External Validation of an Evidential Deep Learning Framework for Brain MRI Lesion Segmentation

Article Impact Level: HIGH
Data Quality: STRONG
Summary of  Npj Digital Medicine. https://doi.org/10.1038/s41746-026-02902-0  
Dr. Yassine Guennoun  et al.

Points

UCSF researchers developed an evidential deep learning ensemble framework to quantify uncertainty during automated 3D meningioma segmentation on post-contrast T1-weighted brain MRIs.
Model training utilized 1,655 post-contrast T1-weighted brain MRIs from 788 patients, specifically incorporating postoperative scans to account for treatment-related tissue changes.
Internal testing on 68 MRIs from 43 patients demonstrated high segmentation accuracy with a median Dice similarity coefficient of 0.93.
External validation across 353 independent patients confirmed strong cross-institutional model generalizability, maintaining a median Dice similarity coefficient of 0.92.
Generated spatial uncertainty heatmaps aligned with neuroradiologist-identified ambiguous boundaries, producing well-calibrated volumetric measurements to support clinical decision-making.

Summary

This study evaluated the implementation of an Evidential Deep Learning (EDL) ensemble framework to generate calibrated uncertainty estimates for 3D meningioma segmentation on post-contrast T1-weighted brain magnetic resonance imaging (MRI). Meningiomas represent the most common primary brain tumor, accounting for over 33% of intracranial tumors and nearly 50% of primary central nervous system neoplasms. Standard 2D linear metrics often fail to reflect asymmetric volumetric progression. Led by Andreas Rauschecker at the University of California, San Francisco, the study addressed clinical adoption barriers caused by unquantified automated segmentation errors in complex anatomical regions.

Model training incorporated 1,655 post-contrast T1-weighted MRIs from 788 patients, including complex postoperative scans exhibiting treatment-related signal alterations that increase epistemic uncertainty. Architecturally homogeneous and heterogeneous EDL ensembles were evaluated on an independent internal test cohort consisting of 68 MRIs from 43 patients. Segmentation accuracy was quantified using the Dice similarity coefficient, spatial alignment between EDL-generated uncertainty heatmaps and neuroradiologist-annotated boundary ambiguities, and the mathematical calibration of volumetric credible intervals.

Validation on the internal test set demonstrated high segmentation accuracy with a median Dice similarity coefficient of 0.93. External validation across an independent cohort of 353 patients confirmed generalizability, achieving a median Dice similarity coefficient of 0.92. Spatial agreement analysis showed that generated uncertainty maps aligned closely with expert neuroradiologist-identified ambiguous boundaries. The authors conclude that EDL ensembles produce well-calibrated volumetric confidence intervals, enhancing clinical trust and providing a safe, uncertainty-aware framework for longitudinal brain tumor monitoring.

Link to the article: https://www.nature.com/articles/s41746-026-02902-0

References

Guennoun, Y., Nedelec, P., McArthur, M., Bloch, E., Wei, J., Sugrue, L., Calabrese, E., & Rauschecker, A. M. (2026). Segmenting with confidence through uncertainty quantification for brain tumor imaging. Npj Digital Medicine. https://doi.org/10.1038/s41746-026-02902-0