Graph Convolutional Networks Enable Fast Hemorrhagic Stroke Monitoring With Electrical Impedance Tomography.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computational Science

Background:

• Electrical Impedance Tomography (EIT) is a promising imaging modality for stroke monitoring.
• Current EIT reconstruction methods can be computationally intensive, limiting real-time applications.
• Need for faster, high-quality image reconstruction in stroke detection.

Purpose of the Study:

• To develop a rapid image reconstruction technique for EIT-based stroke monitoring.
• To achieve image quality comparable to computationally expensive nonlinear methods.
• To enable efficient, potentially real-time, stroke detection using EIT.

Main Methods:

• Employed a post-processing approach utilizing graph convolutional networks (GCNs).
• Trained a graph U-net on linear difference reconstructions from 2D simulated stroke data.
• Applied the trained network to 3D images from simulated and experimental data, comparing with 3D vs. 2D trained networks.

Main Results:

• Graph U-net post-processing significantly enhanced linear difference reconstruction image quality.
• Achieved image quality comparable to or exceeding time-intensive nonlinear reconstruction methods.
• Reconstruction time reduced from several hours to a few minutes.

Conclusions:

• Combining fast linear difference imaging with graph U-net post-processing offers significant improvements.
• The graph framework allows training on 2D data for 3D volume processing, reducing simulation costs by ~50x.
• This approach is a feasible method for on-line monitoring of hemorrhagic stroke.