Tumor proportion score estimation based on multi-stage recognition for lung squamous carcinoma PD-L1 expression

Expression of programmed death ligand-1 is a meaningful biomarker for identifying treatments to tumor patients involved non-small cell lung cancer (NSCLC). Tumor proportion score (TPS) is an essential indexes to describe the level of PD-L1 expression. Obviously, the manual estimation method is laborious and error-prone. Therefore, we propose a novel tumor proportion score estimation method based on multi-stage recognition for lung squamous carcinoma PD-L1 expression (TPS-MSR). Specifically, TPS-MSR is consists of a segmentation path (SP) and a classification path (CP). Then combine their outputs to counting TC for indirect TPS estimation. Experiments on a large image set show that the TPS predicted by TPS-MSR with mean absolute error of 5.07. This indicates the feasibility and effectiveness of our proposed TPS-MSR for TPS prediction.