Super-resolution of diffusion-weighted images using space-customized learning model.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Diffusion-weighted imaging (DWI) is crucial for noninvasive brain microstructural analysis.
• Clinical DWI faces a resolution-time tradeoff, limiting practical applications.
• Super-resolution techniques for natural images are challenging to apply to high-dimensional, non-Euclidean DWI data.

Purpose of the Study:

• To develop an end-to-end deep learning network for post-processing DWI to enhance spatial resolution.
• To improve the quality and utility of diffusion-weighted imaging data.

Main Methods:

• Proposed a hybrid deep learning approach combining convolutional neural networks (CNNs) for spatial (x-space) and graph CNNs (GCNNs) for diffusion gradient (q-space) domains.
• Utilized a Gaussian kernel in q-space to bridge CNN and GCNN feature representations.
• Developed a space-customized network for high-dimensional DWI data.

Main Results:

• Demonstrated effective improvement in DWI quality on the Human Connectome Project dataset.
• Validated the method's efficacy in enhancing spatial resolution of diffusion-weighted imaging.
• Showcased advantages of the approach in downstream neuroimaging analysis tasks.

Conclusions:

• The hybrid CNN-GCNN model excels at enhancing spatial resolution in DWI scans.
• This deep learning approach is advantageous for feature learning from heterogeneous spatial data in DWI.
• The method offers a promising solution for overcoming resolution limitations in clinical DWI.