| Part Ⅰ Prediction of histological subtypes in non-small cell lung cancer by[18F]FDG PET/CT radiomicsObjective:To explore the predictive value of[18F]FDG PET/CT radiomics model for histological subtypes of non-small cell lung cancer(NSCLC)and to construct combined prediction model in association with clinical and radiological features to guide clinical decisions.Methods:The data of 616 patients,who had been pathologically confirmed to have adenocarcinoma(ADC)and squamous carcinoma(SCC)before treatment within one month and had undergone[18F]FDG PET/CT examination from January 2015 to June 2021 at the First Affiliated Hospital of the University of Science and Technology of China,were retrospectively analyzed,and then divided into the training set and the validation set by the rate of 8:2.After image segmentation and preprocessing,a total of 1,488 radiomics features were extracted from[18F]FDG PET and thin-section CT images using Python.Feature selection was performed using Pearson correlation coefficient,analysis of variance,intraclass correlation coefficient(ICC),and least absolute shrinkage and selection operator(Lasso).Four classifiers,including logistic regression(LR),random forest(RF),support vector machine(SVM)and extreme gradient boosting(XGBoost),were employed to establish a radiomics model.Clinical features and radiological features were used to establish a clinical model and a radiological model,and in association with the optimal radiomics model,a combined model was constructed,and a nomogram,based on radiomics label score(Rad-score)and clinical and radiological features,was then built in the validation set for personalized prediction.The predictive performance of the model was evaluated by the receiver operating characteristic(ROC)curve and the calibration curve,and the clinical application value of the model was evaluated through decision curve analysis(DCA).Results:8 radiomics features were associated with histological subtypes.The XGBoost machine learning model was the optimal radiomics feature model,and the areas(AUC)under the ROC curve in the training set and validation set was 0.881(95%CI,0.847-0.914)and 0.851(95%CI,0.771-0.931),higher than the LR,RF and SVM models.Based on the weight calculation of each variable in the XGBoost model,Rad-score was calculated,the value of which was lower in the ADC group than in the SCC group,and the difference was statistically significant.The clinical model based on clinical features(gender,CEA level,and tumor location)had AUCs of 0.848(95%CI,0.813-0.882)and 0.857(95%CI,0.791-0.923)in the training set and validation set,respectively.The radiology model constructed on the radiology features(GTV,SUVmax,and pleural indentation)had AUCs of 0.849(95%CI,0.808-0.890)and 0.838(95%CI,0.760-0.917)in the two sets,respectively.The AUCs of the combined model in both sets are 0.934(95%CI,0.911-0.958)and 0.935(95%CI,0.888-0.982),respectively.The DeLong test showed that the prediction performance of the combined model was higher than that of the XGBoost model(P=0.008 and P=0.038),the clinical model(P<0.001 and P<0.001)and the radiology model(P<0.001 and P=0.001)in both sets.The nomogram demonstrated good predictive performance,and Hosmer-Lemeshow test indicated that the combined model did not deviate from the fit in the training set(χ2=6.736,P=0.241)and in the validation set(χ2=3.337,P=0.648),and DCA showed that the combined model had the optimal clinical utility,with a larger area under the decision curve than in the other models.Conclusion:The XGBoost classifier was considered as the best machine learning method,and the combined model,based on[18F]FDG PET/CT radiomics,clinical and radiological features,can be used as a non-invasive means to predict a NSCLC histological subtype and possess potential clinical applications.Part Ⅱ Prediction of EGFR mutations in non-small cell lung cancer by[18F]FDG PET/CTObjective:To explore the predictive value of[18F]FDG PET/CT radiomics model for epidermal growth factor receptor(EGFR)mutation status in non-small cell lung cancer(NSCLC),and to construct a combined prediction model in association with clinical and radiological features,in order to guide targeted therapy.Methods:The data of 313 patients who had received EGFR test and[18F]FDG PET/CT examination before treatment within one month from January 2015 to June 2021 at the First Affiliated Hospital of the University of Science and Technology of China were retrospectively analyzed,which were then divided into the training set(250 cases)and the validation set(63 sets)by the rate of 8:2.After image segmentation and preprocessing,a total of 1,488 radiomics features were extracted from[18F]FDG PET and thin-section CT images using Python.Feature selection was performed using Pearson correlation coefficient,analysis of variance,intraclass correlation coefficient(ICC),and least absolute shrinkage and selection operator(Lasso).Four classifiers,including logistic regression(LR),random forest(RF),support vector machine(SVM)and extreme gradient boosting(XGBoost),were employed to establish a radiomics model.Clinical features and radiological features were used to establish a clinical model and a radiological model,and in association with the optimal radiomics model,a combined model was constructed,and a nomogram,based on radiomics label score(Rad-score)and clinical and radiological features,was