Dual-scale categorization based deep learning to evaluate programmed cell death ligand 1 expression in non-small cell lung cancer

Abstract

In precision oncology, immune check point blockade therapy has quickly emerged as novel strategy by its efficacy, where programmed death ligand 1 (PD-L1) expression is used as a clinically validated predictive biomarker of response for the therapy. Automating pathological image analysis and accelerating pathology evaluation is becoming an unmet need. Artificial Intelligence and deep learning tools in digital pathology have been studied in order to evaluate PD-L1 expression in PD-L1 immunohistochemistry image. We proposed a Dual-scale Categorization (DSC)-based deep learning method that employed 2 VGG16 neural networks, 1 network for 1 scale, to critically evaluate PD-L1 expression. The DSC-based deep learning method was tested in a cohort of 110 patients diagnosed as non-small cell lung cancer. This method showed a concordance of 88% with pathologist, which was higher than concordance of 83% of 1-scale categorization-based method. Our results show that the DSCbased method can empower the deep learning application in digital pathology and facilitate computer-aided diagnosis.