Brain tumour diagnosis has always been one of medicine’s more unforgiving bottlenecks. You have a patient, a tissue sample, and a question that needs answering as quickly as possible: what exactly is this, and what do we do about it?
Currently, the most reliable way to answer that question involves DNA methylation profiling — a molecular test that is genuinely excellent at classifying tumours but requires specialised laboratory equipment, a sufficient quantity of tumour material, and, crucially, about two weeks to return results. In oncology, two weeks is a long time to wait.
A team at the German Cancer Research Center (DKFZ) and Heidelberg University Hospital may have found a faster starting point. Their system, called Hetairos, was published in Nature Cancer in June 2026, and it does something that previously required that full molecular workup: classify central nervous system (CNS) tumours into 102 distinct molecular subtypes. It does this from routine tissue slides — the kind prepared in pathology labs the world over — and it takes approximately 12 minutes.
What Are H&E Slides, and Why Does That Matter?
A quick bit of context. When a tissue sample is taken from a patient, the default preparation involves staining it with haematoxylin and eosin — two dyes that colour different cellular structures, making them visible under a microscope. These H&E slides are the most common diagnostic material in pathology worldwide. They are standard, affordable, and universally available, including in clinical settings that would not have access to methylation sequencing platforms.
Hetairos works directly from digitised versions of these slides, using computer vision to detect subtle patterns in the tissue that correlate with molecular tumour subtypes. The model was trained on more than 11,000 tissue sections from 9,606 patients across 11 medical centres on four continents — a dataset substantial enough to take seriously.
The Numbers Are Striking
In a head-to-head comparison, Hetairos achieved a diagnostic accuracy of 68% on 210 cases using tissue sections alone. Five board-certified neuropathologists reviewing the same cases managed 30%. When expanded to the top three predicted diagnoses, the AI reached 84%, against roughly 50% for the specialists.
These are not small margins. They are also not an argument for replacing neuropathologists — the researchers are clear on this point, and it is worth taking their framing at face value. Hetairos is designed as a triage and decision-support tool: a way to flag likely diagnoses quickly, guide which further tests are worth ordering, and get patients onto appropriate pathways faster. The full molecular testing still happens. It is just no longer the only first step.
“It is not intended to replace molecular analyses, but rather to specifically complement and accelerate them,” said Felix Sahm, a project leader at DKFZ.
Honest About Uncertainty
One detail that deserves credit: Hetairos is built to communicate its own confidence level. In any given case, it will tell you how certain it is. High-confidence predictions occurred in 50–70% of cases across cohorts, and those predictions were correct roughly 90% of the time. That is a useful distinction. An AI that tells you when it is unsure is considerably more clinically usable than one that isn’t.
What This Actually Represents
Brain tumours are notoriously heterogeneous — the WHO currently recognises dozens of CNS tumour subtypes, and correct classification has direct implications for treatment. Getting that classification faster, from widely available materials, has real downstream consequences for patients, particularly in lower-resource settings where molecular diagnostics may not be accessible at all.
That is the argument here, and it holds up. This is not a story about AI rendering specialists redundant. It is a story about a diagnostic gap — between what the gold standard can tell us and how quickly it can tell us — and a tool that might credibly help close it.
Whether Hetairos performs as well when deployed more broadly, outside the controlled conditions of a retrospective training dataset, remains to be seen. But as a proof of concept that AI can extract meaningful molecular information from routine histopathology, the results are hard to dismiss.
Sources
- Jin, D. et al. (2026). Hetairos: AI-based CNS tumour classification from H&E slides. Nature Cancer. nature.com/natcancer
- DKFZ Press Release — Hetairos system announcement, June 2026: dkfz.de
- Capper, D. et al. (2018). DNA methylation-based classification of central nervous system tumours. Nature, 555, 469–474 — foundational paper establishing methylation profiling as diagnostic gold standard.
- WHO Classification of Tumours of the Central Nervous System (5th edition, 2021) — for tumour subtype classification context.
As always — I am not a clinician or researcher, just someone paying close attention. Read the primary paper before drawing conclusions.
