Super-resolved q-space learning of diffusion MRI.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Machine Learning

Background:

• Diffusion MRI (dMRI) non-invasively studies brain structures.
• dMRI quality depends on q-space diffusion gradients.
• High-angular (HA) dMRI requires long scans; reducing gradients causes underestimation.

Purpose of the Study:

• To develop a deep compressive sensing-based q-space learning (DCS-qL) method.
• To estimate HA dMRI from low-angular dMRI data.

Main Methods:

• Designed a deep network by unfolding proximal gradient descent for compressive sensing.
• Employed a lifting scheme for reversible network transforms.
• Used self-supervised regression for signal-to-noise enhancement and semantic-guided patch mapping for feature extraction.

Main Results:

• The DCS-qL approach shows promising performance in reconstructing HA dMRI.
• Accurate estimation of microstructural indices, fiber orientation distribution, and fiber bundles was achieved.
• The method successfully enhanced the signal-to-noise ratio of diffusion data.

Conclusions:

• The proposed DCS-qL method provides more accurate neural structure estimation than existing approaches.
• This technique offers a potential solution for improving clinical applicability of HA dMRI.