DENOISING DIFFUSION MRI: CONSIDERATIONS AND IMPLICATIONS FOR ANALYSIS

Area of Science:

• Medical Imaging
• Neuroimaging
• Biophysics

Background:

• Noise in diffusion MRI (dMRI) data reduces measurement accuracy and precision.
• Existing dMRI denoising methods' performance, particularly regarding bias and noise floor effects, requires objective characterization.
• Gaps persist in objectively assessing the efficacy of various dMRI denoising techniques.

Approach:

• Defined criteria to evaluate dMRI denoising performance, focusing on signal quality, bias reduction, spatial resolution preservation, and agreement with a gold standard.
• Utilized newly acquired complex dMRI datasets with multiple repeats across different signal-to-noise ratio (SNR) regimes.
• Applied and compared exemplar patch-based denoising algorithms (Non-Local Means, Marchenko-Pastur PCA, NORDIC) in both complex and magnitude domains.

Key Points:

• All tested denoising methods reduced noise variance, but not consistently bias from the noise floor.
• Spatial resolution was penalized by all methods, with variations depending on the algorithm and implementation.
• Complex domain denoising demonstrated advantages over magnitude domain denoising across all evaluated criteria.
• Challenges remain in defining a definitive gold standard for dMRI denoising evaluation.

Conclusions:

• dMRI denoising is crucial for improving data quality and enabling advanced applications like ultra-high-resolution imaging.
• Complex domain processing of dMRI data is advantageous for denoising performance.
• Further research is needed to refine denoising strategies and gold standard definitions for dMRI.