MRI denoising using progressively distribution-based neural network.

Area of Science:

• Medical Imaging
• Computer Vision
• Signal Processing

Background:

• Magnetic Resonance (MR) images are susceptible to noise, impacting quantitative analysis.
• Rician distribution commonly models noise in magnitude MR images due to underlying Gaussian noise.
• Existing denoising methods often rely on spatial information, overlooking pixel-level distributions.

Purpose of the Study:

• To develop a novel supervised network for denoising MR images by capturing pixel-level distribution information.
• To improve the accuracy of quantitative measurements from noisy MR data.
• To address limitations of traditional spatial-based denoising techniques.

Main Methods:

• A progressive network learning strategy is proposed, fitting pixel-level and feature-level intensity distributions.
• The network architecture comprises two residual blocks: one for pixel domain fitting (without batch normalization) and another for feature domain matching (with batch normalization).
• The approach utilizes supervised learning to capture pixel-level distribution characteristics.

Main Results:

• Experimental results on synthetic, complex-valued, and clinical MR brain images demonstrate the network's effectiveness.
• The proposed method shows substantially improved quantitative measures compared to existing techniques.
• Visual inspections confirm significant enhancements in image quality after denoising.

Conclusions:

• The developed supervised network offers a promising approach for MR image denoising.
• The pixel-level distribution fitting strategy effectively enhances image quality and quantitative accuracy.
• This method holds great potential for improving diagnostic capabilities in MR imaging.