Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
New AI Framework Decouples Alzheimer’s Pathology from Normal Brain Activity in PET Scans
Researchers developed an interpretable machine learning framework that separates disease-specific abnormalities from normal physiologic uptake in Alzheimer’s PET imaging, generating a clinically relevant biomarker more strongly associated with cognitive outcomes than standard metrics.
Background
PET imaging captures amyloid-β and tau accumulation in Alzheimer disease, but distinguishing true pathologic signal from normal physiologic uptake remains challenging. Standard quantification methods like Centiloid and CenTauRz correlate well with postmortem neuropathology but show weaker associations with patients’ actual cognitive decline. This mismatch highlights the need for biomarkers that better predict clinical outcomes.
Key Findings
- Adversarial decomposition learning (ADL) achieved diagnostic performance with AUC of 0.94 for amyloid-β and 0.98 for tau in distinguishing Alzheimer disease from cognitively normal individuals
- Generated pathologic attribution maps strongly correlated with expert ratings (Spearman ρ = 0.79 for amyloid-β and 0.63 for tau)
- ADAD score demonstrated larger effect sizes and stronger associations with cognitive decline and hippocampal atrophy than standard metrics
- Study analyzed 7,457 amyloid-β and 1,894 tau PET scans from 4,722 participants across multiple international cohorts
Why It Matters
While standard metrics show higher correlations with postmortem neuropathology, ADAD more directly links PET abnormalities with functional and structural clinical outcomes. This outcome-sensitive approach may improve diagnosis, disease monitoring, and therapeutic assessment in individual patients by providing personalized, interpretable pathologic maps aligned with clinical reality.
Limitations
The study uses retrospective data from established cohorts, which may limit generalizability. The comparison between postmortem accuracy and clinical relevance highlights potential trade-offs between neuropathologic correlation and clinical outcome prediction.
Original paper: Decoupling Alzheimer Disease Pathologic Abnormalities at PET with Improved Clinical Relevance by Interpretable Adversarial Decomposition Learning. — Radiology. 10.1148/radiol.252321
