Adversarially Trained Persistent Homology Based Graph Convolutional Network for Disease Identification Using Brain...

Area of Science:

• Neuroscience
• Graph Theory
• Machine Learning

Background:

• Brain diseases cause network alterations.
• Graph convolutional networks (GCNs) analyze brain networks but often miss topological information and are vulnerable to perturbations.
• Existing GCNs focus on region features, ignoring crucial topological and connectivity patterns.

Purpose of the Study:

• To develop a robust and accurate method for classifying brain diseases using neuroimaging data.
• To enhance graph convolutional networks (GCNs) by integrating topological features for improved brain disease identification.
• To address the vulnerability of current GCNs to network property perturbations in brain disease diagnosis.

Main Methods:

• Constructed brain functional/structural connectivity using neuroimaging data.
• Developed an adversarially trained persistent homology-based graph convolutional network (ATPGCN).
• Integrated persistent homology features with GCN readout features for individual-level representation and simulated adversarial perturbations for robustness testing.

Main Results:

• The proposed ATPGCN method demonstrated superior performance in disease identification across three independent datasets.
• ATPGCN outperformed existing classification methods in accurately classifying brain diseases.
• The model proved robust against minor perturbations in brain network architecture, enhancing diagnostic reliability.

Conclusions:

• The adversarially trained persistent homology-based graph convolutional network (ATPGCN) offers a powerful new tool for brain disease classification.
• Integrating topological information and adversarial training significantly improves the accuracy and robustness of brain network analysis.
• This approach enhances the reliability of neuroimaging-based diagnosis for various brain diseases.