Biomarker Investigation Using Multiple Brain Measures from MRI Through Explainable Artificial Intelligence in...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's Disease (AD) is the leading cause of dementia globally.
• Deep Learning (DL) shows promise for AD classification, but its reliance on biological indicators is unclear.
• Explainability of DL models is crucial for clinical trust and diagnostic applications.

Purpose of the Study:

• To compare a novel DL model (BC-GCN-SE) using structural connectivity with an established model (ResNet18) using structural MRI for AD classification.
• To introduce and utilize a novel Explainable Artificial Intelligence (XAI) metric to assess model adherence to known AD biomarkers.
• To explore the complementary insights provided by different neuroimaging modalities (MRI vs. connectivity) in DL models.

Main Methods:

• Developed a novel XAI metric based on gradient-weighted class activation mapping for quantitative explainability assessment.
• Compared BC-GCN-SE (structural connectivity) and ResNet18 (structural MRI) DL models on 132 brain parcels.
• Evaluated classification performance (true positive/negative rates) and analyzed explainability patterns against AD-relevant regions.

Main Results:

• Both models achieved satisfactory classification performance (ResNet18: TP 0.817, TN 0.816; BC-GCN-SE: TP 0.703, TN 0.738).
• XAI analysis identified distinct biomarker involvements: medial temporal lobe for ResNet18 and default mode network for BC-GCN-SE.
• Statistical tests confirmed the relevance of these identified brain regions.

Conclusions:

• Different neuroimaging modalities provide complementary information for DL models in AD classification.
• The developed XAI metric effectively measures model adherence to domain knowledge and biomarkers.
• Findings support the development of more comprehensive and trustworthy DL diagnostic support tools for neurodegenerative diseases.