Sponsored by
Researchers in Germany have developed a deep-learning AI system called “Hetairos” that can predict the molecular classification of brain and spinal cord tumors in minutes using standard histology slides. The system was trained on over 11,000 digitized tissue sections from 9,606 patients across 11 centers on four continents, with ground-truth diagnoses based on DNA methylation profiling, which is currently the gold standard for CNS tumor classification. It can distinguish 102 molecular tumor subtypes, covering nearly the full WHO classification spectrum for CNS tumors.
In a head-to-head comparison with five expert neuropathologists across 210 challenging cases, Hetairos achieved significantly higher diagnostic accuracy (68% vs ~30% for humans), and 84% when considering top-three predictions. In prospective testing, it produced molecular subtype predictions in about 12 minutes, compared to roughly 12 days for standard methylation-based diagnostics.
The AI also includes a confidence scoring system, with high-certainty predictions in 50-70% of cases reaching ~87-88% accuracy, and it can narrow broad differential diagnoses into a small number of likely subtypes. Importantly, it highlights the specific tissue regions driving its decisions, improving interpretability and clinical trust.
While this system shows great potential, additional validation in independent cohorts and further regulatory evaluation will be required before it can be approved for routine diagnostic clinical use.