| Part Ⅰ The value of CT radiomics in the prediction of PD-L1 expression in non-small cell lung cancerObjective To explore the application value of CT radiomic features of tumors in predicting programmed death ligand-1(PD-L1)expression in non-small cell lung cancer(NSCLC).Methods The clinicopathological and CT data of 318 NSCLC patients who underwent lung CT scanning and PD-L1 immunohistochemical examination in affiliated hospital of Jiangnan university from June 2018 to January 2021 were analyzed retrospectively.According to the ratio of 2:1,they were randomly divided into training set(212 cases)and validation set(106 cases).With the lung tumor biopsy or postoperative pathological results as the gold standard,the expression rate of PD-L1 in tumor cells greater than or equal to 1%was positive,while the expression rate of PD-L1 less than 1%was negative.There were 110 positive cases and 102 negative cases of PD-L1 expression in training set.There were 54 positive cases and 52 negative cases of PD-L1 expression in validation set.Using ITK-SNAP software,two doctors with 5 years and 10 years of experience in lung imaging diagnosis independently delineated the 2D region of interest(ROI)along the edge of tumor with the largest cross section on the lung-window CT images,and the doctor with 10 years of experience repeated the delineation step after a week interval.Interclass and intraclass correlation coefficients(ICCs)were used to test the consistency of the features extracted by the two doctors.A total of 464 radiomic features in the ROI of lung tumor were extracted by Python-based pyradiomics software package,including four categories:①morphological features;②features of first-order gray histogram;③second-order and high-order texture features;④features based on wavelet transform.The extracted radiomic features were screened by LASSO regression,and the optimal features subset were selected by 10-fold cross-validation,and the radscore of each patient was obtained by logistic regression.Furthermore,combined with clinicopathological features,a predictive model of PD-L1 expression in lung cancer was constructed by multivariate logistic regression analysis,which was displayed as a nomogram with visualization and interpretability.The area under the curve(AUC),sensitivity,specificity,and accuracy of the prediction model were calculated by using the receiver operating characteristic(ROC)curve,in contrast with artificial neural network(ANN)and support vector machine(SVM)methods in predicting the PD-L1 expression of lung cancer.Finally,the calibration curve and clinical decision curve were obtained by comparing the prediction results of nomogram model with the actual clinical observation results.Results The consistency for the radiomic features of tumors extracted by these two doctors was good,and intraclass and interclass correlation coefficients(ICCs)were 0.86 and 0.88,respectively.The features firstly extracted by the doctor with 10 years of experience in diagnosis were taken as the record result.In the training set of 212 patients,11 optimal radiomic features were finally selected,including 2 original shape features,5 wavelet firstorder features and 4 wavelet texture features.The AUC for the constructed radscore in predicting PD-L1 expression of lung cancer was 0.737,and the sensitivity,specificity,and accuracy rate were 74.3%,62.7%,and 68.9%respectively.In terms of clinicopathological factors,the clinical stage,pathological type and the degree of differentiation of cancer cells were related to the expression of PD-L1 in tumors.The prediction model was established by logistic regression based on these three factors,and the AUC for the prediction model was 0.690.For the constructed nomogram prediction model based on radscore and clinicopathological factors,the AUC,sensitivity,specificity,and accuracy rate were 0.812,77.1%,72.5%,and 75.0%,respectively.In the training set,the sensitivity,specificity,and accuracy for ANN machine learning method in predicting the PD-L1 expression of lung cancer were 75.0%,71.1%,and 73.1%,respectively.The sensitivity,specificity,and accuracy for SVM machine learning method were 75.2%,60.8%,and 68.4%,respectively.In 106 patients in the validation set,the AUC for radscore in predicting the PD-L1 expression of lung cancer was 0.668,and the sensitivity,specificity,and accuracy rate were 70.9%,57.7%,and 64.2%respectively.The AUC for clinical pathological prediction model was 0.750.In the nomogram model,the AUC,sensitivity,specificity,and accuracy rate were 0.779,74.5%,65.4%,and 69.8%,respectively.The calibration