Lost in the Folds: When Cross-Validation Is Not a Deep Ensemble for Uncertainty Estimation
A new research paper titled "Lost in the Folds" highlights a common misunderstanding in AI research regarding uncertainty estimation in medical image segmentation. The study reveals that using K-fold cross-validation (CV) to form ensembles, often mislabeled as deep ensembles (DE), can lead to inaccurate interpretations of uncertainty. DE, which use the same training data but different random seeds, are found to be better for reliability tasks like failure detection, while CV ensembles are more suited for modeling ambiguity. AI
IMPACT Clarifies best practices for uncertainty estimation in AI, impacting reliability and ambiguity modeling in medical imaging.