Physics Informed Neural Networks (PINN) for Low Snr Magnetic Resonance Electrical Properties Tomography (MREPT)

Area of Science:

• Medical Imaging
• Biophysics
• Computational Science

Background:

• Electrical properties (EPs) of tissues are crucial for early cancer detection.
• Magnetic resonance electrical properties tomography (MREPT) non-invasively probes tissue EPs using MRI data.
• Traditional MREPT methods using numerical differentiation (ND) are sensitive to noise in clinical MRI data.

Purpose of the Study:

• To develop a novel Physics-Informed Neural Network MREPT (PINN-MREPT) model.
• To address limitations of existing MREPT techniques, particularly noise sensitivity and data requirements.
• To enhance the accuracy and robustness of MREPT for potential clinical applications.

Main Methods:

• Proposed a PINN-MREPT model integrating a canonical analytic MREPT model.
• Incorporated a reference padding layer with known EPs to provide necessary collocation points for network optimization.
• Embedded an optimizable diffusion coefficient within the analytic MREPT model for enhanced noise robustness.

Main Results:

• PINN-MREPT demonstrated superior reconstruction accuracy and sensitivity to tumor-like tissues compared to traditional MREPT methods.
• The model showed high correlation in line profiles within regions of interest, even at varying noise levels.
• Experimental phantom measurements validated the advantages of the proposed PINN-MREPT approach.

Conclusions:

• The developed PINN-MREPT model significantly outperforms conventional MREPT methods in accuracy and noise resilience.
• The inclusion of a diffusion term is critical for achieving noise-robust MREPT reconstructions.
• This work represents a significant advancement towards the clinical implementation of MREPT for improved cancer diagnostics.