curve showed that the nomogram results in predicting PD-L1 expression in lung cancer were in good agreement with the actual clinical observation results,the C index was 0.824.Decision curve analysis(DCA)showed that patients could get higher net benefits within the threshold probability range from 0.1 to 0.8 by the nomogram model in predicting the expression of PD-L1 in lung cancer.Conclusion CT radiomic quantitative features of tumors may predict PD-L1 expression in NSCLC,and the nomogram model combined with clinicopathological factors could improve its predictive efficiency and had high clinical application value.In contrast,ANN and SVM machine learning methods did not show better ability to predict the expression of PD-L1 in lung cancer.Part II The value of CT radiomics in the prediction of immunotherapy efficacy in non-small cell lung cancerObjective To investigate the application value of CT radiomic features of tumors in predicting the immunotherapy efficacy in non-small cell lung cancer(NSCLC).Methods The clinicopathological and lung CT data of 70 NSCLC patients who received immunotherapy in the first affiliated hospital of Soochow university from June 2018 to June 2021 were retrospectively analyzed as a training set in CT radiomics;Meanwhile the clinicopathological and lung CT data of 37 NSCLC patients who received immunotherapy in affiliated hospital of Jiangnan university from the same period were collected as the validation set in CT radiomics.According to iRECIST,in the training set 34 cases were effective and 36 cases were ineffective for immunotherapy.In the validation set 18 cases were effective and 19 cases were ineffective for immunotherapy.With ITK-SNAP software,using the same method as Part I,464 quantitative radiomic features in pretreatment tumor ROI on the lung window CT images were extracted,and the optimal feature subset and radscore were obtained by screening and dimension reduction.Combined with clinicopathological factors,the model in predicting the efficacy of immunotherapy for lung cancer was constructed by multivariate logistic regression analysis,which was displayed as a nomogram.The area under the curve(AUC),sensitivity,specificity,and accuracy of the prediction model were calculated by using the receiver operating characteristic(ROC)curve.Furthermore,they were compared with support vector machine(SVM)and decision tree(DT)machine learning methods in predicting the immunotherapy effect of lung cancer.The calibration curve and clinical decision curve were obtained by comparing the predicted results of nomogram model with the actual clinical observation results.Results The consistency for the radiomic features of tumors extracted by these two doctors was good,and intraclass and interclass correlation coefficients(ICCs)were 0.85 and 0.89,respectively.In the training set of 70 patients,5 optimal radiomic features were selected,including 1 original shape feature,1 wavelet first-order feature and 3 wavelet texture features.The AUC for the constructed radscore in predicting the efficacy of immunotherapy in lung cancer was 0.736,and the sensitivity,specificity,and accuracy rate were 66.7%,73.5%,and 70.0%respectively.In terms of clinicopathological factors,the expression of PD-L1 in tumor and pathological type were related to the immunotherapy effect of lung cancer.The AUC for the clinicopathological prediction model was 0.732.For the nomogram model combined the radscore and clinicopathological factors,the AUC,sensitivity,specificity,and accuracy were 0.863,72.2%,82.4%,and 78.6%,respectively.In the training set,the sensitivity,specificity,and accuracy rate for SVM were 70.7%,76.5%,and 72.9%,respectively.As well,the sensitivity,specificity,and accuracy for DT were 71.9%,79.4%,and 75.7%,respectively.In the validation set of 37 patients,the AUC for radscore in predicting the immunotherapy effect of lung cancer was 0.719,and the sensitivity,specificity,and accuracy rate were 57.9%,77.8%,and 67.6%respectively.The AUC for clinicopathological prediction model was 0.651.The AUC for the nomogram model was 0.816,and the sensitivity,specificity,and accuracy rate were 73.7%,66.7%,and 70.3%,respectively.The calibration curve showed that the results of nomogram for predicting the immunotherapy effect of lung cancer were in good agreement with the actual clinical observation results,C index value was 0.897.Decision curve analysis(DCA)showed that patients could get higher net benefits within the threshold probability range from 0.1 to 1.0 by the nomogram model in predicting the immunotherapy effect of