ATPGCN

ATPGCN integrates graph convolutional networks (GCNs), persistent homology, and adversarial training to learn disease-specific representations from brain functional and structural connectivity for disease identification.

Key Features:

• Graph Convolutional Networks (GCNs): Uses GCNs to embed non-Euclidean brain network structures and capture complex connectivity patterns.
• Persistent Homology: Extracts topological features from algebraic topology that characterize global and local connectivity patterns in the connectome.
• Feature Integration: Combines persistent homology features with readout features from the global pooling layer of the GCN to form individual-level representations.
• Adversarial Training: Incorporates adversarial perturbations into training to improve model robustness against targeted disruptions.
• Risk ROI Perturbations: Simulates perturbations targeting clinically relevant risk regions-of-interest (ROIs) to evaluate and enhance resilience to network changes.
• Connectivity Construction: Constructs brain functional and structural connectivity matrices using various neuroimaging modalities.
• Robustness Evaluation: Assesses model stability under perturbations to ensure consistent performance despite minor alterations in network properties.

Scientific Applications:

• Disease identification: Classifies brain disorders by learning disease-specific connectivity representations from functional and structural connectomes.
• Cross-dataset validation: Demonstrated superior disease classification performance across three independent datasets compared to existing methods.
• Neuroimaging research and clinical diagnostics: Provides topological-informed and GCN-derived representations for studies and applications in neuroimaging and clinical diagnosis of brain disorders.

Methodology:

Construct brain functional/structural connectivity from various neuroimaging modalities; extract persistent homology features and combine them with readout features from the GCN global pooling layer to learn individual representations; incorporate adversarial perturbations targeting clinically relevant risk ROIs into the training loop to evaluate and improve robustness.