Denoising and fast diffusion imaging with physically constrained sparse dictionary learning

Area of Science:

• Medical Imaging
• Biophysics
• Computational Neuroscience

Background:

• Diffusion-weighted imaging (DWI) visualizes water diffusion in tissues.
• High b-value DWI and Diffusion Spectrum Imaging (DSI) suffer from noise and long acquisition times.
• Current denoising methods have limitations in preserving DWI data quality.

Purpose of the Study:

• To develop a novel method for denoising DWI data.
• To reduce the number of measurements required for DSI acquisitions.
• To maintain or improve data quality in DWI and DSI.

Main Methods:

• Utilized sparse dictionary learning constrained by physical signal properties (symmetry, positivity).
• Learned a dictionary of diffusion profiles across multiple DW images simultaneously.
• Applied the method to simulated data and two real DSI datasets.

Main Results:

• Dictionary learning outperformed existing methods (mirror symmetry, Gaussian denoising, non-local means) in signal estimation.
• Demonstrated the ability to generate high-resolution DSI data with fewer acquired images using a pre-learned dictionary.
• Achieved effective denoising and enabled faster acquisitions, requiring approximately 40 measurements for high b-value DSI.

Conclusions:

• Dictionary learning provides an effective approach for denoising DWI and accelerating DSI acquisition.
• This technique enhances signal estimation and data quality in diffusion imaging.
• The method holds significant potential for advancing connectomics research using DSI.