Deep learning-based Desikan-Killiany parcellation of the brain using diffusion MRI.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Accurate brain parcellation is crucial for advanced neuroimaging analyses.
• Current methods often rely on anatomical MRI, introducing potential errors and limiting versatility.
• Direct parcellation in diffusion MRI (dMRI) space is highly desirable.

Purpose of the Study:

• To develop a novel deep learning framework for direct brain parcellation using only dMRI data.
• To implement a hierarchical, two-stage segmentation network for accurate parcellation based on the Desikan-Killiany (DK) atlas.
• To evaluate the performance and robustness of the proposed method compared to existing approaches.

Main Methods:

• A hierarchical, two-stage deep learning segmentation network was developed.
• The framework directly utilizes diffusion MRI-derived parameter maps for parcellation.
• An extensive ablation study identified optimal diffusion parameters (fractional anisotropy, trace, sphericity, maximum eigenvalue) for enhanced accuracy.

Main Results:

• The proposed method achieved higher Dice Similarity Coefficients on the Human Connectome Project dataset compared to state-of-the-art methods.
• Robustness was demonstrated on the Consortium for Neuropsychiatric Phenomics dataset across different resolutions and protocols, showing more homogeneous parcellations.
• The framework successfully performs direct parcellation based on the Desikan-Killiany atlas using only dMRI data.

Conclusions:

• This novel deep learning framework enables accurate and practical brain parcellation directly from diffusion MRI data.
• The method eliminates the need for anatomical MRI and complex registration steps, advancing neuroimaging analysis.
• The publicly available implementation facilitates wider adoption and further research in dMRI-based brain mapping.