Why cross-validation ensembles aren't what they seem for uncertainty

When predicting segmentation uncertainty in medical images, researchers often use disagreement across ensemble members. The catch: calling a 5-fold cross-validation ensemble a "deep ensemble" conflates two different signals. Cross-validation mixes data-subset effects with random-seed variability, while true deep ensembles (same training data, different seeds) isolate seed variability alone. On three datasets across three imaging modalities, deep ensembles better detect failures and calibrate confidence, while cross-validation correlates more with inter-rater disagreement. The takeaway: pick ensemble type by your goal—deep ensembles for reliability, cross-validation for modeling annotation ambiguity.