Abstract 6247: Artificial intelligence predicts 2021 WHO glioma subtypes from whole slide images

Abstract Diagnosis of diffuse glioma according to the WHO 2021 classification criteria mandate the integration of histologic features with molecular profiling. However, molecular profiling is expensive, time-demanding, and when not available leads to the ‘not-otherwise-specified' status. We seek interpretable AI-based classification of glioma, as oligodendroglioma, astrocytoma, or glioblastoma, from H&E-stained slides alone. We identified 2, 114 multi-institutional whole slide images (WSIs), from two independent retrospective glioma collections, following reclassification according to the WHO 2021 criteria: a) TCGA-GBM/TCGA-LGG (nWSI=1, 320, npatients=654) & b) EBRAINS (nWSI=npatients=794). TCGA data are used for model development, whereas EBRAINS as hold-out data. Each WSI undergoes comprehensive curation to account for any artifacts, such as tissue foldings, glass reflections, and pen markings. We then conduct a quantitative performance evaluation across: i) eight pathology-specific AI foundation models (FM) and an ImageNet-trained AI model that facilitate robust feature extraction, and ii) nine multiple-instance learning (MIL) approaches that are used to aggregate features into slide-level representations, to differentiate across the three glioma classes. Finally, we take into account magnification level combinations of the WSIs (2.5x, 5x, 10x, 20x), in an attempt to mimic the approach that expert neuropathologists follow to histologically assess tissue slides. Our approach yields AUCTCGA=0.979 over a 10-fold cross-validation schema, and generalizable performance on the independent validation (AUCEBRAINS=0.963), for the best performing FM and MIL in the multi-magnification setting. Our key findings indicate: i) domain-specific FMs outperform the ImageNet model, ii) MILs yield larger performance contributions when used with ImageNet models than with FMs, iii) Fusion of multiple magnifications adds value on both development and validation datasets. Interpretability analysis through attention heatmaps highlights distinct identifiable morphology features for each glioma class. Oligodendroglioma show uniform, round nuclei with perinuclear halos, microcystic and gemistocytic cells. Astrocytomas contain cells resembling astrocytes with irregular, elongated shaped nuclei and fibrillary cytoplasmic processes. Glioblastomas display highly heterogeneous cellular composition, pleomorphic astrocyte cells, microvascular proliferation, multinucleated giant cells, and pseudo palisading necrotic regions. Determination of glioma classes directly from routine clinically acquired H&E slides can obviate the need for molecular profiling, expedite conclusive diagnosis and hence clinical decision-making, even in underserved regions. Interpretability analysis towards distilled human-identifiable features can contribute in disease understanding. Citation Format: Shubham Innani, W. Robert Bell, MacLean Nasrallah, Bhakti Baheti, Spyridon Bakas. Artificial intelligence predicts 2021 WHO glioma subtypes from whole slide images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6247.