Magnetic Resonance Electrical Properties Tomography Based on Modified Physics- Informed Neural Network and...

Area of Science:

• Biomedical Imaging
• Computational Electromagnetics
• Machine Learning

Background:

• Magnetic Resonance Electrical Property Tomography (MREPT) noninvasively maps tissue electrical properties using MRI data.
• Conventional MREPT methods face challenges with artifacts and numerical errors due to simplified assumptions and differentiation.
• Existing deep learning methods are either data-hungry or limited in validation scope.

Purpose of the Study:

• To develop a novel, model-driven deep learning method for MREPT.
• To improve the accuracy and robustness of electrical property reconstruction in MRI.

Main Methods:

• Utilized physics-informed neural networks (PINNs) with fully connected networks (FCNNs) for MREPT.
• Employed automatic differentiation to compute spatial gradients of electrical properties.
• Optimized FCNNs using the convection-reaction MREPT equation residual as the loss function, incorporating multi-constraints.

Main Results:

• Demonstrated the method's efficacy on a 3D head model, digital phantom, and experimental phantom.
• Successfully reconstructed spatial distributions of electrical properties with improved accuracy.
• Validated the model-driven approach in realistic and experimental scenarios.

Conclusions:

• The proposed PINN-based MREPT method offers a robust and accurate solution for electrical property reconstruction.
• This approach overcomes limitations of conventional and existing deep learning MREPT techniques.
• The method shows promise for advanced biomedical imaging applications.