Noise reduction in magnitude diffusion-weighted images using spatial similarity and diffusion redundancy

Area of Science:

• Medical Imaging
• Neuroscience
• Image Processing

Background:

• Diffusion-weighted imaging (DWI) is valuable in clinical settings.
• DWI faces challenges with low signal-to-noise ratio (SNR), particularly at high resolutions or diffusion sensitivities.

Purpose of the Study:

• To develop and validate a denoising method for magnitude DWI.
• To address the low SNR problem in DWI, especially for high-quality brain imaging.

Main Methods:

• A two-module denoising approach: pre-denoising using local kernel principal component analysis (KPCA) and post-filtering with patch-based non-local mean (NLM).
• The method mines diffusion redundancy and non-local self-similarity within DWI data.

Main Results:

• The proposed method significantly improves whole-brain SNR in both simulated and in vivo datasets.
• Enhanced SNR leads to better diffusion metrics estimation, crossing fiber discrimination, and fiber tractography.
• The method demonstrates superior performance, especially for denoising diffusion data acquired with sensitivity encoding (SENSE).

Conclusions:

• The developed denoising method offers significant practical utility for acquiring high-quality whole-brain diffusion data.
• This advancement is critical for improving the reliability and scope of numerous neuroscience studies.