Task-specific uncertainty can be more useful for overseeing high-stakes medical AI than a single composite trust score. A study of a multi-task deep-learning system for MRI-based glioma diagnosis evaluated tumor segmentation alongside predictions for IDH mutation status, 1p/19q co-deletion status and tumor grade.
Across the evaluated tasks, uncertainty estimates supported meaningful error detection. Predictive-probability calibration depended on the dropout rate, with moderate rates producing the most reliable probabilities in the evaluation. Deep Ensemble and Monte Carlo Deep Ensemble estimates offered comparable operational utility to Monte Carlo Dropout, but no evaluated uncertainty-quantification method consistently dominated across all tasks and metrics.
For selective prediction, the proposed composite trust score did not consistently outperform classification uncertainty. In this evaluation, task-specific predictive uncertainty was the more informative operational indicator for classification reliability. That result does not establish a universally preferred score or method beyond the evaluated system, tasks and datasets.
Segmentation reliability required a different view of uncertainty. Uncertainty aggregated within the predicted tumor region showed the strongest and most stable negative association with segmentation accuracy, while full-brain aggregation showed weak associations. The study used adult-type glioma data from in-house and publicly available datasets, with 1,466 training cases and 236 test cases across the reported sources. Its convergence analysis examined average uncertainty magnitude as a function of Monte Carlo sample count and dropout rate, not complete case-wise or repeated-run stability.