Cross-domain Fiber Cluster Shape Analysis for Language Performance Cognitive Score Prediction

Area of Science:

• Neuroimaging
• Computer Graphics
• Computational Neuroscience

Background:

• Shape analysis is crucial for understanding morphology and function in computer graphics and brain imaging.
• Brain white matter connections' shape may correlate with cognitive functions.
• Diffusion magnetic resonance imaging (dMRI) tractography reconstructs these connections.

Purpose of the Study:

• Investigate the predictive relationship between the 3D shape of brain white matter connections and human cognitive function, specifically language performance.
• Introduce and evaluate a novel framework, SFFormer, for this prediction task.

Main Methods:

• Reconstructed brain connections as 3D point sequences using dMRI tractography.
• Extracted 12 shape descriptors alongside traditional dMRI connectivity and microstructure features.
• Developed the Shape-fused Fiber Cluster Transformer (SFFormer) model with a multi-head cross-attention feature fusion module.

Main Results:

• The transformer-based SFFormer model effectively predicts subject-specific language performance.
• Fusion of shape, microstructure, and connectivity features significantly improves prediction accuracy.
• Demonstrated the informativeness of both the SFFormer model and its feature fusion approach.

Conclusions:

• The 3D shape of brain connections is a significant predictor of human language function.
• The SFFormer framework offers a powerful tool for integrating multimodal neuroimaging data for cognitive function prediction.