Researchers at the German Cancer Research Centre (DKFZ), Heidelberg University and Heidelberg University Hospital have developed an AI system capable of classifying brain tumours within minutes using standard microscopic tissue samples. The system, called Hetairos, can distinguish between more than a hundred molecular subtypes of tumours in the central nervous system, representing a potential breakthrough in the diagnosis of brain and spinal cord tumours.
According to the researchers, the technology could significantly reduce the time required for a diagnosis and make specialist diagnostics more accessible in regions where advanced molecular analyses are not available.
Molecular diagnostics
Tumours of the brain and spinal cord are among the most complex forms of cancer. In recent years, it has become clear that a reliable diagnosis often cannot be based solely on the microscopic characteristics of tumour tissue, but also on the molecular properties of the tumour.
An important technique in this regard is DNA methylation analysis. This method is considered the gold standard for classifying many brain tumours, as it reveals subtle biological differences between tumours. The disadvantage is that such analyses require specialised laboratories, expensive equipment and sufficient tumour material. Furthermore, it often takes two weeks for the results to become available.
For many hospitals worldwide, this poses a significant limitation. In some countries, the necessary infrastructure is completely lacking. The results of the research were recently published in the scientific journal Nature Cancer.
Trained on thousands of patient records
The research team developed Hetairos with the aim of deriving molecular information directly from routinely stained histological slides. More than 11,000 digitised tissue sections from 9,606 patients were used to train and validate the system. The data came from eleven medical centres spread across four continents.
Based on this dataset, the AI model learned to distinguish between 102 different molecular subtypes of tumours in the central nervous system, covering virtually the entire spectrum of the current WHO classification. A key feature of Hetairos is that the system not only makes a diagnostic proposal but also indicates how confident it is in that prediction. In 50 to 70 per cent of cases, the system provided a diagnosis with a high degree of certainty. In this group, accuracy ranged between 87 and 88 per cent.
Even when the AI was less certain about a diagnosis, the system was often able to significantly narrow down the number of possible tumour types. This allows pathologists to target additional investigations more effectively.
AI outperforms experienced specialists
The researchers also compared Hetairos’ performance directly with that of human experts. Five experienced neuropathologists were presented with 210 cases and had to make a diagnosis based solely on the microscopic images.
The AI system achieved an accuracy of 68 per cent, compared to an average of 30 per cent for the specialists. When the three most likely diagnoses were taken into account, Hetairos’s score rose to 84 per cent, whilst the neuropathologists achieved around 50 per cent.
According to the researchers, this demonstrates that modern AI systems are capable of recognising very subtle morphological patterns that are difficult to detect even for experienced specialists. However, they emphasise that very rare tumour types still pose a challenge for the model. As more data becomes available, they expect performance to improve further.
From twelve days to twelve minutes
In a prospective study, Hetairos was used in parallel with standard diagnostic procedures. The system analysed 210 tumour samples without influencing the final diagnosis or treatment. Whereas full molecular diagnostics took an average of twelve days, Hetairos generated a classification within twelve minutes on standard computer hardware. Including preparation and digitisation of the tissue sections, results could often be available within one to two days.
The researchers explicitly view Hetairos as a complement to existing molecular diagnostics, not as a replacement for them. The technology can offer valuable support, particularly in situations where insufficient tumour material is available or molecular tests do not provide a clear-cut result. In addition, the method can reduce the costs of diagnostics. Whilst DNA methylation analyses often cost hundreds of euros, Hetairos makes use of histological specimens that are already available.
According to the research team, the study highlights the potential of AI-supported digital pathology to make high-quality diagnostics faster, cheaper and more accessible worldwide. For patients with brain tumours, this could ultimately lead to a faster diagnosis and, consequently, a quicker start to appropriate treatment.
Brain tumor diagnosis
In 2025, researchers at Harvard Medical School developed PICTURE, an AI-powered diagnostic tool designed to support brain tumor diagnosis during surgery. The system can accurately distinguish between glioblastoma and primary central nervous system lymphoma (PCNSL), two cancers that require very different treatments. In international testing, PICTURE achieved more than 98% accuracy, outperforming both human pathologists and existing AI models. A key feature is its built-in uncertainty detector, which alerts clinicians when the AI is not confident in its assessment.
The tool can also identify dozens of other central nervous system tumors and flag unusual cases for expert review. Researchers believe PICTURE could improve intraoperative decision-making, expand access to specialist-level neuropathology and support pathology training. Further validation in more diverse patient populations is still needed before broader clinical adoption.
