Learning site-invariant features of connectomes to harmonize complex network measures

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biostatistics

Background:

• Multi-site diffusion MRI data present significant challenges for connectome analysis due to scanner and protocol variability.
• Site-specific information confounds efforts to combine data and derive robust biological insights.

Purpose of the Study:

• To develop a data-driven method to isolate site-invariant connectome features from multi-site diffusion MRI data.
• To maintain relevant biological information, such as age-related network changes, while removing site-specific artifacts.

Main Methods:

• A conditional, variational autoencoder model was developed with additional prediction tasks for patient age and network modularity.
• The model constructs a latent space uncorrelated with imaging site but correlated with age and modularity.
• The approach was tested by projecting connectome data from the VMAP and BIOCARD studies to a common site domain.

Main Results:

• The projected network modularity data exhibited statistically similar means across different sites (p < 0.05).
• Positive correlations between patient age and network modularity were successfully preserved after harmonization.
• The model demonstrated the ability to isolate site-invariant features and re-inject site context.

Conclusions:

• The proposed method effectively harmonizes multi-site diffusion MRI data, enabling more reliable connectome analyses.
• This approach facilitates the integration of diverse datasets while preserving crucial biological relationships.
• The technique holds promise for advancing large-scale neuroimaging studies and understanding brain aging.