In Heidelberg, Germany, a team of researchers has unveiled an AI system that transforms brain tumor diagnosis from a two-week waiting game into a 12-minute answer. The system, called Hetairos, learns from more than 11,000 digitized tissue sections across 9,606 patients from 11 medical centers on four continents and can identify 102 distinct molecular tumor subtypes — nearly the entire spectrum of the WHO classification for central nervous system tumors.
For decades, accurate brain tumor diagnosis has required DNA methylation analysis, a gold standard molecular test that demands specialized laboratories, expensive equipment, and patience. Results typically arrive after about two weeks, if the necessary technology is even available in a patient's region. That gap in access and time has meant delayed treatment decisions and unequal care worldwide. Hetairos, developed by Moritz Gerstung at the German Cancer Research Center (DKFZ) and Felix Sahm at Heidelberg Medical Faculty and University Hospital, was designed to change that by extracting molecular information from routine tissue slides that any pathology lab already uses.
The numbers speak clearly. When the AI expressed high confidence in its diagnosis — roughly 50 to 70 percent of cases — it achieved 87 to 88 percent accuracy. In a head-to-head comparison with five experienced neuropathologists from international centers, Hetairos scored 68 percent accuracy on 210 cases while the human specialists averaged 30 percent. When both were asked to identify the three most likely diagnoses, the AI reached 84 percent while the specialists managed about 50 percent. "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.
Perhaps most crucially, even when Hetairos expressed uncertainty, it proved useful. Rather than leaving pathologists to sort through all 102 possible subtypes, the system narrowed the field to a few likely candidates, dramatically simplifying the next steps of diagnosis. In a prospective study running parallel to routine clinical practice, the AI delivered results in 12 minutes on standard computer hardware after tissue sections were digitized — compared to an average of 12 days for complete molecular diagnostics through traditional methods.
The implications extend far beyond Heidelberg. Brain and spinal cord tumors are extraordinarily diverse, with many requiring molecular analysis to diagnose reliably. In many regions of the world, the specialized labs and equipment that perform DNA methylation testing simply don't exist. Hetairos, relying only on slides prepared and stained using routine protocols, offers a path to democratizing accurate diagnosis.
Researchers acknowledge remaining challenges. Rare tumor types still pose difficulties for the AI, where experienced pathologists remain competitive. But Moritz Gerstung expects performance to improve as the system trains on larger and more diverse datasets. The study, published in Nature Cancer, marks a milestone in how artificial intelligence can reshape cancer care — not by replacing human expertise, but by amplifying it and making life-saving diagnosis accessible to more patients, faster, across more of the world.