Atlas-free Brain Network Transformer.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Medical Imaging

Background:

• Current brain network analysis relies on fixed atlases, causing spatial misalignment and functional heterogeneity.
• These limitations hinder the reliability and interpretability of brain network studies.

Purpose of the Study:

• To develop a novel atlas-free brain network transformer (BNT) for individualized brain parcellations.
• To overcome the limitations of traditional atlas-based approaches in brain network analysis.

Main Methods:

• Utilized subject-specific resting-state fMRI data for individualized brain parcellations.
• Computed ROI-to-voxel connectivity features in a standardized voxel-based feature space.
• Employed the BNT architecture to generate comparable subject-level embeddings.

Main Results:

• The atlas-free BNT outperformed state-of-the-art atlas-based methods in sex classification and brain-connectome age prediction.
• Demonstrated significant improvements in precision, robustness, and generalizability of brain network analyses.
• The approach showed superior performance compared to elastic net, BrainGNN, Graphormer, and the original BNT.

Conclusions:

• The atlas-free BNT offers a more reliable and interpretable approach to brain network analysis.
• This advancement has the potential to enhance neuroimaging biomarkers for personalized precision medicine.
• The method provides a robust foundation for improved clinical diagnostic tools.