LSOR: Longitudinally-Consistent Self-Organized Representation Learning.

Area of Science:

• Neuroimaging
• Deep Learning
• Machine Learning

Background:

• Interpretability of deep learning models in longitudinal brain MRI analysis is crucial.
• Self-organizing maps (SOM) visualize high-dimensional latent spaces but struggle with stability and clinical relevance in self-supervised settings.
• Existing SOM methods do not inherently capture clinically significant information like brain age.

Purpose of the Study:

• To introduce a novel self-supervised SOM approach for generating interpretable, high-dimensional representations from longitudinal brain MRIs.
• To stratify these representations by brain age without requiring demographic or cognitive data.
• To enhance the stability and clinical relevance of SOM in deep learning for neuroimaging.

Main Methods:

• Proposed Longitudinally-consistent Self-Organized Representation learning (LSOR), a self-supervised SOM method utilizing soft clustering for enhanced training stability.
• Developed a technique to align longitudinal MRI trajectories with SOM cluster reference vectors, stratifying the latent space by brain age.
• Applied LSOR to the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.

Main Results:

• LSOR successfully generated a stable and interpretable latent space stratified by brain age from longitudinal brain MRIs.
• The method demonstrated comparable or superior accuracy to state-of-the-art approaches on downstream tasks.
• Achieved high performance in classification (mild cognitive impairment progression) and regression (ADAS-Cog score prediction).

Conclusions:

• LSOR provides a robust and interpretable framework for analyzing longitudinal brain MRI data using deep learning.
• The approach effectively integrates brain age stratification into the representation learning process.
• LSOR offers a promising tool for advancing neuroimaging research and clinical applications, particularly in Alzheimer's disease.