Background: The WHO 2021 classification criteria for adult-type diffuse glioma integrate histology with molecular profiling for conclusive diagnosis. Since molecular profiling can be expensive and time-consuming, often necessitating outsourcing or leading to the “not otherwise specified (NOS) label,” this study develops an AI-driven WHO 2021 classification of gliomas solely from H&E whole-slide images (WSIs). Methods: Our pipeline is based on a multi-institutional dataset reclassified per WHO 2021 guidelines. This dataset includes (a) Primarily US-based TCGA-GBM/TCGA-LGG (n = 1,320) for model training, independently evaluated on two hold-out sets, (b) Austria-based EBRAINS (n = 794), and (c) India-based IPD-Brain (n = 304). Each WSI undergoes pre-processing followed by quantitative benchmarking across (i) eight pathology foundation models (FM), (ii) nine aggregation methods (AM), and (iii) 15 combinations of magnification levels through a late fusion approach. Model interpretability conducted through heatmaps highlights distinct, identifiable morphology features. Results: Our best-performing combination of FM, AM, and multi-magnification achieved an AUC of 97.95% on the training cohort, 96.30% on EBRAINS (set 1), and 92.61% on IPD (set 2). The results yield the following key insights: (1) domain-specific FMs outperform ImageNet-based models, (2) AMs, while theoretically promising yield larger performance improvements when used with ImageNet-based feature extractor rather than FMs, and (3) Fusion of multiple magnifications adds value in performance. Conclusion: Determining glioma diagnosis directly from H&E slides can obviate the need for molecular profiling, expedite conclusive diagnosis, and, hence, clinical decision-making. These findings motivate the development of advanced domain-relevant FMs and the design of more adaptable slide-level aggregation techniques.

Background/Objectives: Adult-type diffuse gliomas (ADGs) are aggressive brain tumors with highly variable clinical outcomes. Prognostic assessment directly from routine hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) could support personalized decision-making. Methods: To our knowledge, this is the first systematic benchmarking of pathology foundation models (FMs) and multiple instance learning (MIL) methods for overall survival (OS) risk stratification in ADGs under the 2021 World Health Organization (WHO) classification. A retrospective, multi-institutional cohort of 817 subjects from The Cancer Genome Atlas (TCGA) was analyzed. Following tissue segmentation and patch-level curation, we evaluated 77 pathology FM-MIL combinations (11 FMs × 7 MIL methods), together with seven ResNet-50 baseline configurations (84 encoder-MIL configurations in total), using patient-level 10-fold cross-validation. A perturbation-based evaluation assessed attention mechanism faithfulness. Results: UNI2 paired with MambaMIL achieved the highest cross-validated concordance index (C-index = 0.77), whereas its time-dependent AUC was modest (0.60–0.61 at 12–36 months), and several combinations performed comparably. Kaplan–Meier analysis demonstrated significant separation between high- and low-risk groups (log-rank p = 9.16 × 10⁻¹⁷; hazard ratio [HR] = 2.90, 95% confidence interval [CI]: 2.23–3.76), with a median OS of 16.6 months (95% CI: 15.0–20.4) versus 63.5 months (95% CI: 43.9-NR), respectively. Tertile stratification confirmed a monotonic OS gradient from 94.5 months to 14.7 months (p = 3.25 × 10⁻¹⁷). Risk stratification within individual WHO 2021 types was not significant. Most domain-specific FMs outperformed the ImageNet-pretrained baseline, and the removal of highly attended patches reduced performance, supporting attention faithfulness. Conclusions: The AI-derived risk score remained associated with OS after adjustment for individual clinical and molecular variables, but independent prognostic value was not demonstrated in the fully adjusted model (HR = 1.32, 95% CI: 0.94–1.85, p = 0.104), and its discrimination largely reflects differences between WHO 2021 types. This benchmarking framework provides guidance for FM-MIL design in computational pathology applications targeting survival prediction; external validation is required before clinical application.