Functional Connectivity Prediction With Deep Learning for Graph Transformation.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Network Science

Background:

• Inferring functional connectivity (FC) from structural connectivity (SC) is crucial for understanding brain networks and mental health.
• Existing models often use linear approaches or heuristics, failing to capture the complex, nonlinear, and stochastic nature of the SC-FC relationship.
• Factors like age and health significantly influence the SC-FC relationship and require consideration.

Purpose of the Study:

• To develop a novel framework, SF-GAN, for accurately mapping structural connectivity (SC) to functional connectivity (FC).
• To incorporate subject-specific metafeatures and model the inherent stochasticity in the SC-FC relationship.
• To provide interpretability for identifying influential subgraphs and metafeatures within the SC-FC mapping.

Main Methods:

• Developed a novel SC-to-FC generative adversarial network (SF-GAN) framework.
• Utilized a graph neural network-based generative model with edge convolution and deconvolution layers.
• Integrated subject profile information using sparse-regularized layers for metafeature selection.
• Proposed a post hoc explainer based on multilevel edge-correlation-guided graph clustering for subgraph analysis.

Main Results:

• The SF-GAN framework significantly outperforms existing state-of-the-art methods in mapping SC to FC.
• The model effectively captures the nonlinear and stochastic nature of the SC-FC relationship.
• The post hoc explainer successfully identifies influential subgraphs in SC and their impact on FC, along with important metafeatures.

Conclusions:

• The proposed SF-GAN provides a powerful and interpretable method for mapping structural to functional brain connectivity.
• This approach advances the understanding of complex brain network dynamics and has implications for diagnosing and treating mental diseases.
• The framework's ability to integrate metafeatures and provide explainability offers new avenues for personalized neuroscience research.