Researchers at the German Cancer Research Center (DKFZ), Heidelberg University’s Medical Faculty, and Heidelberg University Hospital have developed an AI system called Hetairos that can classify central nervous system (CNS) tumors using routinely prepared and stained tissue sections. The research, published in Nature Cancer, showed that the system can classify 102 molecular tumor subtypes of CNS cancers using digitized H&E slides, delivering diagnostic findings in minutes rather than the days or weeks often required for molecular testing.
“The study shows that artificial intelligence is capable of deriving molecular information directly from routine tissue sections and thus fundamentally changing cancer diagnostics,” said lead author Darui Jin, PhD, a postdoctoral fellow at DKFZ.
CNS tumors encompass a broad range of diseases with substantial molecular and morphological diversity. DNA methylation profiling is currently the gold standard for classifying many brain tumors, but the tests require specialized labs employing expensive analytical tools and require sufficient amount of tumor material, which can be difficult to obtain in some instances. Results from these tests typically take about two weeks to return results. The reliance on molecular testing has also created barriers to diagnosis in some clinical settings.
“Faster and more widely accessible methods are therefore needed,” the researchers wrote. Existing alternatives such as nanopore sequencing also require specialized instruments and tissue preparation. Routine H&E histopathology remains the most widely available diagnostic material worldwide.
The development of Hetairos has been made possible from significant advances in computer vision and digital pathology. Prior research has established that AI algorithms can detect molecular features in tissue samples and classify tumors into molecularly defined categories. The researchers noted that previous research has shown that AI models can estimate methylation signals from standard H&E slides and predict specific molecular alterations, but “an artificial intelligence (AI)-based diagnostic solution for H&E slides that covers the entire spectrum of CNS tumors, as currently only possible with methylation testing, is still missing.”
To develop Hetairos, the team trained and validated the system using more than 11,000 digitized tissue sections from 9,606 patients treated at 11 medical centers across four continents. Diagnoses used for training were primarily established through DNA methylation diagnostics. The resulting model was designed to classify tumors into 102 molecular subtypes that span nearly the entire WHO classification spectrum for CNS tumors.
An important element of the system is it can estimate the level of confidence of its predictions.
“Crucial to Hetairos’s applicability across cohorts are its realistic confidence estimates, which help judge its prediction accuracy,” the researcher wrote. “Depending on the cohort, Hetairos made high-confidence predictions in 50–70% of cases, and those predictions were found to be the correct diagnosis in nearly 90% of instances.”
As it has become custom, the system was also tested directly against human experts. Five board-certified neuropathologists reviewed 210 cases using only tissue sections. Hetairos achieved an accuracy rate of 68%, compared with 30% for the specialists. When the three most likely diagnoses were considered, the AI tool was 84% accurate versus about 50% for the neuropathologists.
“The results show that modern AI systems are now capable of recognizing extremely subtle morphological patterns that are difficult even for experienced specialists to distinguish,” said Felix Sahm, a project leader at DKFZ.
A prospective evaluation of Hetairos in clinical practice analyzed 210 tumor samples alongside routine diagnostics without, but its results weren’t used to influence patient management. While complete molecular testing required an average of about 12 days, Hetairos generated results in approximately 12 minutes after digitized slides were available. Including slide preparation and scanning, findings could often be produced within 24 hours to two days.
This much speedier diagnosis could help clinicians initiate targeted treatments sooner and guide additional testing, but rather than replacing molecular testing, the developers envision Hetairos as a triage and decision-support tool.
“We developed Hetairos primarily as a tool to support diagnostics,” Sahm said. “It is not intended to replace molecular analyses, but rather to specifically complement and accelerate them. The technology could make an important contribution, particularly in countries or regions with limited resources, as it is based on standard tissue sections used worldwide.”
Join host Jonathan D. Grinstein, PhD, North American Editor for Inside Precision Medicine, as he uncovers the stories behind the pioneers driving the precision medicine revolution.
Stay up to date with the latest episodes of Behind the Breakthroughs by subscribing to the IPM eNewsletter.
A contributing partner:
Copyright © 2026 Sage Publications or its affiliates, licensors, or contributors. All rights reserved, including those for text and data mining and training of large language models, artificial intelligence technologies, or similar technologies.
