An Unsupervised XAI Framework for Dementia Detection with Context Enrichment.

Area of Science:

• Neuroimaging and Artificial Intelligence
• Clinical Decision Support Systems
• Explainable Artificial Intelligence (XAI)

Background:

• Explainable Artificial Intelligence (XAI) enhances the transparency and trustworthiness of AI predictions in clinical decision support systems, particularly for brain imaging analysis.
• Limited validation of XAI explanation quality hinders its clinical adoption, necessitating robust evaluation frameworks.
• Convolutional Neural Networks (CNNs) are powerful tools for analyzing brain MRI scans but require interpretable outputs for clinical trust.

Purpose of the Study:

• To introduce and evaluate a framework for assessing XAI methods in dementia research by combining neuroanatomical features with CNN relevance maps.
• To refine XAI explanation spaces and explore different approaches for generating clinically relevant explanations for AI-driven diagnostic tools.
• To determine the potential of validated XAI methods in improving the diagnostic efficiency of AI-based decision support systems for dementia.

Main Methods:

• A CNN was trained on brain MRI scans from 3253 participants across six cohorts (ADNI, AIBL, DELCODE, DESCRIBE, EDSD, NIFD).
• Clustering analysis used morphological features as proxy ground truth to benchmark explanation space configurations.
• Three post-hoc XAI methods were implemented: model simplification, explanation-by-example, and textual explanations, followed by qualitative clinical evaluation.

Main Results:

• Morphology-enriched explanation spaces demonstrated improved clustering performance, enhancing both homogeneity and completeness.
• Model simplification explanations effectively distinguished between participants who would convert to dementia and those who remained stable.
• Explanation-by-example visualized potential cognitive trajectories, while textual explanations provided rule-based summaries of pathological findings.

Conclusions:

• The study successfully refined XAI explanation spaces and demonstrated the utility of various explanation generation approaches.
• The evaluated XAI methods show promise for enhancing diagnostic efficiency within AI-based decision support systems for dementia research.
• Clinical assessments confirmed the potential of these XAI techniques, highlighting both challenges and opportunities for future applications.