Interpretable artificial intelligence based determination of glioma IDH mutation status directly from histology slides

Abstract

Background: Isocitrate dehydrogenase (IDH) mutation status is a diagnostic requirement for glioma with associated prognostic and therapeutic implications. Clinical routine visual assessment of tissue is insufficient to determine IDH status conclusively, mandating molecular workup that is unavailable everywhere. Methods: We developed an interpretable Artificial Intelligence (AI)-based approach for determining IDH status directly from H&E-stained glioma slides. Our study is based on 2442 multi-institutional whole slide images (WSIs) from 3 independent retrospective glioma collections, following their reclassification according to the WHO 2021 criteria: (1) TCGA-GBM/TCGA-LGG (n WSI = 1534, n patients = 799), (2) University of Pennsylvania Health System collection (UPHS, n WSI = n patients = 114), and (3) EBRAINS (n WSI = n patients = 794). Method development is based on TCGA, whereas UPHS and EBRAINS are independent hold-out datasets for model validation. Six pathology-specific foundation AI models and an ImageNet-pretrained AI model facilitate robust feature extraction. Features are aggregated into slide-level representations via an interpretable multiple-instance learning (MIL) mechanism to differentiate IDH-wildtype from IDH-mutant cases and generate attention heatmaps correlating with identifiable morphologic characteristics. Results: Our approach yields AUCTCGA = 0.96 over a 10-fold cross-validation schema and generalizable performance on independent validation (AUCUPHS = 0.97, AUCEBRAINS = 0.95). Interpretability analysis reveals high attention regions in IDH-wildtype tumors exhibiting significant pleomorphism and microvascular proliferation, while IDH-mutant tumors show dense nodular cell concentrations, microcysts, and gemistocytic cells. Conclusions: Accurate H&E-based determination of glioma IDH mutation status can expedite conclusive diagnosis and clinical decision-making and even facilitate it in underserved regions. Finally, interpretability analysis of distilled human-identifiable features can further improve our understanding of the disease.