Graph Neural Network Learning on the Pediatric Structural Connectome.

Area of Science:

• Neuroscience
• Machine Learning
• Brain Imaging

Background:

• Sex classification using structural connectome data is a key benchmark in brain graph learning.
• Graph neural networks (GNNs) excel at learning from graph data, but their use in pediatric populations is underexplored.

Purpose of the Study:

• To investigate the capability of GNNs to learn pediatric connectomic patterns.
• To explore training techniques and architectural designs for GNNs in pediatric sex classification.
• To compare GNN performance against other machine learning models in both adult and pediatric datasets.

Main Methods:

• Utilized adult (BRIGHT) and pediatric (HCP-D) connectome datasets.
• Trained GNN models (GCN simple, GCN residual), a deep neural network (MLP), and standard ML models (RF, SVM).
• Conducted architecture exploration (depth, pooling, skip connections) and assessed adversarial robustness.

Main Results:

• GNNs outperformed other models in both adult and pediatric populations.
• Adult GNNs achieved 85.1% accuracy for adult sex classification.
• Augmenting pediatric data with adult data yielded comparable accuracy (83.0% pediatric, 81.3% adult).
• Simple GCN demonstrated superior adversarial robustness compared to residual GCN and MLP.

Conclusions:

• GNNs hold significant potential for understanding sex-specific neurological development and disorders.
• Data augmentation is crucial for overcoming challenges with small pediatric datasets.
• Tradeoffs exist between GNN model complexity, accuracy, and adversarial robustness.