Multimodal connectivity-based cortical segmentation with graph neural networks.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Manual segmentation of the brain cortex from MRI is time-consuming and requires expertise.
• Automated segmentation methods are needed to improve efficiency and accuracy.
• Graph Neural Networks (GNNs) show potential for complex data analysis tasks.

Purpose of the Study:

• To evaluate the performance of different GNN architectures for brain cortex segmentation.
• To investigate the impact of multimodal data (sMRI and dMRI) on segmentation accuracy.
• To compare GNN-based segmentation with FreeSurfer for predicting demographic/clinical data.

Main Methods:

• Trained three GNN architectures: Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Graph U-Net.
• Utilized structural MRI (sMRI) and diffusion MRI (dMRI) data for multimodal segmentation.
• Evaluated segmentation performance using FreeSurfer-derived labels and Dice scores.
• Compared GNN and FreeSurfer segmentation for demographic/clinical data prediction.

Main Results:

• The GAT architecture achieved Dice scores competitive with existing non-graph methods.
• Incorporating structural connectivity from dMRI significantly improved segmentation accuracy compared to sMRI alone.
• GNN models trained on combined sMRI and dMRI attributes outperformed those trained solely on sMRI.
• Neither GNN-based nor FreeSurfer segmentation showed superiority in predicting demographic/clinical data.

Conclusions:

• GNNs, especially GAT, are effective tools for automated brain cortex segmentation.
• Multimodal data integration (sMRI and dMRI) enhances the performance of GNN-based segmentation.
• GNN and FreeSurfer approaches demonstrated comparable utility in predicting demographic/clinical outcomes.