Automated characterization of noise distributions in diffusion MRI data

Area of Science:

• Medical Imaging
• Signal Processing
• Computational Neuroscience

Background:

• Accurate noise distribution estimation in diffusion MRI is crucial for quantifying acquisition uncertainties.
• Standard Rician or noncentral chi distributions often fail with advanced imaging methods like parallel imaging.
• Information about acquisition specifics (e.g., coil sensitivity) is typically unavailable for noise modeling.

Purpose of the Study:

• Introduce two automated methods to estimate noise distributions in magnitude diffusion MRI data.
• Develop a robust framework independent of detailed acquisition parameters.
• Assess the performance of the proposed methods for bias correction and denoising.

Main Methods:

• Utilized Gamma distribution moments and maximum likelihood equations to estimate noise parameters.
• Incorporated a rejection step for artifact robustness and automation.
• Validated methods using simulations with controlled noise, phantom data, and in vivo datasets across multiple vendors and parallel imaging techniques.

Main Results:

• Simulations demonstrated that assuming Rician distribution leads to misestimation with parallel imaging.
• Real data showed signal leakage in multiband imaging also causes noise misestimation.
• Proposed methods provide stable parameter estimation with lower variability compared to existing algorithms.
• Bias correction and denoising tasks showed noise reduction at high b-values.

Conclusions:

• The proposed automated Gamma distribution-based methods accurately estimate noise in magnitude diffusion MRI.
• The framework is robust to artifacts and parallel imaging variations, outperforming existing methods.
• These algorithms enhance quantitative MRI by providing reliable noise characterization for bias correction and denoising.