Self-supervised learning of neighborhood embedding for longitudinal MRI

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Deep learning models represent Magnetic Resonance Imaging (MRI) as latent features for downstream tasks.
• Existing methods improve performance by associating latent representations with factors of interest, such as brain aging.
• Previous work encoded brain aging using self-supervised learning on longitudinal MRIs.

Purpose of the Study:

• To refine latent representations of brain aging by replacing linear modeling with neighborhood-consistent, age-consistent, and progression-consistent embeddings.
• To introduce Longitudinal Neighborhood Embedding (LNE) for improved analysis of brain aging trajectories.
• To develop a computationally tractable mini-batch sampling strategy for LNE.

Main Methods:

• Longitudinal Neighborhood Embedding (LNE) was developed to ensure age-consistency and progression-consistency in local neighborhoods of the latent space.
• A mini-batch sampling strategy was proposed to approximate global neighborhoods efficiently.
• LNE was evaluated on three downstream tasks: chronological age prediction, Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI) classification, and alcohol consumption impact assessment.

Main Results:

• LNE demonstrated superior accuracy on downstream tasks compared to existing self-supervised methods.
• Visualizations of smooth trajectory vector fields confirmed the method's effectiveness in capturing brain aging dynamics.
• The method successfully predicted chronological age, distinguished between Normal Control (NC), AD, sMCI, and pMCI, and identified alcohol consumption patterns.

Conclusions:

• Longitudinal Neighborhood Embedding (LNE) effectively extracts information related to brain aging from MRI data.
• The method shows promise for studying the impact of neurodegenerative disorders and substance use on brain aging.
• LNE offers a powerful new tool for analyzing longitudinal neuroimaging data.