Abstract BACKGROUND Infiltrating gliomas are the most common primary adult brain tumors. Isocitrate dehydrogenase (IDH) mutation is a key driver of gliomagenesis in 25–30% of infiltrating gliomas and correlates with favorable prognosis when compared with the histologically-similar but biologically-distinct IDH-wildtype glioblastoma. We sought an interpretable computational pipeline to predict IDH status from digitized histopathology glioma sections (Whole Slide Images - WSI) and identify associated cell-level morphology. Materials and METHODS We identify 799/1122 TCGA-GBM and TCGA-LGG tumors (1534 WSIs) with available IDH status. 379 of these tumors (756 WSI) are IDH-wildtype and 420 tumors (778 WSI) are IDH-mutant. Each H&E-stained WSI undergoes comprehensive curation to remove artifactual content before classification. Our interpretability mechanism is based on weakly-supervised attention-based multiple-instance-learning. IDH-associated morphology in each WSI is evaluated by a HoverNet model from the 50 highest-attention tiles. Classification performance was assessed using 10-fold cross-validation, with training (80%), validation (10%), and test (10%) sets stratified on the patient/tumor level, without controlling for grade. An additional out-of-sample hold-out test set of 82 IDH-wildtype and 32 IDH-mutant tumors from the University of Pennsylvania Health System was further used for independent evaluation. RESULTS Quantitative performance evaluation revealed AUC=0.92 and Balanced Accuracy=0.84 on the TCGA test set, and AUC=0.88 and Balanced Accuracy=0.80 on the independent hold-out UPenn set. When comparing identified IDH-wildtype with IDH-mutant tumors (and while excluding glioblastoma with lower-grade histology), the former shows significantly (p< 0.00005) higher distribution of neoplastic and necrotic cells and significantly (p< 0.00003) lower distribution of inflammatory and connective cell type tissue. CONCLUSION Histological prediction of IDH status represented an ideal prototype to interrogate WSI towards identifying clinically-relevant tumor biomarkers. Detailed interpretation of results reveals computational predictions being driven by human-identifiable features at the patch level. This study robustly identifies morpho-IDH correlates and could enable further prognostic refinement of patients.
Read full abstract