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

Hetairos is a histology-based artificial intelligence model for predicting central nervous system tumor methylation subtypes - Nature Cancer

Here, we describe Hetairos, an AI-based model that classifies whole-slide images of H&E-stained FFPE tissue sections into 102 subtypes of CNS tumors. Hetairos was trained and evaluated on 9,606 tumors, comprising more than 11,000 slides from 11 different institutions across four continents. Hetairos provided high-confidence predictions on 50-70% of the slides, with an accuracy of 0.87. The algorithm outperformed human neuropathologists in a side-by-side assessment of 210 cases selected from the full spectrum of CNS tumors. Lastly, we prospectively evaluated Hetairos in a routine diagnostic setting involving 210 tumor samples. Our results suggest that AI-based predictions can lead to faster diagnoses and help resolve cases with ambiguous results in methylation analysis or insufficient genomic material. Hetairos is trained to predict 102 tumor subtypes Hetairos was built using 6,115 slides from 4,961 patients with CNS tumors at the Department of Neuropathology, University Hospital Heidelberg (UKHD). All cases were annotated with molecular data to comply with or exceed the requirements of the 2021 WHO classification (CNS5+). This cohort includes CNS tumors from all age groups and is intended to largely mirror the incidence of tumor subtypes, with deliberate enrichment of some rare tumor types. Twenty percent of the UKHD dataset was used for internal validation. Hetairos was subsequently validated on ten external cohorts from four continents, comprising an additional 4,645 tumors and 5,498 slides (Fig. 1a). For each tumor, one or multiple slides containing H&E-stained tissue sections scanned at a magnification of at least 20× (approximately 0.5 µm per pixel) were available, together with the paired methylation-based molecular classification (Methods). Methylation classifications were generated using version 12.8 of the Molecular Neuropathology methylation classifier. The predicted classes were further grouped -- incorporating expert neuropathological input -- into a simplified classification system of 102 diagnostically relevant tumor subtypes and 34 superfamilies, covering the full spectrum of CNS tumor subtypes (Methods, Fig. 1b and Supplementary Table 1). These 102 subtypes consolidate a range of provisional tumor subtypes from the full set of 184 methylation-based classes, which are either very rare or whose clinical relevance has yet to be established. The aggregated methylation-based subtypes were used as the ground truth for training Hetairos. In line with epidemiology, the distribution of tumor classes in the UKHD dataset followed a long-tailed pattern, with around 30% of cases belonging to the superfamily of adult-type diffuse gliomas, including isocitrate dehydrogenase (IDH)-wild-type glioblastoma (Fig. 1b). Another 50% of cases originated from the superfamilies of ependymomas, meningiomas, low-grade glial/glioneuronal tumors, pediatric high-grade gliomas, and medulloblastomas. The Hetairos model predicted each of the 102 CNS tumor subtypes from H&E whole-slide scans (Fig. 1c). Because the scanned slides were too large to be processed in a single step, they were first tiled into nonoverlapping areas of approximately 128 × 128 µm, followed by computational feature extraction using the Prov-GigaPath vision transformer model. The resulting 1,536-dimensional features per tile were aggregated into slide-level embeddings to predict each of the 102 CNS tumor subtypes using a modified TransMIL model (Methods). This aggregation step enabled Hetairos to learn which image tiles are predictive and to focus on the most informative image areas. Additionally, patient age and anatomical tumor location can be incorporated to enhance model performance (see the ablation results in Supplementary Table 2). The model outputs the probability of each of the 102 tumor subtypes. By considering only local subsets of tiles, Hetairos can generate prediction maps that highlight the image areas most indicative of given tumor classes. More details are provided in the Methods. Accurate classification of CNS tumors and identification of ambiguous cases Over the past decade, the discovery of new tumor classes has been primarily based on molecular analyses, including methylation profiling. This approach consistently reveals clusters of tumors with shared underlying biological relationships that often appear morphologically unrelated. Intriguingly, a