Multimodal Self-Supervised Learning for Early Alzheimer's: Cross-Modal MRI-PET, Longitudinal Signals, and Site...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomarker Discovery

Background:

• Alzheimer's disease (AD) diagnosis is challenging due to diverse imaging, scanner variability, and disease progression.
• Robust multimodal approaches integrating structural, molecular, and longitudinal data are needed for clinical settings.

Purpose of the Study:

• To develop a multimodal self-supervised learning (SSL) framework for Alzheimer's disease (AD) identification.
• To create transferable and domain-robust biomarkers for early AD diagnosis and prediction of mild cognitive impairment (MCI) to AD conversion.

Main Methods:

• A multimodal SSL approach learning joint MRI and PET feature representations.
• Integration of contrastive learning, scanner harmonization, and missing modality-aware fusion.
• Evaluation on six public datasets: ADNI, OASIS-3, AIBL, BioFINDER, TADPOLE, and MIRIAD.

Main Results:

• State-of-the-art performance on benchmark tasks, achieving 93.0% BACC and 0.96 AUC for AD vs. CN classification on ADNI.
• Strong cross-cohort generalizability (78.0% BACC on OASIS-3, 77.5% BACC on AIBL).
• Effective longitudinal prognosis for MCI to AD conversion (0.85 AUC, 0.82 C-index on TADPOLE) and high test-retest consistency (ICC = 0.91 on MIRIAD).

Conclusions:

• The multimodal SSL framework provides effective, transferable, and domain-robust biomarkers for early AD diagnosis.
• The approach accurately predicts MCI to AD progression with strong cross-dataset generalization.
• This method addresses real-world cohort diversity challenges in neuroimaging analysis.