Hetairos

Two central nervous system tumors can look nearly identical under hematoxylin and eosin and still belong to different methylation classes, with different prognoses and treatment paths. DNA methylation profiling separates them — the 2021 WHO classification defines many CNS entities by it — but it costs roughly €400 per case, needs more tissue than a stereotactic biopsy often yields, and takes about two weeks. Hetairos learns that taxonomy's decision boundaries from the slide already under the microscope, predicting 102 methylation-defined CNS tumor subtypes directly from a scanned H&E section of formalin-fixed, paraffin-embedded tissue.

The model was developed at the German Cancer Research Center (DKFZ) with the Department of Neuropathology at Heidelberg University Hospital, Beihang University and Heidelberg University, and published in Nature Cancer in June 2026. The name expands to "H&E Trained AI for Recognition of Oncology Slides"; an earlier project name, Paion ("Precise AI enabled neuro-ONcology"), survives as a commented-out line in the repository header.

The trained artifact is a slide-level aggregator, not a new image encoder. Tiles are embedded by Prov-GigaPath, which stays frozen, and Hetairos learns how to pool those embeddings into a subtype call; UNI is a supported substitute and performed similarly in the authors' experiments. One globally trained model — a ten-member ensemble of the aggregator — was applied unchanged to ten external cohorts and to prospective routine diagnostics.

#Key Features

• 102-subtype output at diagnostic granularity: The label space consolidates the 184 methylation classes of the Heidelberg CNS tumor methylation classifier v12.8 into 102 clinically relevant subtypes and 34 superfamilies, covering pediatric and adult CNS tumors.
• Calibrated confidence: Predictions above 0.5 are flagged high-confidence; accuracy tracks confidence closely enough to be usable for triage, and stays conservative on unfamiliar cohorts.
• Prediction heatmaps: Sliding 5×5-tile windows produce local predictions that compose into a map of which regions support which subtype, guiding where to sample for follow-up molecular testing.
• Optional clinical covariates: Patient age (sinusoidal encoding) and anatomical tumor location (one-hot, seven compartments) can be concatenated into the slide representation to improve accuracy.
• Commodity hardware: Training and inference run on a single GPU with 11 GB of memory; a slide takes about 12 minutes end to end, at an estimated €1–2 per case.

#Technical Details

Whole-slide images are segmented for tissue, tiled into non-overlapping 256×256-pixel patches at 20× (0.5 µm per pixel), and filtered for blur and low tissue content. Each tile is embedded into 1,536 dimensions by Prov-GigaPath, a ViT-G/14 pretrained with DINOv2. Aggregation uses a modified TransMIL — two transformer blocks with Nyström attention — with the slide matrix randomly split into three submatrices as regularization; each submatrix yields a 512-dimensional class token, and the concatenated tokens form the slide embedding. Local and slide-level MLP heads are supervised against the same label, and a contrastive term pulls sample representations toward exponentially-moving-average class templates. Training uses the methylation classifier's probability vector as a soft label, a Lookahead optimizer at learning rate 1e-5, and Mixup in embedding space. Development used 6,115 slides from 4,961 patients at Heidelberg; ten external cohorts added 4,645 tumors and 5,498 slides. Top-1 accuracy was 0.75 internally and 0.68 externally, with high-confidence accuracy of 0.88 and 0.87. Against five board-certified neuropathologists on 210 slides, Hetairos scored 0.68 top-1 versus their 0.30 average.

#Applications

Hetairos is a triage tool inside the existing neuropathology workflow, not a replacement for it. On the roughly 30% of cases histology and immunohistochemistry cannot resolve, it narrows 102 possible diagnoses to one or three before any molecular test is ordered, turning a broad panel into a single stain or gene assay. It also covers cases the array cannot: of 96 specimens with too little tissue for methylation analysis, largely stereotactic biopsies, it called 76 correctly. In a prospective evaluation at Heidelberg from August 2024 to June 2025, its high-confidence top-1 prediction agreed with the eventual integrated diagnosis in 120 of 133 cases, delivered in minutes rather than the 12-day average turnaround.

#Impact

The work demonstrates that methylation-class granularity is recoverable from routine slides across the whole CNS tumor spectrum, not just within diffuse gliomas or a handful of entities, and that a single fixed checkpoint transfers across scanners, staining protocols and continents when its confidence estimates are honest. Predicted subtypes stratified overall survival in medulloblastoma and ependymoma cohorts in the expected direction. The limits are stated plainly by the authors: high-confidence predictions cover only 50–70% of cases depending on cohort, accuracy in the low-confidence band drops to about 0.45, and the 12 subtypes with mean confidence below 0.25 all had fewer than 20 training examples — so rare entities, frozen sections and molecularly ambiguous cases still need the array. The code is released under MIT; the training and internal validation slides are routine clinical material and are not redistributable.