Longitudinal Feature Disentanglement With Cross-Time Contrastive Learning for Alzheimer's Disease Diagnosis

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Longitudinal structural MRI is crucial for tracking brain atrophy in Alzheimer's Disease (AD).
• Analyzing longitudinal data is complex due to inter-temporal correlations, noise, and redundancy.
• Existing methods struggle to effectively model these dynamic changes for accurate AD progression prediction.

Purpose of the Study:

• To develop a novel deep learning model for enhanced Alzheimer's Disease diagnosis and progression prediction.
• To address the challenges of modeling complex temporal dynamics and noise in longitudinal brain imaging data.
• To disentangle stable, disease-specific features from variable, noise-related features in brain scans.

Main Methods:

• Proposed a Longitudinal Feature Disentanglement Network using a 3D convolutional encoder with weight sharing.
• Employed a Variational Autoencoder (VAE) to disentangle spatial representations into stable and marginal features.
• Incorporated cross-time-point contrastive learning and a longitudinal feature reconstruction module.

Main Results:

• The proposed network achieved superior performance in Alzheimer's Disease diagnosis compared to state-of-the-art methods.
• Demonstrated high accuracy in predicting Mild Cognitive Impairment (MCI) to AD conversion.
• Visualization analysis confirmed the model's ability to identify key brain regions and interpret its decisions.

Conclusions:

• The Longitudinal Feature Disentanglement Network effectively analyzes longitudinal sMRI data for AD.
• The method shows significant potential for improving early diagnosis and predicting AD progression.
• This approach offers a robust framework for leveraging complex neuroimaging data in AD research.