Super-Resolution Diffusion Tensor Imaging using SRCNN: A Feasibility Study^.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• High-resolution diffusion imaging offers detailed anatomical insights but is limited by long scan times and image artifacts.
• Low signal-to-noise ratio (SNR) and geometric distortions are common issues in echo-planar imaging sequences.

Purpose of the Study:

• To propose and validate a deep learning method for achieving submillimeter super-resolution diffusion-weighted (DW) images.
• To assess the efficacy of a super-resolution convolutional neural network (SRCNN) for improving diffusion MRI quality.

Main Methods:

• A 2D-based deep learning method (SRCNN) was employed to generate super-resolution DW images.
• Validation involved numerical simulations comparing results to ground truth and analysis of region-of-interest (ROI) data from healthy volunteers.
• The method's performance was evaluated under varying SNR conditions.

Main Results:

• The deep learning method successfully reproduced anatomical details typically only visible with high-resolution diffusion imaging.
• Diffusion tensor imaging (DTI) derived metrics showed less than 8% error at baseline SNR > 20.
• Region-of-interest analysis revealed comparable diffusion metrics to high-resolution data, with similar distribution patterns in gray and white matter.

Conclusions:

• Deep learning using SRCNN is a feasible and robust approach for generating submillimeter diffusion MRI.
• This technique offers a valuable alternative when acquiring true sub-millimeter diffusion MRI is not practical.
• The enhanced image quality, particularly in cortical gray matter, demonstrates the potential of AI in advanced neuroimaging.