MR Image Denoising and Super-Resolution Using Regularized Reverse Diffusion.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Magnetic Resonance Imaging (MRI) scans are susceptible to noise, degrading diagnostic utility.
• Current deep learning denoising methods often use minimum mean squared error (MMSE) estimation, leading to blurred outputs and poor generalization to real-world data with complex noise.
• Existing models struggle with out-of-distribution data and non-standard noise patterns.

Purpose of the Study:

• To develop a novel MRI denoising method that overcomes the limitations of MMSE-based approaches.
• To introduce a technique capable of enhancing image resolution concurrently with denoising.
• To provide a flexible denoising solution with uncertainty quantification.

Main Methods:

• Implementation of a score-based reverse diffusion sampling network for MRI denoising.
• Training the network using coronal knee MRI scans.
• Testing the network's performance on out-of-distribution in vivo liver MRI data with complex noise mixtures.
• Incorporation of a super-resolution enhancement capability within the same network.

Main Results:

• The proposed method demonstrates state-of-the-art performance in denoising MRI scans.
• The network trained on knee data successfully denoises liver MRI data, showing robustness to out-of-distribution samples and complex noise.
• The integrated super-resolution capability effectively enhances the resolution of denoised images.
• The method allows for flexible control over the denoising level and provides uncertainty quantification.

Conclusions:

• Score-based reverse diffusion sampling offers a superior alternative to MMSE-based methods for MRI denoising.
• The developed network is robust, versatile, and capable of both denoising and enhancing image resolution.
• This approach provides advanced capabilities for medical image analysis, including adaptable denoising and uncertainty estimation.