Splitting uncertainty in medical imaging into two distinct sources

When deep generative models reconstruct medical images from incomplete measurements, their uncertainty estimates conflate two very different problems: ambiguity that is physically irreducible versus errors introduced by the model's inference. This work provides a structural decomposition and cascade formulation that isolates each component, then applies simulation-based calibration tests to expose failure modes invisible to reconstruction-quality metrics alone. Validated on a Gaussian case with a known analytical solution, the method is demonstrated on accelerated MRI and EEG source imaging. The framework is aimed at researchers and practitioners who need trustworthy uncertainty estimates in high-stakes diagnostic settings.