Self-supervised arbitrary-scale super-angular resolution diffusion MRI reconstruction.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Diffusion magnetic resonance imaging (dMRI) is crucial for noninvasive in vivo tissue microstructure investigation.
• Acquiring diffusion-weighted (DW) images along numerous directions is time-consuming, limiting clinical applications.
• Estimating tissue microstructure from limited diffusion directions remains a significant challenge.

Purpose of the Study:

• To propose a self-supervised network, SARDI-nn, for reconstructing diffusion-weighted (DW) images at arbitrary angular resolutions.
• To enable detailed tissue microstructure analysis from DW images acquired with a reduced number of diffusion directions.

Main Methods:

• Developed SARDI-nn, comprising DW image feature extraction (DWFE) and physics-driven implicit reconstruction (IRR) modules.
• Employed dual downsampling for training: first for low-angular resolution (LAR) DW images, second for input/target construction.
• Tested on Human Connectome Project and in-house datasets, comparing with existing methods using PSNR, SSIM, RMSE, and microstructure metrics (DKI, NODDI).

Main Results:

• SARDI-nn achieved superior performance, improving SSIM by 10.04% (upscale 3) and 5.9% (upscale 15) in reconstructed DW images.
• Outperformed supervised methods for microstructure metrics (DKI, NODDI) up to an upscale factor of 6.
• Demonstrated excellent generalizability on external datasets, confirming method robustness.

Conclusions:

• SARDI-nn is the first method capable of reconstructing high-angular resolution DW images at any upscale factor without retraining.
• Facilitates flexible variation of diffusion direction number and upscaling, enabling easy extension to unseen datasets.
• Offers a promising approach for advanced tissue microstructure analysis using limited dMRI acquisitions.