Surface-based Multi-Axis Longitudinal Disentanglement Using Contrastive Learning for Alzheimer's Disease.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate modeling of Alzheimer's Disease (AD) progression is crucial for understanding neuropathology.
• Neuroimaging analysis often struggles to differentiate disease effects from normal aging.
• Current deep learning methods for disentangling longitudinal data have limitations in capturing AD's heterogeneity.

Purpose of the Study:

• To develop a novel Surface-based Multi-axis Disentanglement framework for modeling Alzheimer's Disease progression.
• To enhance the capacity of deep learning models to capture the multifaceted nature and diverse trajectories of AD.
• To improve the separation of aging effects from disease-specific alterations in longitudinal neuroimaging data.

Main Methods:

• Proposed a Surface-based Multi-axis Disentanglement framework using Autoencoder networks.
• Implemented a longitudinal contrastive loss with self-supervision to assign data trajectories to multiple disease axes without explicit labels.
• Evaluated the model on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset (N=1321).

Main Results:

• The multi-axis model demonstrated superior performance in distinguishing between cognitively normal (CN), mild cognitive impairment (MCI), and AD subjects.
• Achieved improved classification of stable MCI versus converting MCI and Amyloid status compared to single-axis models.
• An ablation study confirmed the effectiveness of the contrastive loss in capturing complex AD progression patterns.

Conclusions:

• The proposed multi-axis disentanglement framework effectively models the heterogeneous progression of Alzheimer's Disease.
• This approach offers a more nuanced understanding of disease trajectories compared to traditional single-axis methods.
• The framework shows promise for improving differential diagnosis and tracking disease progression in AD research.