IMG-139. AI-Powered Classification of Tissue Sub-Compartments in Glioblastoma: Toward Quantitative Assessment of Tumor Heterogeneity

BACKGROUND Tumor assessment from routinely available clinical whole slide images (WSIs) of Glioblastoma (GBM) significantly impacts prognosis and treatment response but remains challenging due to heterogeneity. Since current manual assessments are subject to inter-rater reliability and reproducibility, we sought to develop an artificial intelligence (AI) based computational pipeline from hematoxylin & eosin-stained WSIs to identify tumor subregions, offering a promising approach for improved treatment strategies. METHOD The dataset comprises 1 million image patches, with 80% allocated to the development cohort and 20% to the validation cohort from 210 multi-institutional GBM WSIs following WHO 2021 reclassification guidelines. This dataset, available through Brain Tumor Segmentation Pathology 2025 challenge was annotated by board-certified neuropathologists for tumor subregions: cellular tumor (CT), geographic necrosis (NC), cortical infiltration, pseudopalisading necrosis, microvascular proliferation, white matter penetration, regions dense with macrophages, leptomeningeal infiltration (LI), and presence of lymphocytes (PL). Our classification framework was based on ResNet18 convolutional neural network, trained using the cross-entropy loss function and optimized with the Adam optimizer with a batch size of 256 over 50 epochs. RESULT The model demonstrated robust performance, achieving a Matthews Correlation Coefficient (MCC) of 0.71, reflecting balanced predictive capabilities across all classes. The model attained a weighted F1-score of 0.85 (Global Accuracy=80.1%, Specificity=95.9%, Sensitivity=67.1%). The class imbalance, with certain tumor subregions such as CT and NC being far more prevalent than LI and PL, influenced the model’s sensitivity across classes and underscored the need for stratified sampling strategies. CONCLUSION Our preliminary results suggest that the AI-driven approach offers a promising solution to accurately classify morphological patterns in GBM, with the ultimate objective of improving diagnostic accuracy and facilitating quantitative studies that advance our understanding of the disease. Additional independent validation is warranted to ensure the generalizability and clinical robustness of the AI framework.