SurfGNN: A robust surface-based prediction model with interpretability for coactivation maps of spatial and cortical...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Current brain surface models often fail to capture regional variations at the cortical feature level.
• Graph neural networks (GNNs) struggle with high-density graph structures common in cortical surface meshes.

Purpose of the Study:

• To propose an interpretable prediction model, Surface Graph Neural Network (SurfGNN), for brain surface analysis.
• To address challenges in modeling regional heterogeneity and high-density graph structures in cortical data.

Main Methods:

• SurfGNN utilizes topology-sampling learning (TSL) and region-specific learning (RSL) to manage cortical features at multiple scales.
• A score-weighted fusion (SWF) method merges nodal representations for prediction.
• The model was applied to neonatal brain age prediction using 481 subjects' harmonized MR images.

Main Results:

• SurfGNN achieved a mean absolute error (MAE) of 0.827 ± 0.056 in postmenstrual weeks, outperforming state-of-the-art methods by at least 9.0%.
• The model generated feature-level activation maps, identifying key regional variations contributing to predictions.
• The proposed method effectively handles sparse graph representations of cortical surfaces and regional heterogeneity.

Conclusions:

• SurfGNN offers a superior approach for brain surface-based prediction tasks, particularly in developmental assessments.
• The model's interpretability through activation maps aids in understanding morphometric contributions to age prediction.
• This work advances GNN applications in neuroimaging by addressing complex graph structures and regional variability.