visualization of slide-level embeddings for the internal validation cohort using UMAP revealed that Hetairos learned internal representations that similarly cluster different groups of tumors (Fig. 2a). While the clustering is not yet as distinct as that found by methylation analysis, the emerging structures reflect the histopathological similarity of different tumor superfamilies, such as glial tumors, ependymal tumors, meningiomas and medulloblastomas. The visible clustering of subtypes also reflects Hetairos's ability to distinguish different tumor classes. Hetairos assigns a probability to each possible class, which usually centers around a single class or a few related classes. Of greatest interest is typically the class with the highest probability, which we term the top-1 prediction, and the corresponding estimated probability, which we refer to as Hetairos's confidence. Top-1 predictions agreed with the methylation classification results (methylation score > 0.8) in 75% of all internal validation tumors (Fig. 2b) and remained robust in the presence of common histological artifacts (Supplementary Tables 3 and 4); additional performance metrics are provided in Supplementary Table 5. In 87% of cases, the true class label was among Hetairos's three most likely classes (top-3 accuracy). Reassuringly, incorrect predictions were typically assigned lower confidence scores, showing a conservative tendency that is important to avoid confident mispredictions (Fig. 2c). In line with the methylation classifier, it is important for users of Hetairos to know the confidence levels at which the predictions are typically correct. We designated cases with confidence above 0.5 as 'high confidence' and those with confidence below 0.5 as 'low confidence'. High- and low-confidence cases comprised 70% and 30% of tumors in the internal validation cohort, respectively (Fig. 2d). Hetairos's top-1 accuracy among the high-confidence set was 0.88, demonstrating that the model can deliver an accurate and detailed initial diagnosis in the majority of cases, with accuracy further increasing as the confidence threshold rises (Supplementary Table 6). In the low-confidence set, the accuracy dropped, as expected, to 0.46. When combining the three most likely predictions for the low-confidence set, the accuracy was still 0.71, which shows that Hetairos can often meaningfully reduce the set of possible diagnoses from 102 subtypes to just 3 even in low-confidence cases. Such narrowing of probable classes may help guide further diagnostic tests, potentially resolving the differential diagnoses with only a few or even a single immunohistochemistry test, single gene assay or chromosomal hybridization, rather than through high-throughput testing. For half of the errors Hetairos made on the high-confidence cases, the correct class belonged to the same superfamily as the predicted class, resulting in an accuracy of 0.94 at the superfamily level. This pattern reflects the close morphological similarity of these entities and their consistent confusion, even within the hierarchical classification systems (Extended Data Fig. 1a-f). Among the low-confidence set, errors tended to occur across superfamilies and subtypes. In such errors, classes with higher incidence in the training data were predicted more often and with greater confidence. All 12 tumor subtypes with an average confidence below 0.25 had fewer than 20 occurrences in the training set (Figs. 2d,e and 3). Data augmentation strategies, including oversampling and color space transformation, appeared insufficient to further improve performance on these underrepresented classes (Extended Data Fig. 2a-h). These findings highlight the need for large datasets to confidently classify very rare tumor subtypes, such as liponeurocytomas, atypical teratoid/rhabdoid tumors and germinomas. Hetairos's accuracy is preserved in external validation cohorts To be widely applicable, digital pathology models must maintain predictive performance across centers, demographics and processing protocols. To evaluate Hetairos's performance in settings unobserved during training, we assembled ten validation cohorts from different institutions across four continents, comprising 4,645 cases and 5,498 slides covering the incident spectrum of CNS tumor diagnoses (Fig. 4a,b). Hetairos's overall accuracy was lower in external validation cohorts than in the internal validation cohort (68% versus 75%) (Fig. 4c, Extended Data Fig. 3 and Supplementary Table 5). Reassuringly, however, this discrepancy was largely