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

A new study warns that Monte Carlo Dropout, a common technique for estimating uncertainty in medical image segmentation, can produce dangerously misleading confidence scores in brain tumour analysis, particularly for the most clinically critical tumour sub-regions [1]. The research, a two-model case study on 126 patients from the BraTS21 dataset, evaluated a pretrained SegResNet and a locally trained UNet with residual units (UNet-Res) for glioma segmentation in multiparametric MRI [1]. Glioma segmentation is a critical component of treatment planning, and a model that fails silently on treatment-critical sub-regions represents a patient safety risk that overlap-based metrics such as Dice scores cannot expose [1]. The study found that MC dropout preserved segmentation accuracy, with an absolute difference in Dice scores of less than 0.01, while achieving strong uncertainty-error alignment, indicated by an AUROC for entropy of approximately 0.97 [1]. This suggests uncertainty correctly ranks erroneous voxels above correct ones [1]. Entropy-based patient stratification identified a high-uncertainty subgroup with substantially lower segmentation performance, recording a median whole-tumour Dice of 0.835 compared to 0.925 for the low-uncertainty subgroup, supporting the use of uncertainty as a practical triage signal [1]. However, the study's authors caution that global alignment metrics can mask important region-specific differences [1]. Despite similar AUROC scores, the UNet-Res model exhibited near-zero enhancing tumour entropy of 0.054 and an Expected Calibration Error of 0.915, with a Dice score of only 0.714 [1]. This indicates severely miscalibrated confidence on the most clinically critical sub-region, a failure mode invisible to standard Dice and AUROC reporting [1]. The findings demonstrate that strong uncertainty-error alignment is necessary but insufficient for clinical safety, and the authors argue that sub-region-specific calibration assessment must accompany AUROC evaluation when selecting models for clinical deployment [1]. The study underscores a broader challenge in deploying AI in high-stakes medical environments, where standard performance metrics may not capture the full picture of a model's reliability in specific, clinically relevant tasks.