Predictive value of 18F-FDG PET/CT radiomics for the PD-L1 expression level in lung adenocarcinoma patients / 中华核医学与分子影像杂志

ABSTRACT

Objective:To explore the predictive value of 18F-fluorodeoxyglucose (FDG) PET/CT radiomics for the programmed death ligand-1 (PD-L1) expression level in lung adenocarcinoma patients. Methods:A total of 101 patients (43 males, 58 females; median age 60 years) with histologically confirmed lung adenocarcinoma who received pre-treatment 18F-FDG PET/CT from January 2017 to January 2019 in Peking University Cancer Hospital were included retrospectively. There were 44 patients with positive PD-L1 by immunohistochemical assays, and 57 with PD-L1 negative. Patients were assigned to a training set ( n=71) and a validation set ( n=30). Clinical data, PET/CT radiomics parameters, conventional metabolic parameters, and observed CT characteristics of these patients were included in the models. The filter method and embedded method were used in feature selection. Models based on logistic regression, random forest, XGBoost and Light Gradient Boosting Machine (LightGBM) were trained and evaluated, and the optimal parameters to predict the PD-L1 expression as well as the area under curve (AUC) were attained. Results:All models had predictive ability in the prediction of PD-L1 expression, while LightGBM was more powerful than the others, with the precision for positive and negative predictions of 0.85 and 0.76, respectively. Incorporating clinical data and data derived from thin-section CT images (clinical data+ CT) into the LightGBM, the precision, recall and F1-score for positive and negative patients were 0.71, 0.67, 0.69 and 0.69, 0.73, 0.72, respectively, with the accuracy of 0.70 and the AUC of 0.79. As for clinical data+ PET, the precision, recall and F1-score for positive and negative patients were 0.79, 0.73, 0.76 and 0.75, 0.80, 0.77, respectively, with the accuracy of 0.77 and the AUC of 0.80. As for clinical data+ CT+ PET, the precision, recall and F1-score for positive and negative patients were 0.85, 0.73, 0.79 and 0.76, 0.87, 0.81, respectively, with the accuracy of 0.80 and the AUC of 0.83. Features with significant importance in the model (clinical data+ CT+ PET) were as follows maximum standardized uptake value (SUV max), peak of standardized uptake value (SUV peak), CT_shape_Maximum2DDiameterSlice, PET_shape_Elongation, PET_gray level co-occurrence matrix (GLCM)_Correlation, etc. Conclusions:Incorporating clinical data, PET/CT radiomics features and conventional metabolic parameters, the PD-L1 expression can be effectively predicted, which help to assist the selection of patients who may benefit from the immunotherapy.