关键词: 放射组学, 结核,淋巴结, 癌,非小细胞肺, 转移瘤, 诊断,鉴别, 体层摄影术,X线计算机

Abstract:

Objective: To explore the value of CT radiomics in differentiating the mediastinal lymph node tuberculosis and mediastinal lymph node metastasis of non-small cell lung cancer. Methods: From September 2017 to November 2021, CT imaging data of 109 patients with mediastinal lymph node tuberculosis (tuberculosis group) and 65 patients with mediastinal lymph node metastasis of non-small cell lung cancer (metastasis group), who were diagnosed in Shandong Public Health Clinical Center and Qilu Hospital of Shandong University, were retrospectively collected. The CT images were observed and delineated with a double-blind method, and the radiomics features were extracted from the volume of interest (VOI) of the delineated lymph node by using the Radcloud platform. The feature normalization method, univariate and multivariate logistic regression models were used to analyze the characteristics with differential diagnosis capacity and the influence of collinearity between characteristics. Using the selected radiomics characteristics, the patient data of Shandong Public Health Clinical Center was used as a training set to establish a 5-fold cross-validation model of six machine learning methods (including k-Nearest Neighbor (KNN)), Support Vector Machine (SVM), eXtreme gradient boosting (XGBoost), Random Forest (RF), Logistic Regression (LR), and Decision Trees (DT)), and the model with the best diagnostic effect was selected. Then the patient data of Qilu hospital of Shandong University was used as a validation set to verify the diagnostic effect of the model. Results: Two hundred and eighty-one VOIs on CT images were delineated in 174 patients including 196 VOIs in tuberculosis group and 85 ones in metastasis groups. The median (quartile) VOI (1 (1, 8)) of the tuberculosis group was significantly higher than that of the metastasis group (1 (1, 3))(Z=2.827, P=0.005). A total of 1409 radiomics features were extracted, and eight mutually independent radiomics features were selected for modeling after feature standardization, univariate and multivariate logistic regression analysis. After 5-fold cross-validation modeling diagnosis using training group data, the area under the ROC curves of SVM and LR models (0.834 and 0.821, respectively) were superior to the other four models. Furthermore, LR and SVM models were established using training set data with AU values of 0.809 and 0.911, respectively. Then, the external validation was also achieved using the validation set data with AUC values of 0.804 and 0.851, respectively. The results showed that the two models still had a diagnostic effect in extrapolation. Conclusion: Regardless of whether age or gender feature are included or not, the LR model and SVM model established by CT radiomics have good and stable effect in differential diagnosis of mediastinal lymph node tuberculosis and non-small cell lung cancer mediastinal lymph node metastasis, and the SVM model is superior to the LR model.

Key words: Radiomics, Lymph node tuberculosis, Carcinoma, non-small-cell lung, Metastatic tumor, Diagnosis, differential, Tomography, X-ray computed