AI-based histopathological classification of central nervous system tumours

DNA methylation analysis has become an essential diagnostic assay for classifying tumours of the central nervous system (CNS). However, this test requires additional resources and time compared to conventional histopathological diagnosis using haematoxylin and eosin (H&E) stained tissue sections, which are available globally. Here, we propose to reduce time and resource requirements using Hetairos, an artificial intelligence (AI) algorithm that predicts 102 methylation-based CNS tumour subtypes from digital images of H&E slides. Hetairos is built and assessed using over 11,000 slides from 10 centres across four continents. Hetairos produces well-calibrated probabilities for each subtype. Across cohorts, Hetairos identifies 50-70% of cases, which it classifies with high confidence and an accuracy of 0.87 for its highest-rated predictions. Hetairos achieved higher accuracy than four board-certified neuropathologists in a direct comparison using only histology images (0.68 vs 0.28). A prospective evaluation within a routine diagnostic setting confirmed Hetairos’ performance and highlighted its 2-day turnaround time compared to an average of 12 days for molecular testing. Hetairos can assist diagnostic decision-making by limiting the number of likely diagnoses, informing efficient further testing, and by highlighting tissue areas indicative of specific diagnoses. These capabilities suggest Hetairos is a valuable tool assisting in the diagnosis of the full spectrum of paediatric and adult CNS tumours.