predicted, as the fraction of low-confidence predictions increased from 30% to 45% (Fig. 4d). This discrepancy may in part be attributed to modality differences between the external and internal validation sets (Extended Data Fig. 4a,b and Supplementary Table 7). Among the 55% high-confidence cases, the top-1 accuracy remained 0.87 (0.96 at the superfamily level), similar to the performance achieved in the internal validation cohort. Among low-confidence cases, the prediction accuracies were 0.45 for top-1 and 0.71 for combined top-3 predictions, consistent with those of low-confidence samples in the internal validation cohort. This provides evidence that Hetairos recognizes differences in slide characteristics and adjusts its confidence levels accordingly. The change in average confidence per tumor class was -0.10 (range -0.53 to +0.30) and was usually accompanied by a corresponding change in observed accuracy, thereby maintaining calibration (Fig. 4e). While 59 of 79 tumor types exhibited a drop in confidence, there was also an increase in confidence and accuracy for 8 of the 79 tumor types, including medulloblastoma group 3 and schwannoma. The external cohorts have distinct tumor class compositions due to the specialization of the individual centers. Nonetheless, the drop in confidence was of a similar magnitude across subcohorts. The accuracy on the respective high- and low-confidence cases matched or exceeded the model's confidence, indicating that Hetairos's predictions remain conservative (Fig. 4f,g). Lastly, Hetairos was evaluated qualitatively on the EBRAINS Digital Brain Tumour Atlas (DBTA) cohort, which comprises 3,110 slides lacking methylation-based predictions. Instead, Hetairos predictions were compared to the detailed histopathological diagnoses provided (Extended Data Fig. 5a,b). Specific histology class names and their corresponding color codes are detailed in Supplementary Table 8. In this cohort, Hetairos's confidence distribution was similar to that in other external validation cohorts, showing a 50-50% split between high- and low-confidence cases, with corresponding accuracies of 85.6% and 50.2%, respectively. The predicted tumor types showed good agreement with histopathological diagnoses, with discrepancies mostly occurring within tumor families. Additional results on external samples with low methylation scores (<0.8, with an average of 0.45) are provided in Extended Data Fig. 5c,d (note that methylation class annotations may differ from the final diagnoses for these cases). Hetairos outperforms neuropathologists in H&E assessment Identifying and narrowing differential diagnoses from H&E slides is an important first step in the diagnostic workflow and is pivotal for the efficient selection of subsequent diagnostic tests. While many institutes in high-income countries typically have a set of special stains (for example, PAS and reticulin) and immunohistochemistry tests (for example, GFAP, MAP2, NeuN and synaptophysin) readily available at initial inspection to assess tumor lineage, such diagnostic tools are often absent or restricted to a few cases in settings with more limited resources. Hence, we conducted a blinded side-by-side evaluation of 210 slides by five board-certified neuropathologists and Hetairos. The slides were selected to have a similar number of cases from each class (Supplementary Table 9). Neuropathologists were provided with a drop-down list containing the 102 methylation subtypes, which correspond directly to Hetairos's output classes, and were asked to choose and rank their top-3 diagnoses from it. The neuropathologists participating in the evaluation had prior experience with methylation classification and were thus familiar with the provided classes. Based on H&E-stained sections only, Hetairos's accuracy was consistently better than that of neuropathologists, who achieved an average top-1 accuracy of 0.30 (0.18-0.36) compared to Hetairos's 0.68 (see the additional metrics provided in Supplementary Table 5). This gap narrowed when assessing top-3 accuracy, which was 0.50 (0.31-0.70) for humans and 0.84 for Hetairos (Fig. 5a). Human evaluators often appeared to struggle with identifying the single best choice among a large number of granular classes, but they were able to provide a plausible set of diagnoses. While Hetairos provided a calibrated range of confidence levels, it consistently outperformed human accuracy across all confidence intervals. The performance gap between human evaluators and Hetairos narrowed slightly within Hetairos's lower confidence