Alzheimer's Imaging Consortium.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomarkers

Background:

• Alzheimer's disease (AD) diagnosis relies on multi-domain biomarkers.
• Machine learning models can benefit from integrating multiple data modalities.
• Early detection of AD and prediction of dementia progression are critical.

Purpose of the Study:

• To enhance deep learning model predictive power for AD using multi-modal neuroimaging data (T1-MRI and amyloid-PET).
• To capture both amyloid (A) and atrophy (N) brain patterns.
• To derive a dementia risk score (DRS) for predicting future dementia progression at early stages.

Main Methods:

• Utilized multi-modal neuroimaging data from ADNI 1, 2, and GO datasets.
• Trained classification models using T1-MRI and Amyloid-PET data with 5-fold cross-validation.
• Fine-tuned ResNet50 models pre-trained on MedicalNet for independent MRI and PET classification, then fused for multi-modal DRS.

Main Results:

• The fused multi-modal model achieved 97.29% balanced accuracy in classifying AD/CN, outperforming MRI-only (94.53%) and PET-only (86.53%) models.
• For predicting Mild Cognitive Impairment (MCI) to dementia progression, the fused model reached 74.59% balanced accuracy, slightly improving over single-modal models (MRI: 71.17%, PET: 71.79%).

Conclusions:

• Multi-modal deep learning models show significant potential for improving AD prediction accuracy and tracking disease progression.
• T1-MRI and Amyloid-PET data demonstrate complementary strengths in AD prediction.
• Further research is needed to optimize fusion strategies and explore other modalities for enhanced prediction of MCI progression.