Robust multi-coil MRI reconstruction via self-supervised denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning (DL) based reconstruction methods offer high quality for magnetic resonance imaging (MRI).
• Training these DL methods typically requires large, noise-free datasets, which are impractical to acquire.
• K-space data used in MRI is inherently noisy and often multi-coil.

Purpose of the Study:

• To assess the impact of self-supervised denoising as a pre-processing step for DL-based MRI reconstruction.
• To evaluate how denoising affects DL methods trained on noisy K-space data.
• To determine if denoising can improve reconstruction quality and efficiency.

Main Methods:

• Leveraged Generalized Stein's Unbiased Risk Estimate (GSURE) for self-supervised denoising.
• Evaluated two DL reconstruction methods: Diffusion Probabilistic Models (DPMs) and Model-Based Deep Learning (MoDL).
• Tested on accelerated multi-coil MRI reconstruction using T2-weighted brain and fat-suppressed proton-density knee scans at various signal-to-noise ratios (SNRs).

Main Results:

• Self-supervised denoising significantly enhanced MRI reconstruction quality and efficiency.
• Training DL networks with denoised data resulted in lower normalized root mean squared error (NRMSE).
• Denoised data led to higher structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) across different SNRs.

Conclusions:

• Denoising is a crucial pre-processing technique for effective DL-based MRI reconstruction.
• Improving input data quality through denoising enables the training of more efficient DL networks.
• This approach may eliminate the necessity for acquiring noise-free reference MRI scans for training.