Automated interpretation of PD-L1 CPS based on multi-AI models integration strategy in gastric cancer

Ting Han¹Meng Zhuo¹Ziyu Song²Peilin Chen³Shiting Chen³Wei Zhang⁴Yuanyuan Zhou⁵Hong Li⁶Dadong Zhang⁴Lin Xiaolin¹Zebing Liu²Xiuying Xiao^1*

• ¹Department of Oncology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China
• ²Department of Pathology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China
• ³SODA Data Technology Inc., Shanghai, China
• ⁴Department of Clinical and Translational Research, 3D Medicines Inc., Shanghai, Shanghai, China
• ⁵School of Pharmacy, East China University of Science and Technology, Shanghai, Shanghai, China
• ⁶State Key Laboratory of Systems Medicine for Cancer, Shanghai Cancer Institute, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China

Programmed cell death ligand-1 (PD-L1) combined positive score (CPS) evaluation plays a pivotal role in predicting immunotherapy efficacy for gastric cancer. However, manual CPS assessment suffers from significant inter-observer variability among pathologists, leading to clinical inconsistencies. To address this limitation, we developed a deep learning-based artificial intelligence (AI) system that automates PD-L1 CPS quantification for patients with gastric cancer (GC) using whole slide images (WSIs). Our pipeline firstly employs a dual-network architecture for tumor region detection:MobileNet for patch-level classification and U-Net for pixel-level segmentation. Followed by a YOLO-based cell detection model to compute PD-L1 expression on different cells for CPS calculation.Total 308 GC WSIs (internal cohort: 210; external cohort: 98) were utilized. Within the internal cohort, 100 WSIs were utilized for model development, while the remaining 110 WSIs served as an internal testing set for comparative analysis between AI-derived CPS values and pathologist-derived reference standards. The AI-based CPS prediction pipeline was further evaluated for its performance in the external cohort. Notably, the AI-derived CPS demonstrated strong concordance with expert pathologists' consensus in internal cohort (Cohen's kappa = 0.782) and robust performance on the external test set (Cohen's kappa = 0.737). This system provides a standardized decision-support tool for immunotherapy stratification in GC management, demonstrating potential to improve CPS assessment reproducibility.