PriUS framework aligns uncertainty with human perception in medical imaging

Uncertainty quantification is critical for high-stakes medical AI, but most methods reduce it to a single confidence score that lacks spatial interpretability. A new paper from researchers at Fudan University and other institutions tackles this gap with PriUS (Principle-Guided Uncertainty Supervision), a framework that forces uncertainty estimates to behave in human-understandable ways. The team identifies three perception-aligned principles: uncertainty should reflect image contrast between anatomical structures, the severity of image corruption (e.g., noise or artifacts), and the geometric complexity of boundaries. Using evidential learning, they incorporate explicit supervision objectives for each principle during training.

Experiments on three public benchmarks—ACDC (cardiac), ISIC (skin lesions), and WHS (whole heart)—show PriUS achieves uncertainty maps that correlate more strongly with ground-truth ambiguity sources. The method introduces quantitative metrics to measure consistency between predicted uncertainty and image attributes. While maintaining segmentation accuracy on par with state-of-the-art models, PriUS produces spatially meaningful uncertainty that clinicians can trust. This work represents a step toward making AI uncertainty not just a number, but a useful tool for understanding model limitations.