AI classifies brain tumors in 12 minutes, outperforms pathologists

Clinical takeaway: For brain tumor classification, this AI model points toward a future where a 12-minute result from a standard tissue slide replaces a two-week wait for molecular testing, at a fraction of the cost and without specialized equipment.

Accurate classification of brain tumors currently depends on molecular testing that takes weeks, costs hundreds of dollars per case, and requires equipment some hospitals do not have. The results with a new AI model suggests that gap may be closeable with tools already sitting in most pathology labs.

In head-to-head testing against five board-certified neuropathologists on 210 cases, the AI model's first-choice diagnosis was correct 68% of the time, compared to an average of 30% for human specialists. When considering each evaluator's top three diagnoses, the AI scored 84% versus 50% for the neuropathologists.

In high-confidence cases, which represented 50 to 70% of all cases depending on the cohort, accuracy reached 87 to 88%. The model correctly classified 102 molecular subtypes of central nervous system tumors across both pediatric and adult cases, validated on more than 11,000 slides from 11 centers on four continents.

The model was trained on slides from a single academic center and validated on ten external cohorts across four continents, covering 9,606 patients and more than 11,000 slides. Methylation-based classification served as the diagnostic ground truth. Prospective testing ran alongside routine diagnostics at one center for ten months. The model performs less reliably on rare tumor subtypes and on slides from institutions not represented in training data.

Rather than replacing molecular testing, the authors position the model as a tool to accelerate and focus it. AI highlights tissue areas most informative for diagnosis, helping pathologists decide where to direct further testing. It also performed well in cases where standard molecular methods hit their limits, including samples with too little tissue for methylation analysis. The authors estimate the cost of running Hetairos at roughly 1 to 2 euros per case, compared to several hundred euros (roughly $300 to $500) for DNA methylation testing.

The authors call for broader multicenter validation and retraining on more diverse datasets to improve performance on rare subtypes and new institutions. Future extensions may enable the model to work with smartphone-adapted microscopes, potentially expanding access in low-resource settings.

"The study shows that artificial intelligence is capable of deriving molecular information directly from routine tissue sections and thus fundamentally changing cancer diagnostics," said lead author Darui Jin, Ph.D., of the German Cancer Research Center.

Source: Jin et al. Nat Cancer. June 10, 2026. Hetairos is a histology-based artificial intelligence model for predicting central nervous system tumor methylation subtypes