Anatomically and biochemically guided deep image prior for sodium MRI denoising

Area of Science:

• Medical Imaging
• Biophysics
• Computational Imaging

Background:

• Sodium (²³Na) MRI offers metabolic insights but suffers from low signal-to-noise ratio (SNR) and lengthy acquisition times.
• Existing denoising methods risk altering sodium signal distribution, impacting structural fidelity and quantitative accuracy.
• Deep Image Prior (DIP) methods show promise but require enhancement for complex MRI data.

Purpose of the Study:

• To develop and validate DIP-Fusion, a novel framework for robust sodium MRI denoising.
• To leverage complementary proton (¹H) and sodium (²³Na) MRI data for improved denoising performance.
• To enhance the structural fidelity and quantitative consistency of denoised sodium MRI.

Main Methods:

• Proposed DIP-Fusion framework utilizing a fused proton-sodium prior within a directional total variation (dTV) regularization scheme.
• Optimized a variational loss function incorporating data fidelity, fused dTV regularization, gradient consistency, and bias-field correction.
• Evaluated performance against classical (NLM, BM3D, TV) and deep learning (RD-DIP) methods on healthy volunteers and breast cancer patients.

Main Results:

• DIP-Fusion demonstrated consistent improvements over all baseline methods in both healthy subjects and patient datasets.
• Achieved significant PSNR gains (up to +2.36 dB in healthy, +3.35 dB in patients) and SSIM improvements (~5%).
• Showed superior robustness and accuracy compared to RD-DIP, especially under high Rician noise conditions, preserving sodium-specific signal characteristics.

Conclusions:

• DIP-Fusion effectively denoises sodium MRI by integrating multi-modal information, stabilizing DIP optimization.
• The framework enhances image quality, structural integrity, and quantitative reliability in challenging noise environments.
• This approach enables more dependable sodium MRI analysis for both research and clinical applications.