A research team led by Harvard Medical School has developed an AI-powered tool that could transform the way brain tumors are diagnosed in the operating room. The system, named PICTURE (Pathology Image Characterization Tool with Uncertainty-aware Rapid Evaluations), distinguishes between glioblastoma, the most common and aggressive brain tumor, and primary central nervous system lymphoma (PCNSL), a rarer cancer that is often misdiagnosed.
The distinction is critical: glioblastoma requires immediate surgical removal, while PCNSL is best treated with radiation and chemotherapy. A misdiagnosis can lead to unnecessary surgery or delayed treatment. PICTURE addresses this by delivering near real-time results during surgery, helping surgeons and pathologists make faster and more accurate decisions.
High accuracy
The AI tool achieved over 98% accuracy across multiple international test sites, outperforming both human pathologists and existing AI systems. What sets PICTURE apart is its built-in “uncertainty detector”, which alerts clinicians when the model is unsure, avoiding overconfident errors in high-stakes scenarios.
Traditional frozen-section analysis, performed during surgery, provides quick but sometimes imprecise results, with one in 20 cases later revised upon detailed review. PICTURE minimizes this risk by detecting subtle tumor features invisible to the human eye, such as changes in cell density, morphology, and necrosis.
Other cancers
Beyond glioblastoma and PCNSL, the system also identified 67 other central nervous system cancers it had not been specifically trained on, and flagged them for further expert review. In head-to-head comparisons, PICTURE consistently resolved cases where human experts disagreed, sometimes misdiagnosing up to 38% of difficult tumors.
The researchers envision PICTURE as a collaborative decision-support tool, providing real-time insights in operating rooms and pathology labs worldwide. By democratizing access to advanced neuropathology, the technology could particularly benefit hospitals without specialized expertise. It also offers potential as a training platform for young pathologists.
Further validation
While promising, the researchers note that most tumor samples used for training came from white patients, meaning further validation across diverse populations is necessary. Future work aims to expand PICTURE’s scope to additional tumor types and integrate molecular and genetic data for even more precise diagnostics.
With over 300,000 new brain and CNS tumors diagnosed each year, PICTURE could help redefine precision in neuro-oncology, bringing faster, safer, and more personalized care directly into the operating room.
The research has this week been published in Nature Communications. The AI model is publicly available for other scientists to use and build upon, the team said.
