UCSF researchers' AI maps uncertainty in brain tumor MRI scans

On a brain MRI, the most consequential line may be the one an algorithm draws around a tumor—and the places where it admits that line could be wrong. Researchers at the University of California, San Francisco, developed an AI framework that segments meningiomas in three dimensions while producing maps of its own uncertainty, a step toward safer volume measurements and treatment monitoring.

MRI is among the most accurate imaging tests for diagnosing brain tumors, but tumor size is often assessed with subjective judgment or simplified two-dimensional measurements. Those shortcuts can miss the true burden of a slow-growing or irregularly shaped meningioma. A three-dimensional map captures more of the tumor’s space; the challenge is knowing when an automated boundary is dependable.

The UCSF team used an ensemble of Evidential Deep Learning models—AI models combined to estimate uncertainty—to distinguish confident segmentations from ambiguous ones. The framework was trained on 1,655 MRIs from 788 patients, including postoperative scans, where treatment-related changes can resemble tumor tissue and raise uncertainty. On an independent test set of 68 MRIs from 43 patients, the uncertainty maps aligned with regions that neuroradiologists identified as ambiguous, and the volume estimates were well calibrated.

External validation in 353 patients provided evidence that the approach could generalize beyond the training data. Andreas Rauschecker, UCSF assistant professor of radiology and co-chief of Intelligent Imaging Research, said the aim was to quantify how much uncertainty remains in a tumor-volume estimate rather than treat every AI output as definitive. The framework could also be applied to lesion segmentation beyond meningiomas, according to Rauschecker.

So what changes in practice? A clinician could receive not only an automated tumor contour and volume, but also a visual signal showing where to inspect the result most closely—particularly when image quality is uneven, anatomy hides the border or subtle changes over time matter. That could make quantitative monitoring more useful without turning the algorithm into an unquestioned authority. The model was trained exclusively on internal private data and designed for a local workflow; the researchers say multicenter datasets and multiple-rater annotations are still required to fully test its uncertainty against differences between human readers.