range (Fig. 5b). When considering individual tumor subtypes, Hetairos generally outperformed humans for classes represented by more than ten cases in the training cohort (Fig. 5c). For rare tumor subtypes with fewer than ten cases in the training cohort, human pathologists performed similarly to Hetairos. For some diagnoses, such as metastatic melanoma and teratoma -- which Hetairos struggled to diagnose -- human diagnoses were correct in two out of three cases and one out of one case, respectively. Together, these results indicate that Hetairos is currently better at diagnosing all but the rarest types of tumors. AI-assisted diagnosis reaches methylation-level accuracy in 12 min Named after the Greek term for 'companion', Hetairos is designed to assist neuropathologists in diagnostic work. As mentioned previously, the typical diagnostic workflow in neuropathology begins with a morphological assessment of an H&E section, followed by a series of immunohistochemical tests selected to narrow the differential diagnoses (Fig. 6a). Approximately 30% of cases cannot be resolved in terms of tumor classification and subtyping without advanced molecular testing. Most of these can be resolved by DNA methylation array analysis, whereas some require additional testing, such as DNA and RNA sequencing, to identify pathognomonic mutations or fusions. Some specimens are unsuitable for molecular analysis because of limited sample quantity or quality. Hetairos fits into this workflow as a tool to supplement the first-line method -- that is, manual histopathological evaluation using H&E-stained FFPE slides. Owing to Hetairos's WHO 2021-compatible granularity and well-calibrated predictions, it is intended to be used as a triaging tool to efficiently guide further molecular analyses. The report provided by Hetairos highlights tissue areas of the H&E slide that are most informative for diagnostic prediction and those that contribute least, both in a prediction map and with exemplary magnifications (Fig. 6b; see also the example reports in Extended Data Fig. 6 and Supplementary Table 10). These illustrations enable neuropathologists to review Hetairos's decision-making and may also guide the selection of optimal areas for extraction if further molecular testing is needed. For example, areas that are morphologically informative for the diagnosis of craniopharyngioma or meningioma are robustly separated from adjacent connective tissue. Within the spectrum of glial morphology, pilocytic astrocytoma bulk tissue was clearly distinguished from the surrounding reactive gliosis (Fig. 6b). Similarly, Hetairos uniformly identified the prototypical histology of oligodendroglioma (Extended Data Fig. 6a). In astrocytoma, areas considered by Hetairos to be indicative of the second-best diagnosis, glioblastoma, coincided with microvascular proliferation -- a characteristic feature of both diagnoses and a criterion for grade 4 tumors (Extended Data Figs. 6b and 7a,b). Further examples of intratumoral heterogeneity were observed in meningiomas, where Hetairos frequently identified regions assigned to the benign category within intermediate-group cases and vice versa (Extended Data Fig. 8a-d). The biological underpinning of the distinct grading classes found within the same tumor was supported by Ki-67 staining patterns (that is, weaker staining in areas predicted to be benign). Taken together, these examples illustrate that Hetairos has the potential to capture subtle yet clinically relevant and biologically meaningful histologies that exhibit intratumoral heterogeneity and pose diagnostic challenges. Even advanced technologies such as methylation analysis may sometimes fail to provide a clear prediction on their own, thus necessitating additional testing (for example, for fusions or mutations). Within a cohort of 50 samples diagnosed solely based on a combination of molecular assays, Hetairos correctly predicted 27 cases, demonstrating superiority over methylation analysis in some scenarios (Fig. 6c). For 96 specimens in which methylation analysis could not be performed because of limited tissue, particularly stereotactic biopsy samples, Hetairos correctly predicted 76 cases (Fig. 6d). This highlights the robustness and predictive capability of Hetairos in specimens with limited tumor content (Extended Data Fig. 9a,b). To assess the potential clinical utility of Hetairos, the algorithm was prospectively evaluated alongside routine diagnostics from August 1, 2024, to June 1, 2025, at the Department of Neuropathology at UKHD. All cases