Investigating Multiple Physical Priors in Deep Learning for Electrical Properties Reconstruction in MREPT

Area of Science:

• Medical Imaging
• Computational Electromagnetics
• Artificial Intelligence

Background:

• Magnetic Resonance Electrical Properties Tomography (MREPT) enables non-invasive in-vivo conductivity and permittivity mapping.
• Current deep learning (DL) MREPT methods face challenges in reliability and generalization for clinical use.

Purpose of the Study:

• To improve the accuracy and generalization of DL-based MREPT by integrating physics-informed constraints.
• To evaluate the impact of various physical priors on MREPT reconstruction.

Main Methods:

• Developed a physics-informed deep learning framework for MREPT.
• Incorporated gradient similarity, Helmholtz equation, complex arithmetic, curl, and Laplacian operators as constraints.
• Quantitatively analyzed the impact of different physical priors on reconstruction.

Main Results:

• The proposed framework demonstrated enhanced generalization and predictive accuracy.
• Achieved promising results in reconstructing electrical properties of the human head.
• Physics-informed constraints significantly improved the robustness of DL-based MREPT.

Conclusions:

• Integrating multi-physical constraints enhances the robustness and clinical applicability of DL-based MREPT.
• This approach shows potential for accurate non-invasive brain conductivity and permittivity mapping.