A quantitatively interpretable model for Alzheimer's disease prediction using deep counterfactuals.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Medical Diagnostics

Background:

• Deep learning (DL) models predict Alzheimer's disease (AD) but lack interpretability.
• Counterfactual reasoning offers visual explanations but needs quantitative validation.
• Current methods struggle to intuitively link visual maps to neuroscientific validity.

Purpose of the Study:

• To develop a framework for interpretable AD prediction using counterfactual reasoning.
• To quantitatively validate visual explanatory maps from DL models.
• To enhance understanding of brain status in AD progression.

Main Methods:

• Synthesized counterfactual-labeled structural MRIs using a novel framework.
• Transformed MRIs into gray matter density maps to measure volumetric changes in regions of interest (ROIs).
• Developed a lightweight linear classifier to boost ROI effectiveness and quantitative interpretation.

Main Results:

• Achieved predictive performance comparable to existing DL methods.
• Generated an "AD-relatedness index" for each ROI, quantifying disease association.
• Demonstrated the framework's ability to provide intuitive understanding of brain status.

Conclusions:

• The proposed framework enhances the interpretability of AD prediction models.
• Quantitative features derived from counterfactual reasoning provide neuroscientific validity.
• The "AD-relatedness index" offers a valuable tool for assessing AD progression.