A hierarchical Bayesian brain parcellation framework for fusion of functional imaging datasets

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Statistical Modeling

Background:

• Brain organization studies benefit from diverse imaging datasets (task-based, resting-state).
• Single-dataset parcellations are limited by dataset-specific biases.
• Integrating complementary information is crucial for comprehensive brain mapping.

Purpose of the Study:

• To develop a hierarchical Bayesian framework for probabilistic brain parcellation.
• To overcome limitations of single-dataset brain atlases by combining multiple data types.
• To create a more accurate and robust population-based human brain atlas.

Main Methods:

• A hierarchical Bayesian framework with spatial arrangement and dataset-specific emission models.
• Probabilistic modeling of voxel-to-parcel assignments.
• Integration of numerous task-based and resting-state neuroimaging datasets.

Main Results:

• The framework successfully combined information from diverse datasets to create a superior population-based atlas of the human cerebellum.
• The novel atlas outperformed atlases derived from single datasets.
• Individual brain parcellations generated using only 10 minutes of data surpassed existing group atlases in accuracy.

Conclusions:

• Hierarchical Bayesian modeling offers a powerful approach to integrate multimodal neuroimaging data for brain parcellation.
• This framework enables the creation of more accurate population-level and individual-level brain atlases.
• The method demonstrates the potential for improved understanding of brain organization through data fusion.