A Rotational Invariant Neural Network for Electrical Impedance Tomography Imaging without Reference Voltage:...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Science

Background:

• Electrical Impedance Tomography (EIT) is a non-invasive imaging method.
• EIT faces challenges in reconstructing sharp conductivity images due to its non-linear and ill-posed nature.
• Artificial Neural Networks (ANNs) show promise for improving EIT image reconstruction.

Purpose of the Study:

• To develop an ANN capable of solving the EIT inverse problem without a reference voltage.
• To enhance the robustness and clarity of EIT conductivity image reconstruction.
• To evaluate the ANN's performance in real-world scenarios.

Main Methods:

• A novel ANN architecture was designed, incorporating reused dense layers to leverage rotational symmetries in circular domains.
• Training data was expanded using a novel data augmentation technique and a wider conductivity range to prevent bias.
• The ANN's performance was compared against the Gauss-Newton approach and validated on phantom tank measurements and a pig study.

Main Results:

• The proposed ANN demonstrated superior robustness to noise compared to the analytical Gauss-Newton method.
• EIT phantom tank measurements reconstructed by the ANN exhibited clearer images with reduced ringing artifacts.
• Reconstructions from an experimental pig study confirmed the ANN's effectiveness under real-world conditions.

Conclusions:

• The developed ANN can successfully reconstruct EIT images without the necessity of a reference voltage.
• The ANN offers a more robust and artifact-free solution for EIT image reconstruction.
• This approach advances the application of ANNs in biomedical imaging.