A new AI-based model can predict where glioblastoma is most likely to return after initial surgery.
The development could give doctors a chance to treat it before it becomes visible on MRI.
Researchers at UC San Francisco and the University of Michigan developed an AI approach that can predict, with high accuracy, where the cancer is likely to recur first. In glioblastoma, most recurrences occur at or close to the tumour cavity.
The findings, published in Science Advances and supported by the National Institutes of Health, could eventually enable doctors to use targeted treatments in these areas before a new tumour becomes visible on MRI.
First author Sanjeev Herr, MD, a postdoctoral research fellow at UC San Francisco and Drexel University College of Medicine, said: ‘The system has incredible potential to provide neurosurgeons with valuable real-time guidance during tumour removal. It can also generate insights that guide subsequent treatment decisions."
These treatments might include removing additional tissue during the initial surgery when safe to do so, said senior author Shawn Hervey-Jumper, MD, a neurosurgeon at UCSF Health, and Mitchel S Berger, an endowed professor at the Weill Institute for Neurosciences.
He said: ‘Patients with disease in parts of the brain that cannot be removed may undergo other treatments targeted at the predicted sites of progression, like higher-dose focal radiation or drugs infused directly into the tumour, via a catheter placed through the skull.’
The researchers analysed tissue samples collected during glioblastoma surgery from UCSF Health patients, for whom the median time to recurrence was 5.5 months. They developed their model using about 300 samples from 60 patients and tested it separately on about 100 samples from another 20 patients.
They used a technique called stimulated Raman histology (SRH), which produces microscopic images of fresh, unprocessed tissue in less than a minute. Conventional pathology requires tissue to be processed with dyes and stains, which takes longer and is more labour-intensive.
They then analysed the images with FastGlioma, an AI system developed by researchers at UCSF and the University of Michigan. The system scores tissue based on tumour infiltration.
That AI score alone performed about as well as conventional pathology at predicting which areas would later develop recurrent tumours. But when researchers combined the AI score with clinical, imaging and molecular data in six machine-learning models, they found that the best-performing model was significantly better at distinguishing between sites that would and would not develop recurrences.
The AI measure of tumour infiltration was also the strongest individual predictor of recurrence across five of the six models, providing more predictive information than the tumour’s molecular characteristics.
The researchers also tested how precisely the model could pinpoint where recurrence would develop. The model performed well at predicting whether cancer would return within 5 or 10 millimetres of the sampled tissue.
The team chose the first recurrence because patients often receive experimental treatments later in their disease, which can influence tumour growth and make subsequent recurrences more difficult to predict reliably.
Published: 16.10.2026
