A research team led by Harvard Medical School has created an AI tool capable of identifying two similar-looking brain cancers, each with different origins, behaviours and treatments.
With near-perfect accuracy, the tool – PICTURE (Pathology Image Characterisation Tool with Uncertainty-aware Rapid Evaluations) – successfully distinguished between glioblastoma and primary central nervous system lymphoma (PCNSL), a rarer cancer often confused with glioblastoma.
Although both can appear in the brain, glioblastoma originates from brain cells, while PCNSL develops from immune cells.
Their microscopic appearances can be very similar, often leading to misdiagnosis with profound implications for treatment.
Supported partly by the National Institutes of Health, this work is published in Nature Communications.
The AI model is available for public use, allowing other scientists to explore and expand upon it, the team said.
Identifying lookalike tumours in the brain during surgery remains one of the most challenging diagnostic tasks in neuro-oncology, the researchers noted.
An accurate diagnosis during surgery can speed up treatment decisions, such as removing glioblastoma or opting for radiation and chemotherapy for PCNSL.
Misdiagnosis or delays in diagnosing brain cancers can result in unnecessary surgery and postponements in administering proper treatment.
What makes this tool particularly valuable is its capacity to be used during surgery, providing real-time insights to surgeons and pathologists.
Study senior author Kun-Hsing Yu, an associate professor of biomedical informatics at the Blavatnik Institute, Harvard Medical School, and assistant professor of pathology at Brigham and Women’s Hospital, said: ‘Our model can minimise errors in diagnosis by distinguishing between tumours with similar features and assist clinicians in selecting the best treatment based on the tumour’s true identity.’
During brain tumour operations, surgeons usually remove tissue for rapid microscopic evaluation.
This evaluation involves freezing the sample in liquid nitrogen, which may slightly alter cellular details, but allows for a quick, real-time assessment that takes approximately 15 minutes.
Based on this initial result, surgeons decide whether to remove the tumour or leave it in place, opting for radiation and chemotherapy instead. Over the next few days, pathologists conduct a more detailed and reliable examination.
In roughly 1 in 20 cases, the initial diagnosis is revised upon second review, Yu said. This is precisely where the new AI system can be especially useful – reducing uncertainty and lowering the risk of error during critical surgical decisions.
He added: ‘Our model performs reliably on frozen sections during brain surgery and in scenarios where there is significant disagreement among human experts.’
The tool was tested across five hospitals and outperformed both human pathologists and other AI models.
A key feature of this new model is an ‘uncertainty detector’, which not only differentiates cancer types with high accuracy but also indicates when it is unsure, which is essential in high-stakes medical scenarios.
Accurate differentiation of PCNSL from glioblastoma during surgery can enable surgeons to preserve brain tissue instead of removing it.
Patients with PCNSL are then referred for radiation and chemotherapy, which are preferred treatments, whereas glioblastoma requires extensive surgical removal of cancerous tissue.
The model, developed by Yu with co-first authors Junhan Zhao and Shih-Yen Lin, was tested on 2,141 brain pathology slides collected globally, including rare cases from frozen and formalin-fixed samples. It was designed to identify critical cancer features such as tumour cell density, cell morphology, and necrosis.
Testing across five international hospitals in four countries showed that PICTURE outperformed existing AI tools and standard frozen-section assessments, the usual method for real-time tumour typing.
In tests, PICTURE correctly distinguished glioblastoma from PCNSL more than 98% of the time – an accuracy that was consistent across five independent international patient groups.
Additionally, PICTURE identified samples from 67 other CNS cancers that were neither gliomas nor lymphomas.
It could detect tumours it had not previously encountered and, when it did, flagged these cases for human review.
This feature makes the model unique among AI systems, the researchers said.
Unlike other AI tools that operate in a binary fashion – identifying disease A versus disease B – this model can handle multiple subtypes, which is particularly important given the more than 100 brain cancer subtypes, many of which are rare.
PICTURE also outperformed human pathologists in diagnosing complex brain tumours. Human experts showed considerable disagreement on challenging cases, with some tumour types misdiagnosed 38% of the time. PICTURE correctly identified all these challenging cases, providing valuable support when expert opinions differ.
