Learning Cortical Parcellations Using Graph Neural Networks.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Deep learning, particularly convolutional neural networks, has advanced magnetic resonance imaging (MRI) applications.
• Brain imaging analysis increasingly utilizes network- and surface-based approaches better suited for graph representations.
• Existing methods may not fully leverage the potential of graph-based analyses for cortical segmentation.

Purpose of the Study:

• To develop a general-purpose cortical segmentation method using graph neural networks (GNNs).
• To leverage resting-state functional connectivity (rsFC) from MRI for automated cortical parcellation.
• To evaluate the performance and generalizability of the proposed GNN-based segmentation approach.

Main Methods:

• Applied graph neural networks to cortical surface segmentation using rsFC features from MRI.
• Trained GNNs on existing cortical parcellations to learn discrete maps of the human neocortex.
• Optimized algorithm type, network architecture, and training features for segmentation accuracy.

Main Results:

• GNNs accurately learn low-dimensional representations of functional brain connectivity.
• The proposed method successfully maps cortices of new datasets.
• Achieved a mean classification accuracy of 79.91% compared to a prior parcellation.

Conclusions:

• Graph neural networks provide an effective framework for cortical segmentation using MRI-derived functional connectivity.
• The method demonstrates accurate and generalizable parcellation of the human neocortex.
• Hyperparameter choices significantly influence the performance of GNN-based brain segmentation.