Hetairos AI classifies brain tumors in 12 minutes with 87% accuracy, outperforming specialists

Hetairos AI Transforms Brain Tumor Diagnosis with Speed and Precision

A team led by Moritz Gerstung at the German Cancer Research Center (DKFZ) and Felix Sahm at Heidelberg Medical Faculty has developed Hetairos, a histology-based artificial intelligence model that promises to accelerate CNS tumor classification worldwide1

1

. The system analyzes whole-slide images of H&E-stained FFPE tissue sections to identify 102 different molecular subtypes of central nervous system tumors, covering nearly the entire spectrum of the current WHO classification2

2

.

Trained and validated on more more than 11,000 digitized tissue sections from 9,606 patients across 11 medical centers on four continents, Hetairos AI delivers results that could fundamentally change cancer diagnostics1

1

2

2

. The training dataset from the University Hospital Heidelberg included CNS tumors from all age groups, with molecular data annotated to comply with the 2021 WHO classification requirements.

Unprecedented Accuracy in Predicting Central Nervous System Tumor Subtypes

When Hetairos makes high-confidence predictions—which occurs in approximately 50 to 70 percent of cases—it achieves an accuracy rate of 87 to 88 percent2

2

. This performance level represents a significant advance in AI system classifies brain tumors technology. Even when uncertain, the system narrows down possible diagnoses, helping neuropathologists select appropriate follow-up tests2

2

.

The AI model processes standard microscopic tissue sections by first dividing scanned slides into nonoverlapping areas of approximately 128 × 128 µm, then extracting computational features using the Prov-GigaPath vision transformer model1

1

. These 1,536-dimensional features are aggregated into slide-level embeddings that predict each of the 102 tumor subtypes. Hetairos can also incorporate patient age and anatomical tumor location to enhance performance.

AI Outperforms Experienced Neuropathologists in Direct Comparison

In a head-to-head assessment, five experienced neuropathologists from international centers evaluated 210 cases using only tissue sections. Hetairos achieved 68 percent accuracy compared to the specialists' average of 30 percent2

2

. When considering the three most likely diagnoses, the AI scored 84 percent while specialists reached approximately 50 percent2

2

.

"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," says Felix Sahm2

2

. The system's ability to identify molecular subtypes from histology alone addresses a critical need, as traditional methylation analysis requires specialized laboratories, expensive equipment, and sufficient genomic material—resources unavailable in many regions2

2

.

Twelve Minutes Instead of Twelve Days for Brain Tumor Diagnosis

In a prospective evaluation involving 210 tumor samples in routine diagnostic settings, Hetairos generated findings in just 12 minutes on standard computer hardware after digitizing the tissue sections2

2

. Complete molecular diagnostics using traditional methylation analysis typically takes about 12 days2

2

. Including preparation and digitization, results could often be available within 24 hours to two days.

This speed advantage matters particularly when methylation analysis yields ambiguous results or when insufficient tumor material prevents comprehensive genetic testing1

1

. Hetairos can help resolve these difficult cases by providing molecular predictions directly from routine histology. The system also generates prediction maps highlighting image areas most indicative of specific tumor classes, allowing doctors to understand the basis of diagnoses and identify regions suitable for further investigation2

2

.

"The study shows that artificial intelligence is capable of deriving molecular information directly from routine tissue sections and thus fundamentally changing cancer diagnostics," notes Darui Jin, one of the lead authors2

2

. While rare tumor types still pose challenges for the system, researchers expect performance to improve with larger and more diverse datasets2

2

.