Enhancing Confidence in AI Enhanced Brain Tumor Imaging by Measuring Uncertainty

MRIs are among the most accurate imaging tests used to diagnose brain tumors, and ​MRI-based monitoring of tumors ​​guides​​​​ critical treatment decisions. The clinical standard for measuring tumors often relies on either subjective evaluation or simplified two-dimensional (2D) measurements.   

However, for tumors with slow or irregular growth patterns, such as meningiomas, these 2D metrics can fail to capture the true tumor burden, with studies showing that 3D spatial analysis provides a more accurate picture of tumor progression. While deep neural networks can be trained to segment (i.e. delineate) tumors in 3D, there is a level of uncertainty that comes with the automated segmentation, limiting clinician trust and adoption of these diagnostic strategies. Due to tumor characteristics, locations, or imaging quality, that level of uncertainty is inherently higher in some cases than others, but current AI models treat each case the same, as if its outputs are definitive. 

To measure levels of uncertainty when using AI to enhance MRI segmentation, UCSF researchers developed a deep learning framework that generated uncertainty estimates for meningioma segmentation on brain MRI. Their Evidential Deep Learning (EDL) AI model achieved a high level of accuracy and produced well-calibrated, credible measurements of tumor volume, supporting safer clinical AI deployment.  

 Their study recently appeared in npj Digital Medicine.

The researchers focused their study on meningiomas, the most common primary brain tumor, accounting for over one-third of all intracranial tumors and nearly half of all primary brain tumors. While some meningiomas have very well-defined borders and model confidence would be expected to be high, others are adjacent to anatomical structures that obscure the tumor boundary, or demonstrate unusual shapes, leading to uncertainty in a model when measuring tumor volumes or when communicating tumor boundaries to clinicians. ​​​​ 

Their deep learning framework was trained on 1,655 MRIs (788 patients) and included post-operative brain MRIs that added to uncertainty scores because treatment-related changes can exhibit similar patterns to tumor tissue. They also evaluated both homogeneous and heterogeneous AI ensembles on an independent test set of 68 MRIs (43 patients). The algorithm’s performance was assessed from the spatial agreement between uncertainty maps and neuroradiologist-identified ambiguous regions. Their model achieved high accuracy with uncertainty maps aligning with ambiguous regions and well-calibrated volume estimates. External validation in 353 patients confirmed generalizability.  

“This capability has the potential to substantially increase trust in biomedical image segmentation, particularly in applications such as brain tumor volumetrics where decisions are sensitive to boundary ambiguities and subtle longitudinal changes in the context of heterogenous image quality.” It also advances the broader goal of developing transparent, safe, and trustworthy AI for medicine, paving the way for uncertainty-aware quantitative monitoring of tumor dynamics in routine clinical care. 

Rauschecker notes that while their study was designed for deployment within a local workflow and trained exclusively on internal private data, the high segmentation performance on an independent external test set provides compelling evidence of cross-institutional generalization. Nevertheless, he adds that future studies should incorporate multi-center datasets and multi-rater annotations to fully validate model uncertainty against human inter-observer variability.

 References:

1. https://www.nature.com/articles/s41746-026-02902-0

2. Andreas Rauschecker, MD, PhD, UCSF assistant professor of Radiology and Co-Chief of Intelligent Imaging Research. https://radiology.ucsf.edu/people/andreas-rauschecker

(Newswise/MF)