then built in the validation set for personalized prediction.The predictive performance of the model was evaluated by the receiver operating characteristic(ROC)curve and the calibration curve,and the clinical application value of the model was assessed by decision curve analysis(DCA).Results:Among the four classifiers constructed with the 6 radiomics features after feature selection,the areas under the curve(AUC)of the RF model were the highest in the training set and the validation set,which were 0.785(95%CI,0.726-0.844)and 0.776(95%CI,0.662-0.889)respectively.The AUCs of the clinical model constructed on clinical features(smoking history,gender,and pathological type)in the two sets were 0.711(95%CI,0.645-0.776)and 0.758(95%CI,0.627-0.890),respectively.The radiological model constructed on the radiological features(air bronchogram,pleural indentation and spiculation sign)had AUCs of 0.632(95%CI,0.564-0.699)and 0.677(95%CI,0.531-0.822)in both sets,respectively.The AUCs of the joint model in the sets were 0.872(95%CI,0.829-0.915)and 0.831(95%CI,0.723-0.940),respectively.The DeLong test showed that the predictive performance of the combined model in the training set and the validation set was higher than that of the RF model,the clinical model and the radiology model,with statistically significant differences(P=0.010,P<0.001 and P<0.001),and was higher than that of the radiological model in the training set(P=0.017).The nomogram had good predictive performance,and Hosmer-Lemeshow test indicated that the model did not deviate from the fit in the training set(χ2=7.398,P=0.495)and in the validation set:χ2=9.900,P=0.272).In addition,DCA showed the area under the decision curve of the joint model larger than other models,thus possessing the best clinical utility.Conclusion:The RF classifier was considered as the best machine learning method,and[18F]FDG PET/CT radiomics features,in combination with clinical and radiological features,are helpful to predict EGFR mutation status in NSCLC patients,and thus have guiding significance for EGFR-tyrosine kinase inhibitors(EGFR-TKIs)targeted therapy.Part Ⅲ Prediction of one-year progression-free survival in patients with advanced non-small cell lung cancer by[18F]FDG PET/CT radiomicsObjective:To explore the predictive value of the[18F]FDG PET/CT radiomics model for one-year progression-free survival(PFS)in patients with advanced non-small cell lung cancer(NSCLC),and to construct a combined prediction model in association with clinical features in order to assist clinical decisions.Methods:The data of 313 patients with advanced NSCLC who underwent[18F]FDG PET/CT examination before treatment within one month during January 2015 to December 2020 at the First Affiliated Hospital of the University of Science and Technology of China were retrospectively analyzed,and then divided into a training set(250 cases)and a validation set(63 cases)by the ratio of 8:2.After image segmentation and preprocessing,a total of 1,488 radiomics features were extracted from[18F]FDG PET and thin-section CT images using Python.Feature selection was performed using Pearson correlation coefficient,analysis of variance,intraclass correlation coefficient(ICC),and least absolute shrinkage and selection operator(Lasso).The four classifiers of Logistic Regression(LR),Random Forest(RF),Support Vector Machine(SVM)and extreme gradient boosting(XGBoost)were employed to establish the radiomics models respectively,with the optimal model screened out.In addition,the nomogram based on Rad-score and clinical features was constructed in the validation set to achieve personalized prediction.The predictive performance of the model was evaluated by the receiver operating characteristic(ROC)curve and the calibration curve,and the clinical utility of the model was assessed by decision curve analysis(DCA).Results:The 9 radiomics features obtained after feature selection could be used to construct radiomics models.The AUCs of the four model classifiers of LR,RF,SVM and XGBoost in the training set and the validation set were 0.695,0.774,0.687,0.757 and 0.641,0.696,0.688,and 0.723,respectively,with the XGBoost model with the highest AUC in the validation set taken as the optimal radiomics model.Based on the same selecting method,we obtained a clinical model constructed with 6 clinical features(tumor location,clinical stage,CYFA21-1,targeted therapy,surgery,and MTV),the AUCs in the training set and the validation set being 0.748(95%CI,0.688-0.809)and 0.728(95%CI,0.603-0.854),respectively.The combined model based on Rad-score and clinical features could improve the predictive performance,and its AUCs in the training set and validation set were 0.839(95%CI,0.791-0.887)and 0.775(95%CI,0.657-0.893),respectively.The nomogram had good predictive performance.The Hosmer-Lemeshow test indicated that the model did not deviate from the fit in the training set(χ=11.327,P=0.184)and in the validation set(x2=5.271,P=0.728),and DCA analysis showed that the combined model possessed certain clinical value.Conclusion:The XGBoost classifier was considered as the best machine learning method,and[18F]FDG PET/CT radiomics features combined with clinical features can help to predict the 1-year progression-free survival of patients with advanced NSCLC,and serve as a guide for patient management and effective treatment strategies. |
PDF Full Text Request