Effectiveness of MC Dropout Uncertainty for Localizing Segmentation Errors at Tumor Boundaries

Determine whether Monte Carlo Dropout–based uncertainty estimates in 2D brain tumor MRI segmentation reliably identify segmentation errors, particularly along tumor boundaries, by quantifying the strength of association between per-pixel uncertainty values and misclassification errors.

Background

Monte Carlo (MC) Dropout is a widely used approach to approximate Bayesian uncertainty in deep learning models, including for medical image segmentation tasks. In brain tumor segmentation, accurately identifying error-prone regions at tumor boundaries is clinically important because boundary inaccuracies can affect downstream clinical decisions.

The paper motivates its empirical paper by noting that the practical effectiveness of MC Dropout for highlighting segmentation errors—especially at boundaries—has not been clarified. The authors then present an analysis for 2D brain tumor MRI segmentation using a U-Net, showing weak global correlations and negligible boundary correlations between uncertainty and error, suggesting limited utility of MC Dropout for boundary error localization in this setting.