Qualitative analysis of PD-L1 expression in non-small-cell lung cancer based on chest CT radiomics

Abstract

To explore the ability of using chest CT images to predict the immunotherapy response in non-small-cell lung cancer (NSCLC) patients, a radiomics model was constructed to qualitatively analyze PD-L1 expression in NSCLC tumor tissues. We retrospectively collected 54 CT images of primary NSCLC patients with clear PD-L1 expression information, and decomposed each 3D lesion volume data into a set of 2D image sets from 9 fixed orientations to construct a high-quality PD-L1 expression image dataset. Due to the inconsistent size of the lesions, to exclude the potential influence of lesion size on the prediction results, a multi-scale feature extraction method was used to extract 1656 radiomic features for each lesion area at different scales. To automatically select and aggregate the extracted features, an elaborate multi-layer perceptron (MLP) was used to build the PD-L1 expression prediction model. The experimental results showed that our proposed radiomics MLP model can achieve an accuracy of 76.57% and an f1-score of 0.8488 in predicting the positive expression of PD-L1. In addition, we also ranked the importance of the 1656 features, the results showed that the short-run low-gray-enhanced features (glrlmSRLGE) and low-gray run-length enhanced features (glrlmLGRE) based on the grayscale run-length matrix (GLRLM) played a decisive role in predicting PD-L1 expression, and the potential value of glrlmSRLGE and glrlmLGRE features in the qualitative analysis of PD-L1 expression was first revealed.