Learning disentangled representations to harmonize connectome network measures

Nancy R. Newlin, Michael E. Kim, Praitayini Kanakaraj, The Biocard Study Team, Kimberly R. Pechman, Niranjana Shashikumar, Elizabeth E. Moore, Derek B. Archer, Timothy J. Hohman, Angela L. Jefferson, Daniel Moyer, Bennett A. LandmanJournal of Medical Imaging 2025 · DOI

TL;DR

This paper introduces a novel method using conditional variational autoencoders to "harmonize" brain connectome data collected from different sites. By disentangling site-specific variations from true biological signals, their approach creates more robust and comparable network measures, crucial for large-scale, multi-site studies of neurodegenerative diseases like Alzheimer's.

What Problem Does This Paper Address?

Understanding the brain's intricate network of connections, known as the connectome, is vital for studying neurological conditions like Alzheimer's disease and the aging process. However, a major hurdle in this research is the variability introduced by different imaging sites. Factors like scanner hardware, imaging protocols, and software used can create "site-specific effects" that confound the analysis of brain connectivity. These effects can mask true biological differences or create spurious ones, making it difficult to pool data from multiple research centers. Previous attempts to address this often focused on harmonizing raw imaging data or used methods that require traveling subjects or perfect demographic matching, which are impractical for large-scale studies. This paper tackles the challenge of harmonizing the derived connectome network measures themselves, aiming to make data from different sites directly comparable.

Methods

The core of this paper's innovation lies in a conditional variational autoencoder (VAE) architecture. This VAE is designed to learn a "disentangled" latent representation of the connectome data. Here's how it works:

  • Encoder: Takes the connectome data and maps it to a latent space.
  • Site-Conditional Decoder: This is the key. The decoder reconstructs the connectome, but it's "conditioned" on the site information. This allows the model to learn how to generate connectomes that look like they came from a specific site.
  • Disentanglement: The VAE is trained with multiple "prediction heads" that predict patient sex, age, and brain network measures from the latent space. This forces the latent space to capture meaningful biological information while simultaneously minimizing the influence of site-specific factors.
  • Harmonization: During testing, the model can reconstruct the connectome in a "common site domain" of the user's choice. This effectively shifts the data to a standardized site, reducing inter-site variability.

The authors explored two learning schemes: "individual learning" (training a separate model for each network measure) and "co-learning" (training a single model for all measures simultaneously).

Datasets

The study utilized data from two well-established cohorts:

  • BIOCARD (Biomarkers of Cognitive Decline Among Normal Individuals): Scanned on a 3T Philips Achieva scanner at Johns Hopkins University.
  • VMAP (Vanderbilt Memory and Aging Project): Scanned on a Philips 3T Achieva scanner at Vanderbilt University Medical Center.

The data was split into training and testing cohorts. A matched testing cohort was created by pairing participants from BIOCARD and VMAP based on similar demographics (age ±1 year, same sex) to rigorously evaluate harmonization performance without the need for traveling subjects.

Key Insights and Ideas

  • Connectomes are inherently site-biased: Even with matched demographics, raw connectome data shows systematic differences between sites, particularly in intra-hemispheric connections. This highlights the need for harmonization at the network measure level.
  • Disentanglement is key: By forcing the model to predict biological attributes (sex, age) and network measures from the latent space, the method effectively separates site-specific noise from genuine biological signals.
  • Co-learning offers an advantage: The "co-learning" scheme, where all network measures are learned simultaneously, demonstrated superior harmonization performance and better preservation of biological information compared to individual learning.
  • Robustness to demographic matching: A significant advantage is that the model itself doesn't require perfect demographic matching during training or testing, making it more scalable and applicable to diverse datasets.

Key Results

The proposed method successfully harmonized network measures, reducing inter-site differences to small and medium effect sizes (Cohen's D < 0.2). Significant site-wise differences in median values for most network measures were ameliorated.

  • Sex prediction accuracy: Achieved 0.92 ± 0.02 accuracy under the co-learning scheme.
  • Age prediction error: Achieved 2.71 × 10⁻² ± 2.86 × 10⁻³ years of error under the co-learning scheme.
  • Harmonization efficacy: Figures 3 and 4 visually demonstrate the reduction in inter-site variability for various network measures after applying the proposed correction.

Limitations and Future Work

  • Limited to two sites: The current study was conducted using data from only two sites. Future work aims to expand applicability to more sites.
  • Cognitive impairment: The current model was calibrated using data from individuals without cognitive impairment. Future work will extend its utility to patients with cognitive impairment.
  • ComBat comparison: While the paper mentions ComBat as a widely used method, it doesn't provide a direct quantitative comparison of their method's performance against ComBat on the same dataset.
  • Computational intensity: While more efficient than some methods, the VAE approach is still computationally intensive.

The authors suggest that their method holds significant potential for advancing multi-site studies of neurodegeneration and aging by enabling the integration of diverse datasets.

Summary updated

Abstract

Our findings demonstrate that using latent representations significantly harmonizes network measures and provides robust metrics for multi-site brain network analysis.
BibTeX
@article{newlin2025learning,
  author = {Nancy R. Newlin and Michael E. Kim and Praitayini Kanakaraj and The BIOCARD Study Team and Kimberly R. Pechman and Niranjana Shashikumar and Elizabeth E. Moore and Derek B. Archer and Timothy J. Hohman and Angela L. Jefferson and Daniel Moyer and Bennett A. Landman},
  title = {{Learning disentangled representations to harmonize connectome network measures}},
  journal = {Journal of Medical Imaging},
  year = {2025},
  doi = {10.1117/1.jmi.12.1.014004},
}
Figure