Answers are grounded in this paper and cite their sources.
AI model classifies CNS tumours with 88% accuracy, reducing reliance on molecular diagnostics and turnaround time.
BACKGROUND DNA methylation profiling is central to CNS tumour diagnostics, providing molecular classification beyond conventional histopathology. However, its use is constrained by extended turnaround time, high costs and the need for specialised equipment. Recent developments in computational pathology have shown that molecular alterations and tumour subtypes can be inferred directly from H&E-stained histological images. The emergence of pathology foundation models has further advanced digital pathology approaches, particularly for challenging entities like CNS tumours. Building on this, we previously introduced Paieon, a model for CNS tumour classification from H&E-stained slides. MATERIAL AND METHODS Extending our previous Paieon work, we present Hetairos, an AI-based model that classifies whole slide images into 102 CNS tumour classes, derived by consolidating the 184 methylation-based classes from the v12.8 Molecular Neuropathology classifier. We optimised the model structure, expanded the external validation cohort, and performed additional prospective clinical testing. Hetairos leverages the Prov-Gigapath Vision Transformer to extract tile-level features at 20× magnification, then aggregated via a modified TransMIL for slide-level predictions. Hetairos was trained and evaluated using a cohort of 9,583 CNS tumours (>11,000 slides) from 10 institutions across four continents (an expansion of over 3,000 slides compared to the Paieon study). Of these, 6,115 slides from UKHD were used for training and internal validation, and 5,460 for external validation. Hetairos was further benchmarked against four board-certified neuropathologists and prospectively evaluated at UKHD in routine diagnostics. RESULTS Hetairos achieved 88% accuracy on high-confidence predictions, which comprise 70% of the cases in the internal validation. In the multi-centre external validation, it maintained 87% accuracy, with 56% of cases meeting the high-confidence threshold. Compared to Paieon, Hetairos showed a 3% improvement in overall accuracy and an 8% increase in the fraction of high-confidence predictions. In the comparison with neuropathologists, Hetairos outperformed the average diagnostic accuracy of the participating experts by a margin of 40% (68% vs 28%). The prospective study further confirmed Hetairos’ performance, with a 2-day turnaround time compared to 12 days for molecular testing. CONCLUSION Hetairos delivers precise diagnoses at clinically relevant granularity using routinely available diagnostic material, which could support broad applicability. Compared to molecular diagnostic tools, it offers substantial advantages in turnaround time and cost. In summary, Hetairos complements existing CNS tumour diagnostics, with the potential to accelerate diagnosis, reduce reliance for molecular testing, and optimise resource utilisation.
No takes yet. Share an insight, caveat, or question.
Jin et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: