SOM2LM: Self-Organized Multi-Modal Longitudinal Maps.

Area of Science:

• Artificial Intelligence
• Neuroimaging
• Biomedical Data Analysis

Background:

• Longitudinal neuroimaging studies provide complementary data for understanding disease progression.
• Alzheimer's disease (AD) involves early amyloid plaque buildup (visualized by PET) and later brain atrophy (seen in MRI).
• Accurate modeling of multi-modal longitudinal data is crucial for disease trajectory prediction.

Purpose of the Study:

• To propose an interpretable self-supervised model, Self-Organized Multi-Modal Longitudinal Maps (SOM2LM), for analyzing multi-modal longitudinal neuroimaging data.
• To encode each imaging modality into a 2D self-organizing map (SOM) where one dimension represents disease abnormality.
• To regularize across modalities to capture the temporal order of abnormality detection.

Main Methods:

• Developed SOM2LM, a self-supervised model utilizing 2D self-organizing maps for each neuroimaging modality.
• Applied SOM2LM to longitudinal T1w MRI and amyloid PET data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort (N=741).
• Evaluated model performance on downstream tasks including cross-modality prediction and joint-modality prediction of disease progression.

Main Results:

• SOM2LM generated interpretable latent spaces that effectively characterize disease abnormality in Alzheimer's disease.
• The model achieved higher accuracy compared to state-of-the-art methods in predicting amyloid status from T1w-MRI.
• Demonstrated superior performance in joint-modality prediction of mild cognitive impairment converters to AD using both MRI and amyloid PET.

Conclusions:

• SOM2LM offers an effective and interpretable approach for modeling multi-modal longitudinal neuroimaging data in Alzheimer's disease research.
• The model enhances the understanding of disease progression by integrating complementary information from different imaging modalities.
• SOM2LM shows significant potential for improving diagnostic and prognostic accuracy in Alzheimer's disease.