Heat-Kernel Priors Enable On-Manifold Prototypes for Medical Imaging

A new approach to unsupervised representation learning in medical imaging promises to extract clinically meaningful prototypes without relying on noisy expert labels. The paper, “On-Manifold Variational Learning with Heat-Kernel Priors,” from researchers including Jiarui Xing and Tal Zeevi, tackles a fundamental flaw in existing deep latent-variable models: they estimate Gaussian mixture priors using Euclidean averaging, which causes prototypes to drift off the curved data manifold and degenerate as the number of sub-populations increases. The team introduces a geometry-aware Expectation-Maximization (EM) algorithm whose M-step selects each sub-population prototype as the graph medoid with the highest diffusion centrality on a heat-kernel-weighted latent graph, guaranteeing every prototype remains anchored to the manifold. A Dirichlet energy regularizer further enforces geometric smoothness of the latent space, and a per-sub-population uncertainty score enables label-free quality assessment.

On benchmarks for cardiac scar segmentation and brain MRI analysis, the framework outperforms all compared methods, producing the sharpest prototypes reported to date. It remains stable even at large sub-population counts where all baselines degenerate. The manifold-anchored EM is presented as a general-purpose geometric tool that extends standard EM and could be applied to other latent-variable models beyond this medical imaging context. This work has significant implications for unsupervised discovery of disease subtypes, particularly in heterogeneous pathologies where labeled data is scarce or unreliable.