Weakly Supervised IDH-Status Glioma Classification from H&E-Stained Whole Slide Images

Gliomas are the most common malignant primary brain tumors in the Central Nervous System. Isocitrate dehydrogenase (IDH) mutation is a key driver of gliomagenesis in infiltrating gliomas and a diagnostic criterion with prognostic relevance. Determining IDH status from hematoxylin and eosin (H&E) stained Whole Slide Images (WSI) would be a valuable tool in precision oncology as it would expedite diagnosis and downstream patient treatment. We seek a computational approach to predict IDH status from digitized histopathology glioma sections (WSI). The proposed approach first applies a comprehensive curation step to remove artifactual content on each H&E-stained WSI, followed by a classification mechanism based on weakly supervised attention-based multiple instance learning. Quantitative evaluation on 1,534 digitized tissue sections from 799 cases from the publicly available TCGA-GBM and TCGA-LGG dataset revealed AUC of 0.92 and a balanced accuracy of 0.85 on a 10-fold cross-validation (10CV) schema. Histological prediction of IDH status represents an ideal prototype to interrogate WSI towards identifying clinically-relevant tumor biomarkers.