Decoupling Alzheimer Disease Pathologic Abnormalities at PET with Improved Clinical Relevance by Interpretable...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Alzheimer Disease Research

Background:

• Template-based PET metrics for Alzheimer disease (AD) amyloid-β (Aβ) and tau burden oversimplify data, potentially causing imaging-clinical discordance.
• Current artificial intelligence (AI) methods offer richer patterns but often lack biological interpretability.
• There is a need for interpretable AI biomarkers in AD diagnostics.

Purpose of the Study:

• To develop and validate an interpretable deep-learning framework for PET imaging in Alzheimer disease.
• To generate a clinically meaningful AI biomarker by separating AD-specific abnormalities from physiologic uptake using pathophysiologic constraints.
• To enhance the understanding of disease heterogeneity and improve imaging-clinical correlations.

Main Methods:

• Retrospective analysis of Aβ and tau PET scans from multiple cohorts (ADNI, AIBL, GAAIN, author's center).
• Development of an adversarial decomposition learning (ADL) network to generate voxel-level pathologic maps and an AD adversarial decomposition (ADAD) score.
• Evaluation of discriminatory performance using AUC and clinical relevance using cognitive, volumetric, CSF, and neuropathologic measures.

Main Results:

• The ADL framework achieved high discriminatory performance with AUCs of 0.94 for Aβ and 0.98 for tau in external testing.
• Interpretable pathologic attribution maps generated by ADL correlated well with expert rankings (Aβ: ρ=0.79, tau: ρ=0.63).
• The ADAD score showed independent associations with cognitive outcomes and hippocampal atrophy, outperforming traditional metrics in clinical relevance.

Conclusions:

• Pathophysiologically constrained ADL provides interpretable, personalized pathologic maps for AD.
• The AI-derived ADAD score offers a clinically meaningful biomarker linking PET abnormalities to multimodal clinical measures.
• This framework improves the interpretability and clinical utility of AI in Alzheimer disease neuroimaging.