Adaptive Integration of Incomplete Multimodal 3D Neuroimaging for Alzheimer's Prediction and Biomarker Discovery

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Data Analysis

Background:

• Alzheimer's disease (AD) diagnosis is complex, requiring analysis of heterogeneous factors.
• Magnetic resonance imaging (MRI) and positron emission tomography (PET) are crucial noninvasive tools for assessing brain structure and function in AD.
• Existing multimodal neuroimaging approaches face challenges in adaptively integrating incomplete or varied data.

Purpose of the Study:

• To propose V3D-MMoE, an interpretable framework for adaptive integration of incomplete multimodal 3D neuroimaging data.
• To enhance Alzheimer's disease diagnosis prediction and facilitate biomarker discovery.
• To overcome limitations of prior methods in handling missing modalities and varying modality importance.

Main Methods:

• Developed V3D-MMoE, a framework utilizing a sparse mixture-of-experts formulation for adaptive modality importance.
• Implemented modality alignment to improve cross-modal learning between MRI and PET.
• Employed cross-encoders to dynamically manage missing neuroimaging data.

Main Results:

• V3D-MMoE demonstrated superior performance compared to state-of-the-art methods in predicting two-year Alzheimer's disease diagnosis using MRI and PET scans.
• Interpretability analyses identified subject-specific MRI and PET biomarkers aligned with known AD biology.
• Ablation studies confirmed the significant contribution of multimodal neuroimaging integration.

Conclusions:

• V3D-MMoE offers an effective and interpretable approach for Alzheimer's disease diagnosis using incomplete multimodal 3D neuroimaging.
• The framework's ability to handle missing data and adaptively integrate modalities represents a significant advancement.
• The identified biomarkers hold potential for future diagnostic and research applications in Alzheimer's disease.