BIOM-104. AI-BASED QUANTIFICATION OF PD-L1 IMMUNOHISTOCHEMISTRY IN GLIOBLASTOMA: OVERCOMING INTERPRETATION CHALLENGES WITH MACHINE LEARNING

Abstract INTRODUCTION Programmed cell death ligand –1 or PD-L1 is an immune checkpoint protein expressed on inflammatory cells and aberrantly in tumor cells enabling evasion of immune recognition. Accurate quantification is crucial for identifying candidates for anti-PD-L1 immunotherapy. Interpretation of PD-L1 immunohistochemistry (IHC) in Glioblastoma is challenging due to its variable, often incomplete, and faint membranous staining, unlike the distinct staining patterns seen in lung or colon cancer. Current PD-L1 scoring in Glioblastoma using Tumor Proportion Score (TPS) is subjective and prone to inter-observer variability. METHODS We developed an AI-based platform for objective PD-L1 quantification for Glioblastoma from IHC-stained Whole Slide Images (WSI), incorporating 125,347 annotated nuclei. Annotations include faint to partial PD-L1 membranous staining and clearly negative cells such as endothelial cells. The two-stage platform consists of 1) nuclear detection with a BoNuS model trained on 77,789 nuclear annotations and validated on 47,558 annotated nuclei from viable tumor regions of independent cases, and 2) classification of detected nuclei using a pretrained ResNet18 model, with transfer learning and finetuned on PD-L1 stained data for 50 epochs with focal loss criteria to update the model weights. RESULTS The two-stage platform demonstrated strong performance across both detection and classification tasks. The BoNus model achieved an F1 score of 0.7743 with precision of 0.7025 and recall of 0.8679, effectively identifying nuclei in viable tumor regions. The ResNet18 model attained 85.56% accuracy with an F1 score of 0.7362. It exhibited high sensitivity (recall: 0.8306) for detecting PD-L1 positive cells with precision of 0.6919, indicating robust classification despite significant morphological variability. Visual comparison with expert annotations confirmed strong alignment. CONCLUSION This AI-powered platform enables objective and reproducible quantification of PD-L1 IHC in Glioblastoma, addressing the challenges of its staining pattern. Ongoing work includes validation across large multi-site cohorts and refinement to exclude necrotic regions, aiming to support more reliable therapeutic decisions by minimizing inter-observer variability in PD-L1 scoring.