Abstract 2441: Artificial intelligence-based quantification of PD-L1 staining intensity in non-small cell lung cancer: Beyond binary assessment

Abstract Background: Determining the non-small cell lung cancer (NSCLC) patients likely to benefit from treatment based on the proportion of PD-L1-positive cells in immunohistochemistry (IHC)-stained slides, such as the Tumor Proportion Score (TPS), is effective but suboptimal. Manual PD-L1 quantification is labor-intensive, subjective, and restricted to binary assessment (positive/negative). To address these challenges, we developed an artificial intelligence (AI)-powered IHC image analysis pipeline that identifies cancer cells and their components while quantifying the staining intensity of the cancer cell membrane where PD-L1 is bound, on a scale from 0 to 100. We validated the AI-generated PD-L1 expression metrics by comparing them with clinically assessed PD-L1 TPS in NSCLC. Methods: The AI pipeline consists of two deep learning components: cancer area localization and cancer cell staining analysis. The first component extracts patch images from whole-slide images (WSIs) and identifies cancer areas using a segmentation model for cancer localization. The second component classifies pixels of the tile image into the cell nucleus, cytoplasm, or membrane using a cellular compartment segmentation model. This component measures the optical density of Diaminobenzidine (DAB) staining at the target component corresponding to the target antibody for each cancer cell (e.g., PD-L1 on the cell membrane). We collected 366 PD-L1-stained NSCLC WSIs, which were clinically assessed for TPS by expert pathologists and used as the independent set for evaluation. Pearson's correlation coefficient (r) was used to compare clinically assessed TPS (C-TPS) values with PD-L1 staining metrics generated from the AI pipeline, such as TPS (AI-TPS), H-score, and the mean PD-L1 intensity of all tumor cells (Mean PD-L1 Intensity). The Mann-Whitney test was employed to compare the distribution of staining intensity values across NSCLC subtypes. Results: The C-TPS values showed a strong correlation with AI-generated PD-L1 staining intensity metrics: Mean PD-L1 Intensity (r = 0.9042, p < 1e-6), H-score (r = 0.8995, p < 1e-6), and AI-TPS (r = 0.8896, p < 1e-6). Additionally, the C-TPS values exhibited a strong correlation with PD-L1 intensity heterogeneity metrics: the standard deviation of PD-L1 intensity (r = 0.7261, p < 1e-6), interquartile range of PD-L1 intensity (r = 0.6125, p < 1e-6). The Mean PD-L1 Intensity in lung adenocarcinoma was significantly higher than that in squamous cell carcinoma (Mann-Whitney test, p < 0.0001). Conclusion: We demonstrated that the AI-driven PD-L1 staining intensity metrics show a high correlation with clinically assessed TPS values. This AI pipeline extends beyond binary PD-L1 assessment and can be applied to novel targets for PD-L1 evaluation. Citation Format: Yunseob Hwang, Gui Young Kwon, Jeongwon Kim, Jiyoon Jung, DoKyung Kim, Tae-Yeong Kwak, Sun Woo Kim, Hyeyoon Chang. Artificial intelligence-based quantification of PD-L1 staining intensity in non-small cell lung cancer: Beyond binary assessment [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 2441.