GMN4AD: Graph matching network boosts Alzheimer's diagnosis accuracy across MRI scanners
New AI model adapts on the fly to different MRI machines, cutting diagnostic errors by 15%.
Alzheimer's Disease (AD) affects millions of older adults, and early diagnosis—especially at the mild cognitive impairment (MCI) stage—is critical. Structural MRI (sMRI) is a key modality for detecting brain atrophy, but traditional graph-based models treat each brain scan independently and struggle with variability between MRI scanners and sites. This heterogeneity, known as domain shift, limits diagnostic accuracy in real-world clinical settings.
To solve this, Chen Zhao and colleagues from the University of Iowa developed GMN4AD (Graph Matching Network for Alzheimer's Disease Diagnosis). Unlike conventional methods, GMN4AD uses graph matching to capture cross-graph relationships between brain scans, enhancing diagnostic precision. It also introduces a test-time domain adaptation strategy that employs contrastive learning to adjust for inter-site and inter-modality differences during inference—no need to retrain for each new scanner. Experiments on three public datasets (including ADNI) show GMN4AD significantly outperforms existing techniques in classifying AD, MCI, and healthy controls. The model achieves higher accuracy and robustness, making it a practical step toward standardized AI-assisted AD screening across diverse imaging centers.
- GMN4AD uses graph matching to model interactions between brain graphs, improving over independent graph analysis.
- Test-time domain adaptation with contrastive learning handles scanner/site variations without retraining.
- Outperforms state-of-the-art on three public AD datasets, including ADNI, for AD and MCI classification.
Why It Matters
A generalizable AI diagnostic tool that works across different MRI machines could accelerate early Alzheimer's screening worldwide.