Hierarchical graph-guided contextual representation learning for Neurodegenerative pattern recognition in MRI.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Neurodegenerative (ND) diseases are autoimmune disorders affecting the central nervous system.
• Deep learning shows promise in medical imaging but requires interpretability for clinical trust.
• Lesions in ND diseases present complex spatial patterns challenging traditional models.

Purpose of the Study:

• To develop an interpretable deep learning classifier for common autoimmune neurodegenerative diseases.
• To improve the capture of local and global relationships in brain MRI data for diagnosis.
• To enhance clinical acceptance of AI in neurodegenerative disease diagnostics.

Main Methods:

• A Residual Graph Neural Network enhanced Vision Transformer (RG-ViT) was developed.
• MRI data was represented as a graph of interconnected patches for analysis.
• Residual connections were integrated into the GNN framework to preserve features and improve message passing.

Main Results:

• The RG-ViT achieved high accuracy in detecting multiple sclerosis (98.7%), Parkinson's disease (99.6%), and Alzheimer's disease (99.1%).
• The model demonstrated strong generalizability with an F1 score of 99.2% on a combined dataset for global ND disease classification.
• The approach effectively addressed spatial disconnection issues in patch-based MRI analysis.

Conclusions:

• The RG-ViT model offers a highly accurate and interpretable solution for diagnosing autoimmune neurodegenerative diseases.
• This interpretable AI approach can build confidence among medical professionals for clinical application.
• The RG-ViT framework shows significant potential for advancing AI-driven diagnostics in neurology.