On-Manifold Variational Learning with Heat-Kernel Priors
Researchers have developed a novel manifold-anchored variational framework designed to improve unsupervised representation learning for medical imaging cohorts. This new approach utilizes a geometry-aware Expectation-Maximization algorithm, ensuring that learned prototypes remain on the data manifold by selecting them as graph medoids with high diffusion centrality. The framework also incorporates a Dirichlet energy regularizer for latent space smoothness and a per-sub-population uncertainty score for label-free quality assessment. Tested on cardiac scar and brain MRI benchmarks, the method achieved superior accuracy and produced sharper prototypes compared to existing models, maintaining stability even with a large number of sub-populations. AI
IMPACT Enhances unsupervised learning for medical imaging, potentially leading to more accurate diagnoses and discovery of novel pathological subtypes.