Deep learning based segmentation of brain tissue from diffusion MRI.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) segmentation is crucial for analyzing brain microstructure and improving tractography.
• Current methods rely on anatomical MRI (e.g., T1/T2-weighted) registration, which is challenging due to dMRI's distortions and lower resolution.
• Inter-modality registration issues limit the accuracy of existing dMRI segmentation techniques.

Purpose of the Study:

• To develop a deep learning method (DDSeg) for direct dMRI segmentation, eliminating the need for anatomical MRI registration.
• To improve the accuracy of brain tissue segmentation in dMRI data.
• To create a robust segmentation method applicable to dMRI data from various acquisition protocols.

Main Methods:

• A convolutional neural network (CNN) was trained on high-quality Human Connectome Project (HCP) dMRI data.
• A novel augmented target loss function was employed to enhance accuracy at tissue boundaries.
• Diffusion kurtosis imaging (DKI) parameters, derived from a corrected mean-kurtosis-curve method, were integrated with diffusion tensor imaging (DTI) parameters.

Main Results:

• The DDSeg method achieved high tissue segmentation accuracy on HCP data.
• The trained model successfully segmented dMRI data from different acquisitions, even with lower resolution and fewer gradient directions.
• Integration of DKI parameters improved the discrimination between tissue types.

Conclusions:

• DDSeg offers a robust and accurate deep learning solution for dMRI segmentation, independent of anatomical MRI.
• The method effectively handles variations in dMRI acquisition protocols and data quality.
• Incorporating DKI parameters enhances segmentation performance, providing valuable insights into brain microstructure.