NSCLC.Conclusion CT radiomic quantitative features in pretreatment tumors may predict the immunotherapy efficacy in NSCLC,and the nomogram model that combined radiomic features and clinicopathological factors could improve its predictive ability,and had high clinical application value.In contrast,SVM and DT machine learning methods did not show better ability to predict the immunotherapy efficacy in lung cancer.Part III The value of CT radiomics in the prediction of immunotherapyinduced pneumonitis in non-small cell lung cancerObjective To explore the application value of CT radiomic features of baseline lung tissue in predicting immunotherapy-induced pneumonitis(IIP)in non-small cell lung cancer(NSCLC).Methods The clinicopathological and lung CT data of 350 NSCLC patients who received immunotherapy in the first affiliated hospital of Soochow university or affiliated hospital of Jiangnan university from June 2018 to September 2021 were analyzed retrospectively.They were randomly divided into training set(233 cases)and validation set(117 cases)according to the ratio of 2:1.According to the diagnostic criteria of IIP,28 patients with IIP and 205 patients without IIP were included in the training set.There were 22 patients with IIP and 95 patients without IIP in the validation set.With ITK-SNAP software,using the same method as Part Ⅰ,these two doctors independently delineated and extracted 455 quantitative radiomic features by drawing 2D ROI along the edge of pretreatment baseline lung tissue on the lung window CT images,and the optimal feature subset and radscore were obtained by screening and dimension reduction.Combined with clinicopathological factors,the prediction model was constructed by using multivariate logistic regression analysis,which was displayed as a nomogram.Then,the area under the curve(AUC),sensitivity,specificity,and accuracy for the prediction model were calculated by using the receiver operating characteristic(ROC)curve.Furthermore,they were compared with the machine learning methods of random forest(RF)and support vector machine(SVM)in predicting IIP in NSCLC.Finally,the calibration curve and clinical decision curve were obtained by comparing the predicted results of nomogram model with the actual clinical observation results.Results The consistency for the radiomic features of tumors extracted by these two doctors was good,and intraclass and interclass correlation coefficients(ICCs)were 0.83 and 0.87,respectively.In the training set of 233 patients,11 optimal radiomic features were selected,including 1 original texture feature and 10 wavelet texture features.The AUC for radscore in predicting IIP was 0.763,and the sensitivity,specificity,and accuracy rate were 72.7%,70.5%,and 70.8%respectively.In terms of clinicopathological factors,the smoking history and basic lung diseases were related to IIP.The AUC for the clinical model constructed by logistic regression of these two factors in predicting IIP was 0.715.Combined radscore and clinicopathological factors,the nomogram model was constructed.The AUC for the nomogram model was 0.853,and the sensitivity,specificity,and accuracy were 81.8%,77.9%,and 78.5%,respectively.In the training set,the sensitivity,specificity,and accuracy for RF method in predicting IIP were 77.3%,68.4%,and 69.5%,respectively.Meanwhile,the sensitivity,specificity,and accuracy for SVM method in predicting IIP were 72.7%,76.8%,and 76.0%,respectively.In 117 patients of the validation set,the AUC for radscore in predicting IIP was 0.754,and the sensitivity,specificity,and accuracy rate were 72.7%,66.3%,and 67.5%,respectively.The AUC for clinical model in predicting IIP was 0.717.The AUC for the nomogram model in predicting IIP was 0.810,and the sensitivity,specificity,and accuracy rate were 77.3%,71.6%,and 72.6%,respectively.The calibration curve showed that the results of the nomogram for predicting IIP in lung cancer were in good agreement with the actual clinical observation results,C index value was 0.904.Decision curve analysis(DCA)showed that patients could get higher net benefits within the threshold probability range from 0.3 to 0.9 by the nomogram model in predicting the IIP.Conclusion CT quantitative radiomic features based on baseline lung tissue may predict the possibility of IIP in NSCLC patients.The nomogram model combined with clinicopathological factors could improve the prediction efficiency and had high clinical application value.In contrast,RF and SVM machine learning methods did not show better ability to predict IIP in lung cancer patients. |
PDF Full Text Request