q-Space Deep Learning: Twelve-Fold Shorter and Model-Free Diffusion MRI Scans

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Diffusion MRI is crucial for non-invasive neuroimaging but requires long acquisition times.
• Advanced diffusion models offer detailed microstructural insights but are limited by scan duration.
• Current data processing pipelines in diffusion MRI are often suboptimal and time-consuming.

Purpose of the Study:

• To investigate the application of deep learning to streamline diffusion MRI data processing.
• To reduce the acquisition time for advanced diffusion MRI models.
• To enable robust microstructural characterization with minimal data points.

Main Methods:

• Applied deep learning algorithms, based on artificial neural networks, to optimize diffusion MRI data processing into a single step.
• Developed a novel approach to estimate advanced diffusion MRI scalar measures with significantly reduced data points.
• Validated the method's ability to detect abnormalities without relying on traditional diffusion models.

Main Results:

• Achieved a twelve-fold reduction in diffusion MRI scan time.
• Successfully estimated diffusion kurtosis measures using only 12 data points.
• Enabled estimation of neurite orientation dispersion and density measures from just 8 data points.
• Demonstrated the capability to detect abnormalities without conventional diffusion models.

Conclusions:

• Deep learning offers a streamlined and efficient alternative to classical data processing in diffusion MRI.
• Reduced scan times facilitate faster and more robust diffusion MRI protocols for clinical routine.
• This advancement enhances the applicability of advanced diffusion MRI techniques, particularly for pediatric and uncooperative patients.