Unsupervised Learning of Acquisition Variability in Structural Connectomes via Hybrid Latent Space Modeling
A new unsupervised framework removes manual capacity tuning in hybrid latent-space models for structural connectome analysis. The method architecturally anneals encoder outputs before decoding, allowing adaptive balancing of discrete and continuous latent variables during training. This addresses acquisition differences across sites, scanners, and protocols in dMRI that complicate connectome analysis. The framework separates acquisition-related effects from biological variation without requiring manual tuning, overcoming a limitation of previous hybrid models. The approach is evaluated on curated datasets, demonstrating its effectiveness in learning robust representations.
Key facts
- Acquisition differences across sites, scanners, and protocols in dMRI introduce variability in structural connectome analysis.
- Deep learning models can represent high-dimensional connectomes in a low-dimensional space while separating acquisition effects from biological variation.
- Conventional dimensionality reduction methods model all variance as continuous, causing acquisition effects to be absorbed into continuous latent space.
- Recent hybrid latent-space models combine discrete and continuous components but require manual capacity tuning.
- The proposed unsupervised framework removes manual tuning by architecturally annealing encoder outputs before decoding.
- The model adaptively balances discrete and continuous latent variables during training.
- The framework is evaluated on curated datasets.
- The approach addresses a key limitation in existing hybrid models for connectome analysis.
Entities
—