Researchers have developed a new artificial intelligence (AI) tool capable of identifying more than 100 brain tumour subtypes from routine pathology slides, potentially accelerating diagnosis and helping guide treatment decisions for patients with central nervous system (CNS) tumours.
A Shortcut to Brain Tumour Classification
Molecular profiling, particularly DNA methylation analysis, has become a cornerstone of modern CNS tumour classification because it can distinguish between tumour types that appear similar under the microscope. However, methylation testing is costly, requires specialised infrastructure, and can take days or weeks to deliver results. In many parts of the world, access remains limited.
The newly developed AI model, called Hetairos, was designed to bridge this gap by predicting methylation-based tumour classifications directly from standard haematoxylin and eosin (H&E) stained tissue slides. These slides are routinely used in pathology laboratories worldwide and are far more readily available than advanced molecular testing.
The researchers trained and validated Hetairos using more than 11,000 digital pathology slides from 9,606 patients across 11 centres spanning four continents. The model was tasked with identifying 102 distinct CNS tumour methylation subtypes, covering both paediatric and adult brain tumours.
Among cases where the system was most confident in its predictions, Hetairos achieved an accuracy of 0.87. Importantly, the model was able to provide high-confidence classifications for approximately 50–70% of patients, suggesting it could offer clinically useful guidance in a substantial proportion of cases.
Outperforming Expert Neuropathologists
To assess its performance in a real-world diagnostic setting, investigators compared Hetairos with five board-certified neuropathologists who reviewed tumour histology alone, without access to molecular data.
The AI system achieved a diagnostic accuracy of 0.68, compared with 0.30 for the neuropathologists. While molecular testing remains the diagnostic gold standard, the findings suggest the model can recognise subtle histological patterns linked to specific molecular tumour subtypes that may not be readily apparent to human observers.
The researchers emphasised that Hetairos is intended to support, rather than replace, expert pathology review.
Faster Answers for Patients With CNS Tumours
One of the most striking findings emerged during prospective testing in routine clinical practice. While conventional molecular analysis required a median turnaround time of 12 days, Hetairos generated predictions in just 12 minutes.
This dramatic reduction in waiting time could help clinicians narrow differential diagnoses earlier, prioritise confirmatory testing, and potentially begin treatment planning sooner.
Expanding Access to Precision Diagnostics
The authors believe Hetairos could be particularly valuable in healthcare settings where comprehensive molecular testing is unavailable or delayed. By providing rapid predictions from widely accessible pathology slides, the system may help extend the benefits of precision diagnostics to more patients worldwide.
Although the tool does not eliminate the need for molecular confirmation, the study demonstrates that AI can extract clinically meaningful molecular information from routine histology alone. As digital pathology becomes increasingly integrated into clinical workflows, approaches such as Hetairos may play an important role in improving the speed, accessibility, and accuracy of CNS tumour diagnosis.
Reference
Jin D et al. Hetairos is a histology-based artificial intelligence model for predicting central nervous system tumor methylation subtypes. Nat Cancer. 2026;DOI:10.1038/s43018-026-01186-3.
Featured image: 程 加星 on Adobe Stock
