Segmentation Uncertainty Estimation

The field studies how reliably segmentation models identify pixels, objects, and boundaries, including uncertainty from ambiguous images or labels and uncertainty from limited model knowledge. Common approaches include deep ensembles and Monte Carlo dropout, evidential and Bayesian models, generative methods that produce multiple plausible masks, and calibration techniques such as conformal prediction. These estimates can help detect errors and distribution shifts, prioritize regions for human review, and represent disagreement among annotators. A central challenge is that well-calibrated scores do not always identify clinically important ambiguity or improve decisions, and results often vary across datasets, tasks, and review settings.