Department of Neurology, Division of Clinical Neurooncology, University Hospital Essen
Background: The determination of isocitrate dehydrogenase (IDH) mutation status plays a crucial role in the diagnosis of glioblastoma. Depending on the age of the patient and the result of the immunohistochemical analysis, additional DNA sequencing may be required to determine IDH mutation status. As DNA sequencing results can occasionally take several days until available, there is a need for inexpensive and fast non-invasive methods. In this work, we investigated whether IDH mutation detection by artificial intelligence (deep learning) from digitized hematoxylin-eosin (H&E) stained sectional specimens is feasible. Methods: Patients with histologically confirmed glioblastoma from The Cancer Genome Atlas cohort were included if digitized H&E stained whole-slide scans with corresponding information on IDH status were publicly available. The total cohort was subdivided into a training, validation, and test cohort in a ratio of 44:33:23. Whole-slide scans were partitioned into tiles of fixed size and used to train a Resnet-34 convolutional neural network. The evaluation of the trained model was performed once on the test cohort using Receiver Operating Characteristic analysis and Area-Under-The-Curve (AUC) metric. To ascertain which regions of the H&E specimens were decisive for the determination of IDH status, the Grad-CAM method was used. Results: 124 patients were included, 29 of which were IDH mutant. The digitized H&E slides had an average size of 2.5 gigabytes per image file and approximately 1000 tiles per slide were prepared. The prediction AUC of the trained model was 0.94. The duration of IDH prediction was about 3.5 seconds per slide. The Grad-CAM evaluation confirmed that the model mainly used cellular regions to collect decision-supporting information. Conclusions: This pilot study shows the promising potential of deep learning for the prediction of IDH mutation status from digitized H&E scans in glioblastoma. To confirm these data, testing this model on an independent cohort is needed.