Deep Network Regularization for Phase-Based Magnetic Resonance Electrical Properties Tomography With Stein's Unbiased...

Area of Science:

• Medical Imaging
• Electrical Properties Tomography
• Computational Neuroscience

Background:

• Phase-based Magnetic Resonance Electrical Properties Tomography (MR-EPT) estimates tissue conductivity.
• The Laplacian operator in MR-EPT is sensitive to noise, leading to amplification.
• Existing denoising methods struggle with noise amplification and boundary errors.

Purpose of the Study:

• To develop a novel unsupervised denoiser for MRI transceive phase images in MR-EPT.
• To improve the accuracy and reduce noise in conductivity maps derived from MR-EPT.
• To create a blind, fully unsupervised 2D MR-EPT reconstruction algorithm.

Main Methods:

• Utilized Deep Image Prior (DIP) with random CNN initialization for implicit regularization.
• Incorporated Stein's Unbiased Risk Estimator (SURE) for network optimization, enabling unsupervised learning.
• Processed real and imaginary MRI images instead of phase images for theoretical alignment.

Main Results:

• Significantly reduced residual noise in MR-EPT phase maps.
• Outperformed existing denoising techniques in phantom and simulated brain data.
• Produced more accurate conductivity maps with reduced errors in healthy volunteers and patients.

Conclusions:

• The proposed unsupervised denoiser effectively mitigates noise amplification in MR-EPT.
• This method offers a robust, data-efficient solution for improved conductivity imaging.
• Represents the first blind, fully unsupervised approach for 2D phase-based MR-EPT reconstruction.