Confidence is Not Reliability: Rethinking MC Dropout in Brain Tumour Segmentation

2026-06-17

AI Summary

Current brain glioma segmentation models may silently fail in clinically critical subregions, a risk inadequately captured by conventional metrics such as Dice score. This study empirically evaluates the reliability of Monte Carlo Dropout for voxel-wise uncertainty estimation, integrating SegResNet and UNet-Res architectures and analyzing performance through entropy, Expected Calibration Error (ECE), AUROC, and Dice metrics. While global indicators appear favorable—e.g., AUROC ≈ 0.97—they obscure severe miscalibration in the enhancing tumor region, where UNet-Res exhibits an ECE of 0.915 and a Dice score of only 0.714. These findings underscore the necessity of incorporating subregion-specific calibration assessment to ensure safe clinical deployment of segmentation models.

Abstract

Glioma segmentation in multiparametric MRI is a critical component of treatment planning. A segmentation model that fails silently on treatment-critical sub-regions represents a patient safety risk that overlap-based metrics such as Dice scores cannot expose. We ask whether voxel-level uncertainty estimation via Monte Carlo (MC) Dropout can reliably identify segmentation errors in clinically critical sub-regions, and whether calibration failure modes are detectable from standard reporting metrics alone. In an empirical two-model case study on 126 BraTS21 patients, we evaluate a high-performance pretrained SegResNet and a locally trained UNet with residual units (UNet-Res). MC dropout preserved segmentation accuracy ($|Δ\text{Dice}|$ $<0.01$) while achieving strong uncertainty-error alignment (AUROC for entropy (H) $\approx$0.97), indicating uncertainty correctly ranks erroneous voxels above correct ones. Entropy-based patient stratification identified a high-uncertainty subgroup with substantially lower segmentation performance (median whole-tumour Dice $0.835$ vs. $0.925$), supporting uncertainty as a practical triage signal. However, global alignment can mask important region-specific differences. Despite similar AUROC, UNet-Res exhibited near-zero enhancing tumour entropy ($0.054$) and Expected Calibration Error (ECE) of $0.915$, with a Dice of only $0.714$, indicating severely miscalibrated confidence on the most clinically critical sub-region, a failure mode invisible to standard Dice and AUROC reporting. These findings demonstrate that strong uncertainty-error alignment is necessary but insufficient for clinical safety: sub-region-specific calibration assessment must accompany AUROC evaluation when selecting models for clinical deployment.

Problem

• brain tumour segmentation
• uncertainty estimation
• MC Dropout
• calibration
• clinical safety

Innovation

• MC Dropout
• Uncertainty Calibration
• Brain Tumor Segmentation
• Expected Calibration Error (ECE)
• Sub-region-specific Evaluation