that required molecular testing and met the inclusion criteria were included without any further selection. During this time, Hetairos was used to predict a total of 210 cases (Fig. 7a and Supplementary Table 11). Its results were compared to independent integrated diagnoses established by a combination of morphological assessment, immunohistochemistry, methylation classification, next-generation sequencing panel and RNA sequencing. Hetairos predictions were not made available to the neuropathologist and did not influence diagnostic or treatment decisions. On average, it takes 12 days from the receipt of the neurosurgical specimen to an integrated diagnosis. As illustrated in Fig. 7a, this corresponds to approximately 16 days for cases requiring molecular testing. Hetairos, running on consumer-grade hardware, took an average of 12 min to process a slide and generate the report. This indicates that, together with the time taken for staining and scanning, Hetairos substantially shortens the turnaround time, with results usually available within 24 h or up to 2 days after sample receipt, depending on fixation time. Among cases that could not be resolved by histology or immunohistochemistry alone, Hetairos yielded 63% high-confidence predictions and 37% low-confidence predictions. Hetairos's top-1 predictions were found to agree with the eventual integrated diagnosis in 120 of 133 high-confidence cases (90.2%; Fig. 7a), highlighting Hetairos's ability to deliver near-methylation accuracy within a substantially shorter timeframe. For cases with high methylation scores and high confidence, Hetairos achieved an accuracy of 94.3% (100 of 106), while for those with low Hetairos confidence, the accuracy was 45.5% (35 of 77). In this subgroup with low methylation scores, Hetairos and the methylation classifier showed comparable accuracy against the integrated diagnosis (Fig. 7b). Notably, among the low-methylation-score cases that were concordant with Hetairos's high-confidence predictions, the accuracy reached 88.9% (16 of 18; Fig. 7c), further demonstrating how Hetairos can aid diagnostic decision-making when molecular results are inconclusive. Granular tumor classification stratifies survival Multiple studies have underscored the need for a granular classification of CNS tumors to reflect considerable differences in survival. To demonstrate the prognostic utility of Hetairos's detailed classification, we used data from 353 patients with CNS tumors in the MNP 2.0 trial, for whom digital H&E images, methylation classification and survival data are available. Among these data, Hetairos classified 165 cases into one of the four WHO-defined subtypes of medulloblastoma: WNT activated, sonic hedgehog (SHH) activated, group 3 and group 4 (Fig. 8a). Compared to the methylation class, the accuracies were 89% and 51% for high- and low-confidence predictions, respectively. Despite some misclassification among low-confidence predictions, Cox proportional hazards models confirmed that Hetairos's subtypes exhibited notable differences in overall survival (P = 0.03). The 3-year overall survival rates were 58% for tumors classified as group 3 medulloblastoma, 81% for SHH-activated medulloblastoma, 88% for group 4 medulloblastoma and 100% for WNT-activated medulloblastoma. These results are in agreement with previous findings. Similarly, Hetairos predicted 90 cases as five different WHO subtypes of ependymoma (Fig. 8b). The accuracy of high-confidence and low-confidence predictions was 100% and 48%, respectively. Overall survival among the predicted subtypes differed in the expected manner (P = 0.07). Group PFA posterior fossa ependymoma and ZFTA fusion-positive supratentorial ependymoma had poorer prognoses, with 3-year survival rates of 89% and 68%, respectively, compared to the other three ependymoma subtypes, all of which had 3-year survival rates of 100%. Lastly, we applied Hetairos to classify samples histologically diagnosed as high-grade gliomas into more detailed subtypes. These subtypes were grouped into high- and low-risk categories based on independent prior knowledge (Methods). Notwithstanding the challenges of accurately classifying glioma subtypes, the survival curves exhibited the expected trend of worse overall survival in high-risk groups (3-year survival: 38% versus 55%; P = 0.2; Fig. 8c). Additionally, a multivariate survival analysis confirmed that Hetairos provides stronger prognostic value than the clinical baseline alone, offering a more effective alternative in the absence of molecular testing (Fig. 8d).

