In Frontiers in oncology
Purpose : To establish and validate a radiomics model to estimate the malignancy of mediastinal lymph nodes (LNs) based on contrast-enhanced CT imaging.
Method : In total, 201 pathologically confirmed mediastinal LNs from 129 patients were enrolled and assigned to training and test sets. Radiomics features were extracted from the region of interest (ROI) delineated on venous-phase CT imaging of LN. Feature selection was performed with least absolute shrinkage and selection operator (LASSO) binary logistic regression. Multivariate logistic regression was performed with the backward stepwise elimination. A model was fitted to associate mediastinal LN malignancy with selected features. The performance of the model was assessed and compared to that of five other machine learning algorithms (support vector machine, naive Bayes, random forest, decision tree, K-nearest neighbor) using receiver operating characteristic (ROC) curves. Calibration curves and Hosmer-Lemeshow tests were used to assess the calibration degree. Decision curve analysis (DCA) was used to assess the clinical usefulness of the logistic regression model in both the training and test sets. Stratified analysis was performed for different scanners and slice thicknesses.
Result : Among the six machine learning methods, the logistic regression model with the eight strongest features showed a significant association with mediastinal LN status and the satisfactory diagnostic performance for distinguishing malignant LNs from benign LNs. The accuracy, sensitivity, specificity and area under the ROC curve (AUC) were 0.850/0.803, 0.821/0.806, 0.893/0.800, and 0.922/0.850 in the training/test sets, respectively. The Hosmer-Lemeshow test showed that the P value was > 0.05, indicating good calibration, and the calibration curves showed good agreement between the classifications and actual observations. DCA showed that the model would obtain more benefit when the threshold probability was between 30% and 90% in the test set. Stratified analysis showed that the performance was not affected by different scanners or slice thicknesses. There was no significant difference (DeLong test, P > 0.05) between any two subgroups, which showed the generalization of the radiomics score across different factors.
Conclusion : The model we built could help assist the preoperative estimation of mediastinal LN malignancy based on contrast-enhanced CT imaging, with stability for different scanners and slice thicknesses.
Dong Mengshi, Hou Gang, Li Shu, Li Nan, Zhang Lina, Xu Ke
benign, classification, machine learning, malignant, mediastinal lymph nodes, radiomics