UCSF

Diffuse midline gliomas (DMGs) have infiltrative, poorly defined boundaries that make tumor measurement difficult and variable. Volumetric assessment may be more informative than bidimensional response measurements, but manual volumetry is time-consuming and automated methods typically provide a single volume without indicating reliability. We evaluated whether evidential deep learning (EDL) with model ensembling could generate calibrated 95% uncertainty intervals for automated DMG volumes.Two five-model EDL ensembles based on SegResNet and nnU-Net v2 were evaluated. Voxel-wise probability distributions were propagated into tumor-volume estimates and 95% intervals. Analyses included voxel-wise calibration, true-volume coverage, relative interval width, and comparison of SegResNet uncertainty with physician difficulty ratings and spatial uncertainty annotations.The EDL nnU-Net v2 Ensemble approached 95% coverage but produced intervals likely too broad for clinical use, whereas the EDL SegResNet Ensemble provided the strongest balance between coverage and precision. SegResNet intervals widened with physician-rated difficulty, while aleatoric and epistemic uncertainty localized near physician-identified uncertain regions and tumor boundaries. These findings suggest evidential ensembling can provide automated DMG volumetry with calibrated uncertainty estimates that reflect clinically recognizable sources of difficulty.