Explainable multiplex graph propagational network with multimodal neuroimage integration for dementia subtype...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Dementia diagnosis relies on diverse data, with neuroimaging offering a minimally invasive approach.
• Machine learning, particularly graph neural networks (GNNs), shows promise in enhancing diagnostic accuracy by analyzing brain connectivity.
• Current GNN methods struggle to capture global brain connectivity and lack clinical explainability.

Purpose of the Study:

• To propose a novel, explainable method for diagnosing dementia subtypes using multimodal neuroimaging data.
• To address the limitations of existing GNNs in capturing global connectivity and maintaining model transparency.
• To develop a clinically applicable tool for dementia diagnosis.

Main Methods:

• Developed the Explainable Multiplex Graph Propagational Network (EMGPN) integrating multimodal neuroimages.
• Utilized multiplex graphs to represent both local and global brain connectivity across modalities.
• Employed a transparent architecture without hidden layers for enhanced explainability.

Main Results:

• EMGPN demonstrated an average performance improvement of 8.6% over existing methods in dementia subtype diagnosis.
• The model successfully generated explainable outputs, including region-specific modality contributions.
• Subtype-specific brain region importance maps were produced, aiding clinical interpretation.

Conclusions:

• EMGPN offers a robust and explainable approach to neuroimaging-based dementia diagnosis.
• The method enhances diagnostic performance while providing crucial transparency for clinical decision-making.
• EMGPN shows significant potential for clinical application in identifying dementia subtypes.