A nonlocal maximum likelihood estimation method for Rician noise reduction in MR images

Area of Science:

• Medical Imaging
• Signal Processing
• Computational Analysis

Background:

• Magnetic resonance (MR) image denoising is crucial for accurate clinical diagnosis and automated analysis.
• Rician noise in MR magnitude images is signal-dependent at low signal-to-noise ratios (SNR), complicating noise removal.
• Existing methods struggle with random fluctuations and bias introduced by Rician noise.

Purpose of the Study:

• To develop an effective method for estimating noise-free signals from MR magnitude images.
• To address the challenges posed by Rician noise in MR imaging.
• To improve the accuracy of MR image analysis through advanced denoising techniques.

Main Methods:

• Modeling MR images as random fields with similar neighborhoods sharing distributions.
• Proposing a nonlocal maximum likelihood (NLML) estimation method for Rician noise reduction.
• Comparing NLML performance against nonlocal means and local maximum likelihood (LML) methods.

Main Results:

• NLML provides a more accurate recovery of true signals from Rician noise compared to the nonlocal means algorithm, enhancing SNR, contrast, and reducing method error.
• NLML demonstrates superior performance over LML in preserving and defining sharp tissue boundaries, as measured by a sharpness metric.
• NLML exhibits better overall performance in terms of method error compared to LML.

Conclusions:

• The proposed NLML method is highly effective for postacquisition denoising of MR images corrupted by Rician noise.
• NLML offers significant advantages over conventional methods in terms of image quality, boundary preservation, and estimation accuracy.
• This technique holds promise for enhancing the reliability of MR image-based diagnostics and computational analyses.