Self-Supervised Longitudinal Neighbourhood Embedding.

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Image Analysis

Background:

• Longitudinal MRI analysis tracks brain changes due to aging and disease.
• Machine learning on this data typically requires extensive, costly ground-truth labels.
• Self-supervised learning offers a promising alternative to reduce labeling dependency.

Purpose of the Study:

• To develop a novel self-supervised representation learning strategy for longitudinal MRI data.
• To reduce the reliance on manual annotations in analyzing brain structure and function changes.
• To effectively model subject progression and capture global and local brain morphology.

Main Methods:

• Proposed Longitudinal Neighborhood Embedding (LNE), a self-supervised method inspired by contrastive learning.
• Constructed graphs in latent space to ensure subject trajectory similarity and directional consistency.
• Applied LNE to longitudinal T1w MRIs from healthy subjects and the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort.

Main Results:

• LNE generated a smooth trajectory vector field, capturing global brain morphological changes.
• Demonstrated superior performance over existing self-supervised methods in downstream tasks.
• Successfully extracted information related to normal aging and neurodegenerative disorders.

Conclusions:

• LNE is a powerful self-supervised method for analyzing longitudinal neuroimaging data.
• The approach effectively models brain aging and the progression of neurological diseases.
• LNE offers a label-efficient strategy for extracting valuable insights from longitudinal MRIs.