Dominant-Current Deep Learning Scheme for Electrical Impedance Tomography.

Area of Science:

• Medical Imaging
• Computational Electromagnetics
• Machine Learning

Background:

• Deep learning shows promise in electrical impedance tomography (EIT) but struggles with complex target shapes.
• Existing methods often fail to accurately reconstruct targets with sharp corners or edges using standard training data.

Purpose of the Study:

• To develop novel iterative and deep learning-based inversion methods for EIT.
• To enhance the recovery of challenging inclusions like triangular, rectangular, or lung shapes in EIT imaging.
• To demonstrate the effectiveness of these methods using diverse training data.

Main Methods:

• Proposed an iterative bases-expansion subspace optimization method (BE-SOM) utilizing induced contrast current (ICC) and total variation regularization.
• Developed a convolutional neural network (CNN) based deep learning scheme using dominant ICC features for multi-channel input.
• Trained the CNN with random circular or elliptical inclusion data to handle complex shapes.

Main Results:

• Both proposed methods demonstrated significant improvements in reconstructing targets with sharp corners and edges.
• Numerical and experimental data, including realistic phantoms with simulated pathologies, validated the methods.
• The techniques proved capable of fast, stable, and high-quality EIT imaging.

Conclusions:

• The developed iterative and deep learning methods overcome limitations in EIT imaging of complex geometries.
• These advancements hold promise for quantitative imaging in potential clinical applications.
• The study highlights the potential of novel deep learning schemes for improved EIT performance.