Sparse DSI: learning DSI structure for denoising and fast imaging.

Area of Science:

• Medical Imaging
• Biophysics
• Computational Neuroscience

Background:

• Diffusion spectrum imaging (DSI) utilizes multiple diffusion-weighted images (DWI) to visualize complex water diffusion in biological tissues.
• Accurate modeling of diffusion profiles is crucial for high-resolution DSI data.
• Current methods face challenges with noise and acquisition time.

Purpose of the Study:

• To develop a novel sparse coding method for DSI data analysis.
• To improve the estimation of diffusion signals by incorporating physical constraints.
• To enable faster DSI acquisition and enhance data quality.

Main Methods:

• Sparse coding constrained by signal symmetry and positivity was used to learn a dictionary of diffusion profiles.
• The method was applied to DWI data from two subjects, learning jointly across all acquired images.
• Performance was evaluated by predicting undersampled DWI data.

Main Results:

• The proposed sparse coding method achieved better prediction of unseen DWI data compared to classic symmetry procedures, even when using only half the acquired data.
• Effective signal estimation was demonstrated even with significantly reduced DWI measurements (e.g., 40 measurements).
• The approach scales to full brain data and shows robustness when diffusion profiles are estimated from a different subject's undersampled data.

Conclusions:

• Sparse coding offers a powerful approach to model complex diffusion signals in DSI.
• This method can significantly reduce acquisition time and improve the quality of high-resolution DSI data.
• The technique holds promise for more efficient and accurate neuroimaging studies.