Preoperative prediction of visceral pleural invasion in non-small cell lung cancer based on contrast-enhanced CT radiomics

LIANG Junjun; CHEN Xiaobo; HU Hengxiao; WANG Kuanhong; CHEN Xin

doi:10.19300/j.2024.L21302

Abstract：Objective To explore the value of a combined predictive model based on CT radiomics quantitative features of non-small cell lung cancer(NSCLC),along with conventional imaging subjective features and clinical information,in preoperatively predicting visceral pleural invasion(VPI)in NSCLC.Methods A retrospective study included 385 NSCLC patients with clear pathological results of VPI status,confirmed by surgical pathology from two hospitals.Their preoperative chest enhanced CT images and clinical data were analyzed.A total of 311 patients from one hospital were used as the training set(144 with VPI and 167 without VPI),and 74 patients from the other hospital were used as the validation set(24 with VPI and 50 without VPI).Subjective CT imaging features were assessed.A semi-automatic segmentation method was used to obtain the tumor volume of interest and extract radiomics features,the optimal radiomics features were selected,and a radiomics model(radiomics signature)was constructed based on logistic regression.Univariate and multivariate logistic regression methods were used to identify clinical and conventional imaging features associated with VPI,and a clinical-conventional imaging predictive model was established.The selected clinical,conventional imaging features,and radiomics signature were combined to construct a joint model using logistic regression.The area under the receiver operating characteristic(ROC)curve(AUC)was used to evaluate the efficacy of the three models in predicting VPI.Results A total of 944 radiomics features were extracted,with 14 radiomics features and 2 clinical-conventional imaging features ultimately selected as related to VPI.In the training and validation sets,the AUCs of the radiomics model were 0.785 and 0.717,respectively;the AUCs of the clinical-conventional imaging predictive model were 0.721 and 0.690,respectively;and the AUCs of the joint model were 0.829 and 0.748,respectively.The joint model had the highest AUC value,with an optimal cutoff value of 0.514 based on the maximum Youden index.The joint model performed well in the training set and validation set(accuracy:0.756 vs.0.743;sensitivity:0.771 vs.0.750;specificity:0.743 vs.0.740).Conclusions The combined model constructed from preoperative CT radiomics features,conventional imaging subjective features,and clinical information helps predict VPI in NSCLC patients,aiding clinical decision-making and improving patient prognosis.