Identifiable Bayesian Representations for Heterogeneous Medical Imaging

Abstract

Medical images exhibit pervasive heterogeneity arising from acquisition protocols, scanner properties, reconstruction pipelines, modality and contrast mechanisms, and anatomical variability across subjects and scan coverage. While deep learning has achieved strong performance in many medical image analysis tasks, robustness under compounded heterogeneity remains fragile. This dissertation argues that such fragility reflects a representational limitation: when task-relevant generative properties and observational variability are not organized in an identifiable manner, models may rely on unstable observational cues as surrogates, leading to degraded generalization as heterogeneity intensifies. To address this challenge, we develop a unified perspective based on Bayesian representation learning and explicit latent role specification. Using latent variable models and variational inference, we construct mechanisms that preserve task-relevant invariants while suppressing observational variability, a requirement termed identifiable invariant preservation. We show that strengthening the identifiability of latent organization provides a practical pathway to both interpretability and improved predictive performance across progressively more demanding regimes of heterogeneity. The dissertation substantiates this thesis through three projects. First, we study supervised intracranial arterial calcification segmentation from multi-contrast brain MRI under intensity-level appearance heterogeneity. Because calcification is dark and often weakly expressed in MRI, segmentation depends on fragile contextual cues that are easily perturbed by scanner- and protocol-dependent fluctuations. A variational Bayesian formulation that restricts representational complexity yields more stable internal organization and improved segmentation accuracy. Second, we address unsupervised multimodal groupwise image registration under compounded contrast/modality variability and registration-compatible geometric heterogeneity. We formulate registration as hierarchical Bayesian inference that disentangles common anatomy from image-specific geometry, enabling intrinsic multimodal similarity and stable alignment without intensity-based heuristics. Third, we study unsupervised domain adaptation for segmentation in a correspondence-free regime with unpaired source and target domains. We introduce a probabilistic anatomical manifold that provides global canonicalization through a structured latent decomposition, inducing architecture-emergent adaptation without an explicit alignment loss and yielding a unified procedure across source-accessible and source-free settings. Together, these contributions demonstrate that interpretable, identifiable latent organization is not merely an explanatory preference, but a practical mechanism for robust medical image learning under increasing heterogeneity. By developing Bayesian, disentangled formulations that progressively strengthen latent role specification across tasks, this dissertation provides a unified methodological pathway that improves both generalization and semantic interpretability in challenging real-world imaging regimes.