PATH-67. AI-based WHO 2021 classification of adult-type diffuse glioma solely from H&E-stained slides

INTRODUCTION Conclusive diagnosis of adult-type diffuse glioma (ADG) based on WHO 2021 criteria requires integrative histopathologic assessment with molecular profiling. Molecular profiling is expensive, time-demanding, and when unavailable leads to the ‘Not-Otherwise-Specified’ status. We sought interpretable artificial intelligence (AI)-based classification of ADG into oligodendroglioma, astrocytoma, or glioblastoma, solely from H&E-stained whole slide images (WSI). METHODS We reclassified international retrospective publicly available ADG collections (nWSI=2,418, npatients=1,657) according to the WHO 2021 criteria. Training was based on the multi-national TCGA-GBM/TCGA-LGG (nWSI=1,320, npatients=654) data collections. Independent validation was based on the a) Austrian EBRAINS (nWSI=npatients=794), and b) Indian IPD-Brain (nWSI=304, npatients=209) datasets. Following comprehensive WSI curation, we benchmarked i) eight pathology-specific foundation models (FM) and one AI model pre-trained on ImageNet for patch-based feature extraction, ii) nine multiple-instance learning (MIL) approaches that aggregate features into slide-level representations, and (iii) 15 combinations across magnification levels (2.5x, 5x, 10x, 20x), in our attempt to mimic routine neuropathologic review. RESULTS The best-performing FM-MIL combination in the multi-magnification setting yields accurate (AUCTCGA=0.979 - 10-fold cross-validation) and generalizable performance (AUCEBRAINS=0.963, AUCIPD=0.926). Our key findings indicate: i) FMs outperform the ImageNet model, ii) MIL approaches improve performance only when used with ImageNet-trained models, iii) multi-magnification fusion improves performance in both discovery and replication cohorts. Attention-based interpretability analysis highlights distinct identifiable morphology features. Oligodendroglioma shows perinuclear halos, microcystic and gemistocytic cells. Astrocytoma reveals astrocytes with irregular, elongated shaped nuclei, and fibrillary cytoplasmic processes. Glioblastoma display microvascular proliferation, and pseudo palisading necrosis. CONCLUSIONS Our findings yield a promising AI-based framework for the identification of ADG subtypes directly from H&E-stained WSI alone that can facilitate expedited conclusive diagnosis even when molecular profiling is not available. Interpretability analysis, following expert neuropathological assessment, reveals distilled human-identifiable morphological features that can contribute to further our understanding of disease.