A newly developed artificial intelligence system called FastGlioma can predict where glioblastoma—the most common and lethal malignant brain tumor in adults—is likely to return after initial surgery, according to a study published September 25 in Science Advances. Researchers at the University of California, San Francisco and the University of Michigan found that the AI model analyzes unprocessed tissue samples in under a minute, potentially allowing neurosurgeons to target aggressive treatments before recurrence becomes visible on a standard MRI.
AI Predicts Glioblastoma Recurrence Using Unprocessed Tissue
Glioblastoma carries a median survival rate of approximately 17 months after diagnosis. Even when surgeons remove all visible tumor mass and patients undergo subsequent therapies, the cancer almost always returns, typically at or near the original tumor cavity. To address this challenge, researchers evaluated an optical microscopy technique known as stimulated Raman histology (SRH). SRH generates microscopic images of fresh, unprocessed tissue in less than 60 seconds, bypassing the time-consuming dyes and staining procedures required in conventional pathology.
The study analyzed tissue samples collected from UCSF Health patients whose median time to recurrence was 5.5 months. Developers trained the FastGlioma AI system on roughly 300 samples taken from 60 patients and conducted independent testing on approximately 100 samples from another 20 patients. The AI system scores tissue based on patterns of tumor infiltration. According to the findings, the AI score alone performed comparably to standard pathology in forecasting which regions would later develop recurrent tumors.
Combining AI Scores with Clinical Data
To improve predictive accuracy, researchers integrated the AI infiltration score with clinical, imaging, and molecular data across six distinct machine-learning models. The best-performing model successfully distinguished between brain sites that developed recurrences and those that did not. Across five of the six models, the AI measure of tumor infiltration emerged as the strongest individual predictor of recurrence, surpassing the predictive value of the tumor’s traditional molecular characteristics.
The model demonstrated high precision in mapping where cancer would return within 5 or 10 millimeters of the sampled tissue. Investigators focused specifically on the first recurrence because subsequent tumor growth is frequently influenced by experimental treatments administered later in the disease course, making later recurrences harder to model reliably.
Potential Clinical Applications in Neurosurgery
The system’s real-time analysis offers actionable insights for surgical teams during operations. Sanjeev Herr, MD, a postdoctoral research fellow at UCSF and Drexel University College of Medicine who served as the study’s first author, noted that the technology provides neurosurgeons with immediate guidance during tumor removal. Shawn Hervey-Jumper, MD, a neurosurgeon at UCSF Health and Mitchel S. Berger, MD, endowed professor at the Weill Institute for Neurosciences, explained that surgeons might use these insights to remove additional brain tissue during the initial operation when safety permits.
For patients with tumors located in regions of the brain that cannot be safely excised, the predictions could direct alternative interventions. These therapies include higher-dose focal radiation or direct drug infusions delivered to the tumor site via a catheter placed through the skull, according to Hervey-Jumper. Co-senior author Todd Hollon, MD, of the Machine Learning in Neurosurgery Laboratory at the University of Michigan, Ann Arbor, emphasized that the primary objective is to delay the initial recurrence and ultimately extend patient survival.
Related reading