AI diagnoses brain tumours in minutes instead of weeks — Researchers

Researchers in Heidelberg, Germany, have developed an artificial intelligence system capable of classifying brain and spinal cord tumours with high accuracy using routine microscope slides, offering a potential breakthrough in the speed and accessibility of cancer diagnosis.

The system, called Hetairos, can distinguish more than 100 molecular subtypes of central nervous system tumours using standard digitised tissue samples stained for conventional pathology. The work, published in Nature Cancer, suggests the approach could significantly reduce diagnostic delays and expand access to advanced tumour classification worldwide.

Brain tumours are highly diverse, and accurate diagnosis increasingly depends not only on how they appear under the microscope but also on molecular profiling, particularly DNA methylation analysis, which is considered the current gold standard. However, these tests require specialised laboratories, expensive equipment and sufficient tumour material, and results can take around two weeks. In many parts of the world, such facilities are not available.

The new AI system aims to address these limitations by predicting molecular tumour subtypes directly from standard histology slides.

Hetairos was developed by researchers led by Professor Moritz Gerstung at the German Cancer Research Centre and Professor Felix Sahm at Heidelberg University and Heidelberg University Hospital. It was trained on more than 11,000 digitised tissue sections from 9,606 patients across 11 medical centres on four continents. The dataset was primarily labelled using DNA methylation-based diagnoses.

In total, the system can classify 102 tumour subtypes, covering nearly the full range of the World Health Organisation classification for central nervous system tumours.

The AI also provides a confidence score for each prediction. In roughly half to 70% of cases, where confidence was high, accuracy reached about 87–88%. Even when less certain, the system was able to narrow down possible diagnoses from over 100 to just a small set of likely options, helping pathologists focus further testing.

In head-to-head comparisons, Hetairos outperformed experienced neuropathologists. When five specialists were asked to classify 210 cases using only tissue slides, the AI achieved 68% accuracy compared with an average of 30% for human experts. When considering the top three possible diagnoses, the AI reached 84% accuracy versus about 50% for specialists.

Researchers say the system is particularly effective at recognising subtle tissue patterns that may be difficult for even experienced clinicians to detect, though it still struggles with very rare tumour types where human expertise remains comparable.

In a parallel clinical study, the AI analysed tumour samples without influencing final medical decisions. While conventional molecular diagnostics took an average of 12 days, Hetairos produced results in around 12 minutes once slides were digitised. Including preparation time, results could be available within one to two days.

The researchers say the tool could be especially valuable in cases where molecular testing is inconclusive, insufficient tissue is available, or resources are limited. The system also highlights regions of tissue that influenced its decision, offering clinicians a way to interpret and verify its outputs.

Importantly, the developers emphasise that Hetairos is intended to complement, not replace, existing molecular diagnostics.

“It is not intended to replace molecular analyses but to support and accelerate them,” said Professor Sahm. “It could be especially useful in regions where advanced testing is not readily available.”

The system may also reduce costs, as DNA methylation testing can be expensive, while the AI relies on existing histology slides already used in routine care.

Researchers say further validation across larger and more diverse datasets is needed, but the results point to a future in which AI could help deliver faster, more accessible cancer diagnosis on a global scale.