AI in Neurosurgery: What Two Brain Tumor Studies Show

Two peer-reviewed studies tested AI on fresh brain tissue during tumor surgery. What they measured, what they did not, and who disclosed an interest.

Last verified: October 20265 min read

The best-documented use of AI in neurosurgery is reading freshly removed brain tissue during an operation, so that surgeons learn within seconds to minutes whether tumor remains. A multicenter 2020 trial of 278 patients found a deep learning system matched pathologists on diagnosis (94.6 percent accuracy versus 93.9 percent), and a 2024 Nature study of a later model from the same research group, FastGlioma, reported that it missed high-risk residual tumor in 3.8 percent of 129 patients against 24.0 percent for standard adjuncts. That second comparison was a simulation on pathology-confirmed samples, and the tool was not approved to guide surgical decisions at the time of publication.

Key takeaways

  • The 2020 Nature Medicine trial paired a label-free imaging method with deep neural networks and produced a diagnosis in under 150 seconds, versus 20 to 30 minutes for conventional methods.
  • FastGlioma scores tumor infiltration in about 10 seconds and reached an average AUROC of 92.1 percent in a 220-patient prospective cohort.
  • In a head-to-head analysis, FastGlioma missed high-risk residual tumor in five of 129 patients, against 31 of 129 for image-guided and fluorescence-guided methods.
  • No study cited here measured whether patients lived longer or had fewer recurrences after AI-guided surgery.
  • Three authors of the FastGlioma paper hold shares in the company that makes the imaging system.

The problem the tools address

Surgeons removing a diffuse glioma, a tumor that spreads through normal-looking brain, must decide during the operation how much tissue to take. Tumor left behind is linked to earlier recurrence and shorter survival. Standard aids include MRI-based neuronavigation and fluorescence guidance with a compound called 5-ALA. A pathologist can also examine tissue during the operation, but the 2020 paper describes conventional methods as taking 20 to 30 minutes. The AI work discussed here targets that delay and the margin decision.

The 2020 trial: diagnosis from laser images

The technique, called stimulated Raman histology, images fresh tissue with a laser and needs no stain. A team led by Todd Hollon reported in Nature Medicine on January 6, 2020 that deep convolutional neural networks, trained on more than 2.5 million images, could classify the images and give a diagnosis in under 150 seconds. In a prospective multicenter trial of 278 patients, the system's accuracy of 94.6 percent was noninferior to pathologist interpretation of standard histology at 93.9 percent. The authors also highlighted tumor-infiltrated regions on the image. This study addressed diagnosis, meaning which type of tumor, not how much remained at the edge of the cavity.

FastGlioma: scoring tumor at the margin

The second study, by Kondepudi, Hollon, and colleagues, appeared online in Nature on November 13, 2024. FastGlioma is a foundation model pretrained on roughly 4 million label-free microscopy images, then adapted to score how densely tissue is infiltrated by tumor. It processes a sample in about 10 seconds. In a prospective international cohort of 220 adults with suspected diffuse glioma, treated at the University of California San Francisco, New York University, and the Medical University of Vienna, the average AUROC was 92.1 percent. AUROC is a score from zero to 100 percent in which 50 percent is chance and 100 percent is a perfect separation of tumor from non-tumor samples.

The head-to-head comparison and its caveats

In a subset of 129 patients with 624 matched specimens, FastGlioma was compared with MRI signals and 5-ALA fluorescence on the task of separating normal brain from dense tumor.

MethodAUROC, normal versus dense tumor
FastGlioma98.1 percent
5-ALA fluorescence89.0 percent
FLAIR MRI signal76.3 percent
Contrast enhancement on MRI71.8 percent

The study also counted patients in whom high-risk residual tumor would have been missed: five of 129 (3.8 percent) with FastGlioma versus 31 of 129 (24.0 percent) with the standard adjuncts. The authors describe the analysis as simulated, because FastGlioma is not approved by the U.S. Food and Drug Administration or the European Medicines Agency to guide resection, and the predictions were checked against pathology afterward. The comparison was also limited to the extreme contrast of normal brain against dense tumor, so it does not cover intermediate levels of infiltration. Readers can find the full text, including the limitations section, in the open-access version in the free archive of the US National Library of Medicine.

What these studies cannot say

Neither paper shows that surgeons who used the AI removed more tumor in practice, or that patients did better. The first measures agreement with a pathologist, and the second measures agreement with pathology in samples. The FastGlioma data came from three tertiary centers and one imaging system. The paper's competing-interests statement says that Christian Freudiger, Daniel Orringer, and Todd Hollon are shareholders in Invenio Imaging, and the company's imaging system produced the training and test data. That is a disclosed interest, not evidence of bias, but independent replication would strengthen the findings. The regulatory status of any clinical product based on these models in Canada or the United States could not be confirmed for this article, and readers should consult Health Canada or the FDA directly for current status.

Questions to ask about an AI-guided operation

  • Is the tool used for diagnosis, for margin assessment, or for both?
  • Is it cleared for that use by the relevant regulator, or is it investigational?
  • Does the surgeon treat the output as one input alongside MRI, fluorescence, and pathology?
  • Has it been tested in patients and tumor types like the one being treated?

This article gives general information about published research. It is not medical advice and does not describe how any individual should be treated. Questions about a specific diagnosis or operation belong with the patient's surgical team.

FAQ

Does AI perform brain surgery?

No study cited here involved AI operating on a patient. The tools analyze images of removed tissue and present a result to the surgical team, which decides what to do.

How fast are these systems?

The 2020 system produced a diagnosis in under 150 seconds, and FastGlioma scores a sample in about 10 seconds. The 2020 paper describes conventional methods as taking 20 to 30 minutes.

Is FastGlioma approved for use in surgery?

The 2024 paper states it was not approved by the FDA or the European Medicines Agency to guide treatment decisions, which is why its comparison was a simulation. Current regulatory status was not confirmed for this article.

Do the studies show better survival?

No. They report agreement with pathology and detection of residual tumor in samples. Effects on recurrence or survival would need studies that follow patients after surgery.

Last verified: October 2026

Sources

  • Hollon TC et al., Near real-time intraoperative brain tumor diagnosis using stimulated Raman histology and deep neural networks, Nature Medicine, January 2020
  • Kondepudi A et al., Foundation models for fast, label-free detection of glioma infiltration, Nature, published online November 13, 2024 (free archive PMC11711092)

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics