Unsupervised Learning of Acquisition Variability in Structural Connectomes via Hybrid Latent Space Modeling

Gaurav Rudravaram, Lianrui Zuo, Michael Eugene Kim, Karthik Ramadass, Elyssa M. McMaster, Jongyeon Yoon, Aravind Krishnan, Adam M. Saunders, Chenyu Gao, Nancy Newlin, Praitayini Kanakaraj, Lori L. Beason-Held, Murat Bilgel, Laura A. Barquero, Micah D'Archangel, Tin Nguyen, Laurie Cutting, Derek B Archer, Timothy J. Hohman, Daniel Moyer et al. (1 additional authors not shown)

Keywords: Harmonization, structural connectomes, unsupervised representation learning

TL;DR: We introduce a Joint-VAE with architectural annealing that robustly learns acquisition-related variability in structural connectomes.

Abstract:

Acquisition differences across sites, scanners, and protocols in dMRI introduce variability in structural connectome analysis.
This motivates the need for deep learning models that can represent downstream, high-dimensional structural connectomes in a low-dimensional space while explicitly separating acquisition-related effects from underlying biological variation.
Conventional statistical and deep learning approaches for dimensionality reduction typically model all sources of variance as continuous, making it difficult to separate discrete effects, e.g., acquisition- or site-related, from continuous biological variation.
As a result, acquisition-related effects often become absorbed into a continuous latent space.
Recent advances in deep learning have explored hybrid latent space modeling, where discrete and continuous components jointly represent structured variability.
However, existing hybrid approaches generally rely on manual capacity tuning to ensure that the discrete component captures desired variability (e.g., acquisition).
Here, we introduce a principled unsupervised framework that removes the need for such manual capacity tuning by \textit{architecturally annealing} the encoder outputs before decoding, allowing the model to adaptively balance the contributions of discrete and continuous latent variables during training.
To investigate this joint latent space modeling, we curated a large dataset ($N=7,416$; 60\% female) of structural connectomes derived from dMRI scans of participants.
Our dataset spans an age range of 2 to 102 years and encompasses 13 different studies with 25 unique acquisition parameter combinations.
Among these, 5,900 are cognitively unimpaired/neurotypical, 877 are diagnosed with mild cognitive impairment (MCI), and 639 are diagnosed with Alzheimer’s disease (AD).
We compare our approach with a standard VAE, PCA followed by k-means clustering, and hybrid models that impose annealing only through the loss function, showing that the architectural annealing results in stronger site learning (ARI=$0.53$, $p < 0.05$) as compared to the other methods. These results demonstrate that the proposed hybrid continuous–discrete latent space provides a useful unsupervised mechanism for capturing acquisition-related variability in diffusion MRI; by jointly modeling smooth and categorical structure, the Joint-VAE recovers meaningful clusters aligned with scanner and protocol differences.

Primary Subject Area: Unsupervised Learning and Representation Learning

Secondary Subject Area: Application: Neuroimaging

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 323