Risk prediction model for postoperative brain metastasis in IIB-IIIB non-small cell lung cancer: based on radiomics and clinicopathology

Abstract

Purpose To develop and validate a model based on radiomics and clinicopathological features for predicting postoperative brain metastasis (BM) in stage IIB-IIIB non-small cell lung cancer (NSCLC) patients. Materials and methods A total of 333 NSCLC patients operated from October 2015 and December 2019 with postoperative pathological stage IIB-IIIB were included, which were randomly divided into a training and validation cohort. The intratumoral and peritumoral radiomics features from preoperative CT image were extracted and selected using the least absolute shrinkage and selection operator (LASSO). The independent clinical predictors of BM were identified by univariate and multivariate Cox analysis. The radiomics model, clinical model and radiomics combined clinicopathological model were constructed with six different algorithms. Subsequently, we constructed a dynamic nomogram. The performance of the model was evaluated by the area under the curve (AUC), sensitivity, specificity, calibration curve and decision curve analysis (DCA). Results The radiomics model combining intratumoral and peritumoral radiomics features exhibited great predictive performance for BM prediction, with an AUC of 0.888–0.928 in the training cohort and 0.838–0.894 in the validation cohort. The model including the intra- and peritumoral radiomics, T stage, histological type, spiculation and other metastatic sites yielded AUC of 0.947–0.979 in the training cohort and 0.847–0.926 in the validation cohort, with good calibration for all algorithms (p > 0.05). DCA revealed that the combined model obtained a greater net benefit. Conclusion The model that integrates radiomics features with clinicopathological features could aid in early-stage prediction of postoperative BM risk in stage IIB-IIIB NSCLC patients. Dynamic nomogram provides great convenience for clinicians to manage patients.