Learning disentangled representations to harmonize connectome network measures

Area of Science:

• Neuroscience
• Medical Imaging
• Data Science

Background:

• Connectome network metrics are crucial for understanding brain function and disease.
• Current metrics are susceptible to site-specific variations due to imaging protocols and software choices.
• These variations hinder multi-site studies and reliable disease association analyses.

Purpose of the Study:

• To develop a method for harmonizing connectome network measures across different study sites.
• To create site-invariant representations of brain connectomes using advanced machine learning.
• To improve the reliability of connectome data for multi-site brain research.

Main Methods:

• Utilized a conditional variational autoencoder (VAE) framework.
• Applied disentanglement techniques to separate site-specific factors from biological features in connectome data.
• Trained the model on data from 823 patients across two study sites, focusing on aging.

Main Results:

• Successfully generated site-invariant representations of the connectome.
• Demonstrated significant harmonization of network measures across sites.
• Maintained robust associations with biological factors like age and sex.

Conclusions:

• Latent representations derived from the conditional VAE effectively harmonize network measures.
• This approach provides robust metrics for multi-site brain network analysis.
• The findings pave the way for more reliable and reproducible connectome research across diverse datasets.