Automated Tumor Proportion Scoring for Assessment of PD-L1 Expression Based on Multi-Stage Ensemble Strategy.

Tumor Programmed Death-Ligand 1 (PD-L1) expression is a crucial biomarker to identify tumor patients who may have an enhanced response to anti-Programmed Death-1 (PD-1)/PD-L1 treatment. Tumor proportion score (TPS) is an indicator to describe the frequency of PD-L1 expression and is essential for selecting from different tumor therapies. In this paper, we propose a novel deep learning-based framework for automated tumor proportion scoring. Specifically, we introduce the clinical diagnosis process to our framework. The framework consists of a cellular localization network (C-Net) and a regional segmentation network (RNet). The C-Net is dedicated to classifying cells and generating TPS, and R-Net learns to distinguish tumor regions from their normal counterparts. The predictions made by R-Net can, in turn, be used to refine the TPS. We have consolidated the visual TPS from multiple pathologists for clinical verification. Concordance measures computed on a set of WSI provide evidence that our method matches visual scoring from multiple pathologists (MAE = 7.405, RMSE = 11.25, PCCs = 0.9305, SRCC = 0.967).