517P - Differentiating IDH-wildtype and IDH-mutant high grade gliomas with deep learning

Background

Glioblastomas (GBM) are the most frequent primary malignant brain tumours affecting adults. Over 85% of GBM cases are IDH-wildtype (IDHwt), which carry an expected survival of around 1 year. A small number of cases have the morphological features of GBM but harbour an IDH-mutation (IDHmut). These cases, designated grade 4 IDH-mutant astrocytoma under the 5th edition WHO CNS classification, are important to identify due to superior prognosis and implications for clinical management. In practice, most neuropathology departments use immunohistochemistry (IHC) to determine IDH status. IHC is not infallible, however, and requires additional tissue and laboratory work. Multiple studies have demonstrated that Deep Learning (DL) can predict molecular alteration status directly from routine pathology whole slide images (WSI). Validating the use of AI for such applications has the potential to improve diagnostic precision and add further value to a digital pathology workflow. We hypothesise that DL can accurately differentiate IDHwt and IDHmut high-grade gliomas directly from WSI.

Methods

We used an attention-based multiple-instance learning (attMIL) approach and applied it to an international biorepository of digitised neuropathology images. Our primary cohort was obtained from University College London (UCL), through collaboration with BRAIN UK and comprised 774 WSIs. The Cancer Genome Atlas (TCGA) was used as a secondary cohort for external validation and comprised 332 WSIs.

Results

We found that DL was able to predict IDH status in high-grade glioma images. On internal cross validation in the UCL dataset, this gave an area under the receiver operating characteristic curve (AUROC) of 0.92. These findings were sustained on external validation in the TCGA dataset, with an AUROC of 0.79.

Conclusions

In the future, DL-based image analysis could be used as a screening tool in the setting of a digital pathology diagnostic environment to guide Pathologists on the targeted use of IHC.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

TU Dresden.

Funding

JNK is supported by the German Federal Ministry of Health (DEEP LIVER, ZMVI1-2520DAT111) and the Max-Eder-Programme of the German Cancer Aid (grant #70113864), the German Federal Ministry of Education and Research (PEARL, 01KD2104C), the German Academic Exchange Service (SECAI, 57616814) and the European Union (ODELIA). UCLH Biomedical research centre is funded by the National Institute for Health Research (BRC399/NS/RB/101410). SB is also supported by the Department of Health’s NIHR Biomedical Research Centre’s funding scheme. TM was supported by The Brain Tumour Charity (GN-000389 clinical research training fellowship) and by the National Institute of health research (NIHR) with clinical lecturer fellowship (CL-2019-19-001).

Disclosure

J.N. Kather: Financial Interests, Personal, Invited Speaker: Fresenius, Eisai, MSD; Financial Interests, Personal, Advisory Board: Owkin, DoMore Diagnostics, Panakeia, London, UK. All other authors have declared no conflicts of interest.