Modeling diffusion-weighted MRI as a spatially variant gaussian mixture: application to image denoising

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Image Processing

Background:

• High angular resolution diffusion imaging (HARDI) is crucial for understanding brain microstructure.
• HARDI data is susceptible to noise, particularly under clinical time constraints, affecting image quality and subsequent analyses.
• Existing denoising methods often struggle with low signal-to-noise ratio (SNR) data.

Purpose of the Study:

• To develop a general, spatially variant mixture model for HARDI data analysis.
• To apply the model for image denoising and segmentation of diffusion MRI volumes.
• To improve the quality of HARDI data for clinical applications.

Main Methods:

• A Gaussian mixture model was trained using HARDI signal attenuation data.
• A Markov random field prior was imposed on mixture weights for spatial smoothness.
• The model was trained unsupervised using expectation-maximization and optimized with the minimum message length criterion.

Main Results:

• The proposed model-based denoising significantly outperformed common methods at low SNR.
• Tractography quality improved substantially when performed on denoised data.
• Mixture probability maps showed potential for image segmentation tasks.

Conclusions:

• The spatially variant mixture model effectively denoises diffusion MRI data, especially at low SNR.
• This enables the use of faster, less noisy pulse sequences in clinical settings.
• The model enhances the utility of diffusion MRI for both research and clinical practice.