An MRI denoising method using image data redundancy and local SNR estimation

Area of Science:

• Medical Imaging
• Signal Processing
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) is crucial for diagnostics but susceptible to Rician noise.
• Traditional denoising methods using local neighborhoods are suboptimal for 3D MRI data.
• Efficiently removing noise while preserving anatomical details is a significant challenge.

Purpose of the Study:

• To develop an advanced 3D denoising method for MR images under a Rician noise model.
• To improve upon conventional Local Minimum Mean Square Error (LMMSE) estimation by utilizing non-local similar samples.
• To enhance the preservation of anatomical structures in denoised MR images.

Main Methods:

• A novel LMMSE-based method is proposed, modeling MR data as random fields.
• A principled approach selects similar samples from a large data portion, not just local neighborhoods.
• An effective similarity measure based on local statistical moments and automatic parameter selection via signal-to-noise ratio (SNR) are employed.
• A recursive version of the filter is introduced for enhanced performance.

Main Results:

• The proposed method demonstrates superior performance in noise removal compared to state-of-the-art filters.
• Experimental results on synthetic and real MR datasets show significant noise reduction.
• Anatomical structures are well-preserved, indicating high fidelity in the denoised images.

Conclusions:

• The developed LMMSE-based 3D denoising method effectively addresses Rician noise in MR images.
• The approach of utilizing non-local similar samples significantly improves denoising efficacy and structure preservation.
• This method offers a valuable tool for enhancing the quality and diagnostic utility of MR imaging.