Explainable AI Points to White Matter Hyperintensities for Alzheimer's Disease Identification: a Preliminary Study

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neurology

Background:

• Deep learning (DL) models are effective for classification but lack clinical trust due to their "black box" nature.
• Explainable AI (XAI) aims to make DL model decisions transparent and understandable.
• Alzheimer's disease (AD) diagnosis can benefit from advanced computational methods.

Purpose of the Study:

• To investigate the "black box" problem in DL-based Alzheimer's disease classification.
• To apply XAI methods to understand how DL models identify AD from neuroimaging data.
• To explore the role of white matter hyperintensities (WMHs) in DL-driven AD classification.

Main Methods:

• Utilized a pre-trained DL model on MR scans from the OASIS-3 cohort (n=251) comprising early-stage AD dementia and healthy controls (HC).
• Applied the Occlusion Sensitivity XAI method to quantify relevance (RV) of brain tissues in classification.
• Compared RV values in WMH lesions versus healthy tissues for correctly and incorrectly classified AD/HC cases.

Main Results:

• The DL model achieved good performance (AUC: 0.82, TPR: 0.78, TNR: 0.81) in classifying AD vs. HC.
• XAI analysis indicated that the DL model preferentially utilized lesioned brain areas (WMHs) for AD identification.
• A statistically significant difference in WMH contribution was observed for AD recognition, unlike in HC cases (p=0.27).

Conclusions:

• DL models can be trained to utilize known clinical information, such as WMHs, for AD classification.
• Reinforces the significance of WMHs as a neuroimaging biomarker for AD dementia.
• Findings enhance trust in DL approaches for clinical applications in neurodegenerative disease diagnosis.