PL01.4.A COMPUTATIONAL HISTOPATHOLOGY ENABLES HIGH-GRANULARITY DIAGNOSTICS IN CNS TUMOURS

• Published in: Neuro-oncology (Charlottesville, Va.) Va.), 2024-10, Vol.26 (Supplement_5), p.v1-v1
• Main Authors: Shmatko, A, Jin, D, Patel, A, Rahmanzade, R, Wick, W, Krieg, S, Pfister, S M, Jones, D T W, von Deimling, A, Gerstung, M, Sahm, F
• Format: Article
• Language: English
• ISSN: 1522-8517; 1523-5866
• DOI: 10.1093/neuonc/noae144.001

Abstract BACKGROUND The recent WHO classification of CNS tumours has increased the number of molecular markers recommended for making a WHO-compatible diagnosis. Particularly, DNA methylation-based classification pioneered by the Heidelberg MNP classifier has emerged as a powerful tool for CNS tumour diagnostics. However, high costs, both in terms of time and resources, along with the need for skilled personnel and expensive equipment, have limited the adoption of such techniques to larger centres. Computational methods like deep learning have proven useful in tasks such as tumour subtyping, identifying genetic alterations and predicting treatment responses. The recent advent of self-supervised learning and multiple-instance learning (MIL) has enhanced the accuracy and generalisation of these models, simultaneously reducing the dataset size requirements, thus bringing computational histopathology closer to clinical applications. MATERIAL AND METHODS To enable broader accessibility of precision diagnostics for CNS tumours and alleviate the need for molecular profiling, we introduce Paiean (Precision AI-enabled Neuropathology) - a computational histopathology pipeline that predicts the diagnostically relevant classes in line with the Heidelberg MNP classifier using only H&E stained whole-slide images (WSI). We designed a neural network architecture implemented in a MIL paradigm. Paiean employs a CTransPath pretrained feature extractor to identify and analyse relevant histological patterns and was fine-tuned on an in-house dataset of CNS tissue WSIs using the MoCo v3 contrastive objective. We trained the model on a dataset of 8,000 H&E stained FFPE WSIs with matched methylation data from the Dept. of Neuropathology, University Hospital Heidelberg. We validated the model on 5 independent external datasets. Additionally, we compared the model to evaluations by 9 board-certified pathologists from across the world. RESULTS Paiean could reliably predict 108 classes of CNS tumours in 72.19% cases. Moreover, for high-confidence predictions (confidence > 0.5), the accuracy increased to 87.52%. We achieved an accuracy of 65.12% on the multi-center external datasets. The model outperformed the human evaluators by an average of 30.17%. CONCLUSION Our work demonstrates that deep learning can supplement conventional methods for molecular diagnostics in CNS tumours. Paiean requires only H&E WSI, covers 108 diagnostically relevant tumour types and has a turnaround time of u