Development And Validation Of Predictive Model For Lymph Node Metastasis In Non-small Cell Lung Cancer Patients

Part One The risk factors for lymph node metastasis in non-small cell lung cancerObjective:Clinical data and CT morphological characteristics of the patients were used to analyze the risk factors of lymph node metastasis(LNM)in NSCLC patients.Methods:A retrospective analysis was performed on 411 NSCLC patients who underwent radical lung cancer resection and lymph node dissection in our hospital from January 2018 to September 2018.All patients underwent standard enhanced CT scans preoperatively.Among the 411 patients,141 had intrapulmonary and mediastinal lymph node metastasis,and 270 had no lymph node metastasis.A radiologist with 5years of diagnostic experience reviewed all the enhanced CT scans without knowing the pathologic diagnosis of lymph nodes,and divided all lymph nodes into groups with or without lymph node metastasis.According to the clinical data and CT morphological characteristics of non-small cell lung cancer patients,univariate and multivariate analysis was conducted to determine the relevant factors of mediastinal and intrapulmonary lymph node metastasis of non-small cell lung cancer.Clinical data included age,gender,smoking status,drinking status,family history and symptom.CT morphological features included the maximum diameter of tumor,vacuole sign,marginal spiculation,marginal lobulation and pleural indentation.The receiver operating characteristic(ROC)curve was further drawn and the area under the curve(AUC)value was calculated,as well as the sensitivity,the specificity,the accuracy,the positive predictive value,and the negative predictive value.Results:In this study,141 patients had intrapulmonary and mediastinal lymph node metastasis.The maximum diameter and spiculation of the tumor have good predictive performance for lymph node metastasis(P<0.01).The OR values of the two risk factors were1.451(maximum tumor diameter)and 2.115(spiculation),respectively.And the optimal truncation values of the Jordan index were spiculation and the maximum tumor diameter≥2.95cm,respectively.Combined with the two factors,the AUC value was 0.684(95%CI:0.631-0.737),with the sensitivity of 0.461,the specificity of 0.819,the accuracy of 0.773,the positive predictive value of 0.664,and the negative predictive value of0.832.Conclusion:There is a high proportion of intrapulmonary and mediastinal lymph node metastasis in NSCLC patients,which requires accurate preoperative lymph node staging.The model based on maximum tumor diameter and spiculation has good clinical prediction efficacy.Part Two Development and validation of Predictive Models for Stage IA Patients with Non-Small Cell Lung CancerObjective:To develop and validate predictive models using clinical data/CT findings,radiomic features and combination of the both for LNM in clinical stage IA NSCLC patients.Methods and materials:This retrospective study included 649clinical stage IA NSCLC patients from September 2017 to January 2019in our hospital.All patients had a thin-section venous CT scan before surgery.There were 138(21%)of the 649 patients who had LNM after surgery.clinical data/CT findings(such as age,gender,smoking status,maximum diameter of tumor,vacuole sign,marginal spiculation,marginal lobulation and pleural indentation)were identified and collected by a study radiologist.The clinical features were selected by Mann-Whitney U test andχ~2 test,and used to develop a clinical model.We used itk-snap software to manually or semi-automatically outline contrast enhanced computed tomography(CE-CT)images(the primary lesions)of NSCLC patients.396 radiomic features were extracted from the region of interest(ROI)by AK software.Among them,there were 455cases in the training group(97 cases with lymph node metastases,358cases without lymph node metastases),and 194 cases in the testing group(41 cases with lymph node metastases,153 cases without lymph node metastases).Mann-Whitney U test and univariate analysis of variance was used for radiomics feature dimension reduction.The least absolute shrinkage and selection operator(LASSO)algorithm was used for radiomics feature selection.Three models(a clinical model,a radiomics model,and a clinical-radiomics model)were developed to predict LNM in the clinical stage IA NSCLC patients.The area under receiver operating characteristic(ROC)curve(AUC value),accuracy,sensitivity,and specificity were used to evaluate the performance in LNM classification by use of the three models.Results:Two clinical features(maximum tumor diameter and spiculation feature)and seven radiomic features(zone-size Variance,Cluster Shade,Correlation,Haralick Correlation,Long Run Emphasis,Percentile15,and Volume,respectively)had good predictive performance(P<0.01).In the clinical model,the AUC values of the training group and the testing group were 0.739 and 0.614,ACC values were 0.633 and0.428,SEN values were 0.753 and 0.951,and SPE values were 0.601 and0.288,respectively.In the radiomic model,the AUC values of the training group and the testing group were 0.898 and 0.851,ACC values were0.782 and 0.804,SEN values were 0.866 and 0.829,and SPE values were0.760 and 0.797,respectively.In the clinical-radiomic model,the AUC values of the training group and the testing group were 0.911 and 0.860,ACC values were 0.802 and 0.830,SEN values were 0.887 and 0.805,and SPE values were 0.779 and 0.837,respectively.Among them,the AUC values of the training group and the testing group of the clinical model and the radiomic model/clinical-radiomic model were statistically significant,and the delong test shows that the P values<0.001.Conclusion:A radiomics model using the venous phase of CE-CT has potential for predicting LNM in clinical CT-based stage IA NSCLC patients,higher than that of the clinical model,and there was no significant statistical difference from the clinical-radiomic model.