Reproducible and Individualized Striatal Parcellation Based on Multi-Level Contrastive Learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Functional magnetic resonance imaging (fMRI)-based parcellation is crucial for understanding brain organization.
• Current methods lack reproducibility and struggle with individual-specific brain features due to noise and independent analysis.
• Existing approaches often result in poorly defined, discrete brain parcels.

Purpose of the Study:

• To develop a novel method for reproducible and individualized striatal parcellation using fMRI data.
• To address limitations of existing clustering-based parcellation techniques.
• To enhance the understanding of brain organization and individual differences in striatal function.

Main Methods:

• Proposed a Multi-Level Contrastive Learning-Based Graph Convolutional Network (MCL-GCN).
• Integrated spatial information using a spatial graph convolutional network for functional connectivity.
• Employed multi-level contrastive learning (voxel and individual) and a homogeneity loss.

Main Results:

• MCL-GCN demonstrated superior reproducibility and spatial continuity compared to existing methods.
• Achieved comparable functional homogeneity of parcels.
• Identified topological features correlated with cognitive behaviors, highlighting individual-specific information capture.

Conclusions:

• The MCL-GCN method offers a significant advancement in reproducible and individualized brain parcellation.
• The model effectively captures both shared and individual-specific brain features.
• This approach has the potential to deepen our understanding of striatal function and its relation to cognition.