AI Tool Helps Surgeons Distinguish Aggressive Glioblastoma from Other Brain Cancers in Real-Time

Glioblastoma, the most common and aggressive brain tumor, often appears similar to primary central nervous system lymphoma (PCNSL), a rarer cancer with different treatment needs. Misdiagnosis can lead to unnecessary surgery or delays in proper care. Now, a new artificial intelligence (AI) system allows surgeons to differentiate between these look-alike cancers in real time with near-perfect accuracy.

A research team led by Harvard Medical School (Boston, MA, USA) has developed an AI tool called PICTURE (Pathology Image Characterization Tool with Uncertainty-aware Rapid Evaluations). The model was trained to spot critical cancer features such as tumor cell density, cell shape, and necrosis, allowing it to distinguish glioblastoma from PCNSL during operations. What makes PICTURE unique is its uncertainty detector, which alerts doctors when a tumor does not match known patterns and requires human review, ensuring safe integration into high-stakes decisions.

The AI was tested on 2,141 brain pathology slides, including rare frozen and formalin-fixed samples, and evaluated across five hospitals in four countries. The results, published in Nature Communications, showed the model could correctly distinguish glioblastoma from PCNSL more than 98% of the time, outperforming both human pathologists and existing AI tools. Importantly, the model also flagged 67 central nervous system cancers outside its main categories, recognizing when it had not seen a tumor type before.

In addition to accuracy, PICTURE addressed a key weakness in current practice. Traditional frozen-section analysis can take 15 minutes but carries error rates of up to 1 in 20 cases, with misdiagnoses occurring in 38% of difficult tumors. The new system minimizes errors, supports clinicians in uncertain cases, and prevents misclassification of rare tumors. It has shown reliable performance during surgery and in cases where human experts disagreed.

Researchers see broad potential for PICTURE to democratize neuropathology, an area with few specialists unevenly distributed worldwide. By providing decision support in real time, the tool could guide treatment in operating rooms, sparing patients with PCNSL from unnecessary surgery while ensuring aggressive resection for glioblastoma cases. Future plans include expanding the model to cover more brain cancer subtypes and integrating genetic and molecular data for deeper insights.

“Our model can minimize errors in diagnosis by distinguishing between tumors with overlapping features and help clinicians determine the best course of treatment based on a tumor’s true identity,” said study senior author Kun-Hsing Yu. “Our model shows reliable performance on frozen sections during brain surgery and in scenarios with significant diagnostic disagreement among human experts.”

Related Links:
Harvard Medical School