New AI system classifies brain tumors with unprecedented accuracy

German Cancer Research Center (Deutsches Krebsforschungszentrum, DKFZ)Jun 10 2026 Experts in Heidelberg have developed an AI system that can classify brain tumors with unprecedented accuracy using standard microscopic tissue sections. Using digitized standard stains, the system identifies more than 100 molecular subtypes of central nervous system tumors, delivers results within minutes, and could accelerate the diagnosis of brain tumors worldwide. Tumors of the brain and spinal cord are extremely diverse. In recent years, it has become clear that many of these tumors can only be reliably diagnosed if their molecular properties are examined in addition to their microscopic appearance. Of particular importance here is so-called DNA methylation analysis, which is now considered the gold standard for the accurate classification of many brain tumors. However, such tests are complex: they require specialized laboratories, expensive equipment, and sufficient tumor material. In addition, it often takes about two weeks for the results to become available. In many regions of the world, the necessary technologies are not even available. AI learns from over 11,000 tissue sections A new AI system called "Hetairos" is expected to bring about substantial improvements. It was developed by a team led by Moritz Gerstung (German Cancer Research Center, DKFZ) and Felix Sahm (Heidelberg Medical Faculty of Heidelberg University and Heidelberg University Hospital). The goal of the project was to predict which molecular subgroup a tumor belongs to based solely on routinely prepared and stained histological sections. Hetairos was trained and validated using more than 11,000 digitized tissue sections from 9,606 patients. The diagnoses were primarily determined using DNA methylation diagnostics. The data came from eleven medical centers on four continents. In total, Hetairos distinguishes 102 different molecular tumor subtypes, covering nearly the entire spectrum of the current WHO classification of central nervous system tumors. The AI not only evaluates its diagnosis but also indicates how confident it is in it. In approximately 50 to 70 percent of all cases, Hetairos made predictions with a high degree of certainty. In these cases, accuracy was around 87 to 88 percent. Even when the AI was uncertain, it was usually able to significantly narrow down the number of possible diagnoses. Instead of having to distinguish between more than a hundred tumor subtypes, Hetairos often provides neuropathologists with only a few likely candidates. This can significantly simplify the selection of further diagnostic tests. The study shows that artificial intelligence is capable of deriving molecular information directly from routine tissue sections and thus fundamentally changing cancer diagnostics." Darui Jin, one of the lead authors of the study Hetairos outperforms experienced specialists Particularly noteworthy was the direct comparison with human experts. Five experienced neuropathologists from various international centers were given 210 cases and asked to make a diagnosis based solely on the tissue sections. Hetairos achieved an accuracy rate of 68 percent, while the specialists averaged 30 percent. When considering the three most likely diagnoses in each case, the AI scored 84 percent, while the specialists scored 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," says Felix Sahm. "Currently, the diagnosis of very rare tumor types still poses a major challenge for Hetairos; in this regard, experienced neuropathologists appear to be at least on par. However, we expect the system's performance to improve even further with larger and more diverse datasets," adds Moritz Gerstung. Diagnosis in twelve minutes instead of twelve days In a prospective study, Hetairos was used in parallel with routine clinical practice. The system analyzed 210 tumor samples without the AI result influencing the actual diagnosis or treatment decision. While complete molecular diagnostics took an average of about twelve days, Hetairos generated its findings in just twelve minutes on standard computer hardware after digitizing the stained tissue sections. Including preparation and digitization of the tissue sections, results could often be available within 24 hours to two days. Assistance with difficult and unclear cases Hetairos could be particularly valuable in situations where traditional molecular methods reach their limits, when there is insufficient tumor material for genetic testing, or when molecular tests do not yield clear results. In addition, the system highlights the areas in the tissue section that were particularly important for its decision. This allows doctors to understand the basis of the AI's diagnosis and identify which regions may be suitable for further investigation. "We developed Hetairos primarily as a tool to support diagnostics," explains neuropathologist Felix Sahm. "It is not intended to replace molecular analyses, but rather to specifically complement and accelerate them. The technology could make an important contribution, particularly in countries or regions with limited resources, as it is based on standard tissue sections used worldwide." The method could also offer economic advantages. While a DNA methylation analysis typically costs several hundred euros, Hetairos uses existing tissue sections for its analysis. Moritz Gerstung confirms: "Hetairos demonstrates the enormous potential of AI-supported digital pathology to provide rapid and widely available diagnostic methods that were previously only possible with considerable technical effort." Source: German Cancer Research Center (Deutsches Krebsforschungszentrum, DKFZ) Journal reference: Jin D., et al. (2026) Hetairos is a histology-based artificial intelligence model for predicting central nervous system tumor methylation subtypes. Nature Cancer. DOI: 10.1038/s43018-026-01186-3. https://www.nature.com/articles/s43018-026-01186-3