Supervised Descent Learning for Thoracic Electrical Impedance Tomography.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• The absolute image reconstruction in Electrical Impedance Tomography (EIT) is an ill-posed problem.
• Traditional regularization methods for EIT suffer from low accuracy, poor noise performance, and long computation times.
• Integrating prior information into traditional EIT methods is often inflexible.

Purpose of the Study:

• To develop a novel machine learning algorithm for solving the EIT inverse problem, specifically for thorax imaging.
• To improve accuracy, noise resistance, and computational efficiency in EIT image reconstruction.
• To enable flexible integration of prior anatomical information into the EIT imaging process.

Main Methods:

• Developed the supervised descent learning EIT (SDL-EIT) inversion algorithm, inspired by the supervised descent method (SDM).
• Created a training dataset incorporating thorax contours and general lung/heart structures.
• Implemented and evaluated SDL-EIT in 2D and 3D using synthetic and measured thoracic data.

Main Results:

• SDL-EIT demonstrated superior accuracy and anti-noise performance compared to the traditional Gauss-Newton inversion (GNI) method on synthetic data.
• Reconstructions from measured thoracic data using SDL-EIT showed reasonable agreement with Computed Tomography (CT) scan images.
• The SDL-EIT algorithm accelerated image reconstruction and allowed for easy integration of prior information.

Conclusions:

• SDL-EIT is an effective algorithm for inverting measured thoracic data, offering improved performance over traditional methods.
• The algorithm's ability to integrate prior information and its speed make it a promising tool for human thorax imaging.
• SDL-EIT presents a viable machine learning-based approach to address the challenges in EIT image reconstruction.