BDEC: Brain Deep Embedded Clustering Model for Resting State fMRI Group-Level Parcellation of the Human Cerebral...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) is crucial for understanding brain function.
• Existing group-level brain parcellation methods often rely on strong model assumptions.
• Developing robust, assumption-free parcellation techniques is essential for accurate brain network analysis.

Purpose of the Study:

• To develop a novel deep learning-based group-level brain parcellation method using rs-fMRI.
• To overcome limitations of previous parcellation approaches by minimizing model assumptions.
• To create a parcellation technique that enhances functional coherence and generalizability.

Main Methods:

• Proposed Brain Deep Embedded Clustering (BDEC), a deep clustering model.
• Implemented a specialized loss function to maximize inter-class separation and intra-class similarity.
• Applied BDEC to rs-fMRI data for group-level brain parcellation.

Main Results:

• BDEC demonstrated superior performance in functional homogeneity metrics compared to ten existing methods.
• Achieved favorable results in parcellation validity, downstream task performance, and generalization.
• Showcased improved handling of task inhomogeneity in brain parcellation.

Conclusions:

• BDEC effectively captures intrinsic functional brain properties, yielding reliable and generalizable parcellations.
• The BDEC model offers a valuable tool for brain network analysis and dimensionality reduction of rs-fMRI data.
• Contributes to a deeper understanding of the brain's functional organization.