Optimizing Connectivity-Driven Brain Parcellation Using Ensemble Clustering.

Area of Science:

• Neuroscience
• Computational Biology
• Network Science

Background:

• Constructing a unified brain atlas is challenging due to individual variability in brain connectivity.
• Existing methods often lack flexibility or fail to capture the average topological structure of the connectome.

Purpose of the Study:

• To develop a topologically optimal, connectivity-based brain atlas by aggregating individual parcellations.
• To balance flexibility in individual brain representations with an intuitive average topological structure.

Main Methods:

• Utilized a dense connectivity representation and graph-based hierarchical parcellation for individual brains.
• Employed the hard ensemble (HE) algorithm to aggregate individual parcellations into a consensus atlas, approximating a pseudo-Karcher mean.
• Assessed computational stability, graph structure preservation, and biological relevance using metrics like edge weight divergence, symmetry, network stability, and a sex classification task.

Main Results:

• The ensemble parcellation demonstrated high stability across subject sampling.
• Outperformed anatomical atlases and other connectome-based parcellations in network classification tasks.
• Preserved global connectome properties, exhibited symmetry comparable to anatomical atlases, and showed high spatial contiguity without explicit priors.

Conclusions:

• The proposed HE-based ensemble parcellation offers a robust and biologically relevant method for creating unified brain atlases.
• This approach advances the field of connectomics by providing a stable and accurate representation of brain network topology.