Knowledge Distillation Guided Interpretable Brain Subgraph Neural Networks for Brain Disorder Exploration.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Brain disorder analysis benefits from neuroimaging but lacks mechanistic insights and interpretability.
• Current artificial intelligence methods for diagnosis often struggle with limited data and do not explore underlying pathogenic mechanisms.
• Graph Neural Networks (GNNs) show promise for analyzing complex, structured data, including molecular graphs.

Purpose of the Study:

• To develop an interpretable GNN-based method for brain disorder diagnosis using neuroimaging data.
• To address data scarcity and improve diagnostic efficiency by leveraging knowledge distillation (KD).
• To identify specific brain regions and functional connectivities associated with disorders.

Main Methods:

• Brain neuroimaging data was modeled into graph-structured data.
• Knowledge distillation (KD) guided brain subgraph neural networks were proposed.
• Discriminative subgraphs were extracted to identify abnormal brain connectivities.

Main Results:

• The proposed method demonstrated superior prediction accuracy for Parkinson's disease (PD) and attention-deficit/hyperactivity disorder (ADHD) compared to existing brain graph analysis techniques.
• Extracted discriminative subgraphs provided interpretable results consistent with medical research.
• Knowledge distillation effectively alleviated the problem of insufficient training data.

Conclusions:

• The KD-guided brain subgraph neural network approach offers an effective and interpretable method for brain disorder analysis.
• This method enhances diagnostic accuracy and provides insights into the pathogenic mechanisms of neurological disorders.
• The findings encourage further exploration of GNNs and KD in neuroimaging for deeper understanding of brain disorders.