Joint-VAE model separates scanner effects from biology in brain connectomes

A team of researchers led by Gaurav Rudravaram and Bennett Landman at Vanderbilt University developed Joint-VAE, an unsupervised framework that disentangles acquisition-related noise from true biological signals in diffusion MRI (dMRI) structural connectomes. The core innovation is architectural annealing—adaptively balancing discrete and continuous latent variables during encoder training—which eliminates the need for manual capacity tuning common in prior hybrid models. On a curated dataset of 7,416 connectomes spanning ages 2–102 across 13 studies and 25 unique acquisition-parameter combinations (including 877 MCI and 639 Alzheimer's cases), Joint-VAE achieved significantly stronger site-clustering alignment (ARI=0.53, p<0.05) compared to standard VAE, PCA+k-means, and loss-based annealing hybrids.

The model's ability to separate categorical scanner/protocol effects from continuous biological gradients has immediate implications for multi-site neuroimaging studies. By jointly modeling smooth and categorical structure, Joint-VAE preserves interpretability of the continuous latent space (reflecting age, cognition, disease) while explicitly capturing discrete acquisition variability. This enables more reliable cross-site brain network analysis without manual harmonization or data exclusion. The authors released code alongside the paper, making it accessible for researchers working with heterogeneous dMRI datasets. Future work could extend the framework to other imaging modalities or incorporate clinical labels for supervised tuning.