitive approach in clinical settings.
• This approach offers a promising alternative to traditional methods for lung EIT analysis.