Conditional generative diffusion deep learning for accelerated diffusion tensor and kurtosis imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Diffusion MRI techniques like DTI and DKI provide crucial insights into brain microstructure.
• Long acquisition times for diffusion-weighted images (DWIs) limit clinical applicability and patient comfort.
• Developing methods to accelerate DWI acquisition without compromising metric quality is essential.

Purpose of the Study:

• To introduce DiffDL, a generative diffusion probabilistic model.
• To enable the generation of high-quality DTI and DKI metrics from a reduced set of DWIs.
• To address the challenge of prolonged data acquisition in diffusion MRI.

Main Methods:

• DiffDL was trained on Human Connectome Project data using a UNet architecture.
• Training involved pairing high-quality DTI/DKI metrics with subsets of DWIs.
• Model performance was rigorously evaluated against conventional methods and a baseline UNet.

Main Results:

• DiffDL significantly improved the quality and accuracy of fractional anisotropy (FA) and mean diffusivity (MD) maps.
• The model outperformed conventional DKI modeling and a baseline UNet across various acceleration scenarios.
• Quantitative metrics (NMAE, PSNR, PCC) confirmed DiffDL's superior performance and ability to capture full metric ranges.

Conclusions:

• DiffDL shows significant potential to reduce diffusion MRI acquisition times while preserving metric quality.
• Further research is needed to optimize computational efficiency and validate DiffDL in clinical settings.
• The generative approach of DiffDL allows for uncertainty quantification, enhancing its utility.