PIFON-EPT: MR-Based Electrical Property Tomography Using Physics-Informed Fourier Networks.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Electromagnetics

Background:

• Electrical Properties (EP) Tomography (EPT) is crucial for medical diagnostics.
• Reconstructing EP from magnetic resonance (MR) measurements is challenging due to noise and data incompleteness.
• Existing methods struggle with high-frequency details and complex material interfaces.

Purpose of the Study:

• To introduce Physics-Informed Fourier Networks for Electrical Properties Tomography (PIFON-EPT).
• To develop a deep learning method for reconstructing EP from noisy and incomplete MR data.
• To simultaneously denoise, complete MR measurements, and estimate object EP.

Main Methods:

• Physics-informed deep learning using the Helmholtz equation.
• Two neural networks for transmit field denoising/completion and EP estimation.
• Embedding random Fourier features for high-frequency detail learning.

Main Results:

• PIFON-EPT accurately reconstructs EP and transmit fields in simulations (3T and 7T).
• Reconstruction achieved ≤ 5% error for EP and ≤ 1% error for denoising/completion with only 20% noisy data.
• Adapted method improved results at material interfaces using the generalized Helmholtz equation.

Conclusions:

• PIFON-EPT is the first method to simultaneously reconstruct EP and transmit fields from incomplete, noisy MR data.
• The approach enables efficient learning of high-frequency details crucial for accurate EP reconstruction.
• PIFON-EPT offers new possibilities for advancing EPT research and applications.