A Supervoxel-Based Method for Groupwise Whole Brain Parcellation with Resting-State fMRI Data

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Network Science

Background:

• Node definition is critical for human brain network analysis and functional connectivity studies.
• Existing atlases derived from meta-analysis, random, or structural criteria may not be optimal for these applications.
• There is a need for more suitable brain atlases tailored for network analysis.

Purpose of the Study:

• To develop and validate novel brain atlases for parcellating whole-brain resting-state fMRI data.
• To generate atlases suitable for use as nodes in network analysis and functional connectivity studies.
• To compare proposed atlas generation methods with existing state-of-the-art approaches.

Main Methods:

• Combined normalized cut (Ncut) for feature extraction from connectivity matrices and simple linear iterative clustering (SLIC) for brain parcellation.
• Proposed two group-level parcellation approaches: mean SLIC and two-level SLIC.
• Evaluated parcellations using metrics like spatial contiguity, functional homogeneity, and reproducibility across different conditions and subject groups.

Main Results:

• The proposed Ncut-SLIC approaches demonstrated robust clustering performance across various experimental conditions, including different weighting functions, sparsifying schemes, and confounding factors.
• Both mean SLIC and two-level SLIC achieved favorable results in terms of spatial contiguity, functional homogeneity, and reproducibility.
• The generated atlases were shown to be appropriate for network analysis applications.

Conclusions:

• The developed method effectively generates appropriate brain atlases for network analysis by combining Ncut and SLIC.
• The proposed group-level approaches offer reliable methods for creating standardized brain nodes.
• The publicly available data and code facilitate further research in brain network analysis.