Prediction Of Pathological Subtypes And Egfr Mutation Status In Non-small Cell Lung Cancer By ¹⁸F-FDG PET/CT Radiomics

Part One Prediction of solid/micropapillary patterns in lung adenocarc-inoma by ¹⁸F-FDG PET/CT radiomicsObjective:To evaluate the value of pre-therapy ¹⁸F-FDG PET/CT radiomics in identifying solid/micropapillary patterns in lung adenocarcinoma,and to provide imaging information for risk stratification of lung cancer.Methods:A total of 233 patients with invasive lung adenocarcinoma diagnosed and treated in our hospital from January 2015 to November 2019were enrolled.All patients underwent pre-therapy ¹⁸F-FDG PET/CT examination,with clear pathological subtype diagnosis.All patients according to whether containing solid/micropapillary patterns were divided into two groups:solid/micropapillary patterns negative group（groupⅠ）,a total of 134patients;solid/micropapillary patterns positive group（groupⅡ）,a total of 99patients.Patients were randomly divided into two sets（163 cases in the training set and 70 cases in the validation set）at a ratio of 7:3.LIFEx package（http://www.lifexsoft.org）was used to extract the PET/CT radiomic features of the lesions in the same VOI,and a total of 47 PET-based and 45 CT-based features were extracted.The least absolute shrinkage and selection operator（LASSO）algorithm was used for feature screening in the training set to establish the radiomics signature,and the radiomics signature score（rad-score）was calculated for each patient.Logistic regression models（radiomics model,clinical model and complex model）were established by radiomics signature,clinical variables,and their combinations.Receiver operating characteristic curve（ROC）was used to evaluate the predictive performance of the three models.Nomogram based on rad-score and clinical variables were established in the training set.Decision Curve Analysis（DCA）was used to evaluate the clinical value of the three models.Results:A total of 9 radiomics features（5 PET radiomic features,4 CT radiomic features）were selected by Lasso regression to establish the radiomics signature,and the rad-score for each patient was calculated.In both the training and validation sets,the rad-score of the solid/micropapillary patterns positive group was higher than that of the negative group（both P<0.0001）.In predicting the solid/micropapillary patterns,the radiomics model showed good predictive performance,with AUC values of 0.86（95%CI,0.80–0.91）in the training set and 0.85（95%CI,0.77–0.94）in the validation set.Among clinical variables,gender（OR=2.976,95%CI:1.636–5.415,P=0.0004）,CEA level（OR=4.597,95%CI:2.431–8.693,P<0.0001）and lymph node metastasis（OR=2.084,95%CI:1.094–3.971,P=0.0256）were independent risk factors for predicting solid/micropapillary patterns.The AUC values of the clinical model established by them in the training and validation sets were 0.76（95%CI,0.69–0.83）and 0.74（95%CI,0.63–0.86）,respectively.The predictive power of the complex model was significantly improved when the radiomics signature was combined with clinical variables,with the AUC values increased to 0.89（95%CI,0.84–0.94）in the training set and 0.86（95%CI,0.78–0.95）in the validation set,which were significantly higher than that of the clinical model（P<0.0001 and P=0.001,in the training and validation set,respectively）.When the complex model was compared with the radiomics model,there was no significant difference in the training set（De Long test,P=0.103）and the validation set（De Long test,P=0.859）.Nomogram based on rad-score and clinical variables（gender,CEA level,lymph node metastasis）in the training set had good predictive performance.Hosmer-Lemeshow test was performed in the training and validation set,and there was no statistically significant difference between the predicted and the observed values.Decision curve analysis（DCA）showed that both the complex model and the radiomics model had high net benefits in predicting the solid/micropapillary patterns.Conclusion:The predictive model based on pre-therapy ¹⁸F-FDG PET/CT radiomics features has a good predictive value for the solid/micropapillary patterns in lung adenocarcinoma,and has a guiding significance for clinical risk stratification and the selection of individualized treatment plan.Part Two Value of pre-therapy ¹⁸F-FDG PET/CT radiomics in pre-dicting EGFR mutation status in patients with non-small cell lung cancerObjective:To assess the predictive power of pre-therapy ¹⁸F-FDG PET/CT-based radiomic features for epidermal growth factor receptor（EGFR）mutation status in non-small cell lung cancer.Methods:We included consecutive patients with histologically proven lung adenocarcinoma,who had under gone pre-therapy ¹⁸F-FDG PET/CT scan and were tested for genetic mutations between January 2015 and January 2019in our department.248 lung cancer patients were included,including 133EGFR mutant and 115 EGFR wild-type cases.The patients were randomly divided into two sets,in the ratio of 7:3,with 175 cases assigned to the training set and 73 to the validation set.The LIFEx package（version 4.00,http://www.lifexsoft.org）was used to extract texture features of PET/CT images of lesions in the same VOI,and a total of 47 PET-based and 45CT-based radiomic features were extracted.Then,the LASSO algorithm was used for feature screening to establish the radiomics signature in the training set,and the radiomics signature score（rad-score）was calculated for each patient.Logistic regression models（radiomics model,clinical model and complex model）were established by radiomics signature,clinical variables,and their combinations.Receiver operating characteristic curve（ROC）was used to evaluate the predictive performance of the three models.Nomogram based on rad-score and clinical variables were established in the training set.Results:Eventually,10 radiomics features（5 PET features,5 CT features）were extracted by LASSO algorithm to build the radiomics signature.The EGFR-mutation group had higher rad-score than the EGFR-wild type group in both the two sets（both P<0.0001）.The radiomics model showed a significant ability to discriminate between EGFR mutation and EGFR wild type,with area under the ROC curve（AUC）equal to 0.79（95%CI,0.73–0.86）in the training set,and 0.85（95%CI,0.76–0.94）in the validation set.Among clinical variables,gender（OR=3.9,95%CI:2.1–7.3,P<0.001）and smoking history（OR=0.1,95%CI:0.1–0.3,P<0.001）were significant predictors of EGFR mutation.The AUC values of the clinical model established by gender and smoking history in the training and validation set were 0.75（95%CI,0.68–0.82）and 0.69（95%CI,0.58–0.81）,respectively.The complex model,based on the rad-score and clinical variables,had higher AUCs,namely 0.86（95%CI,0.80–0.91）and 0.87（95%CI,0.79–0.95）in the training and validation set,respectively.There were statistically significant differences in AUC between the complex and the clinical model in both the two sets（both P<0.0001）.In the training set,the AUC of the complex and the radiomics models were also significantly different（P=0.0194）.However,in the validation set,the AUC of the complex model was slightly higher than that of the radiomics model,but there was no statistically significant difference（P=0.6974）.The Hosmer–Lemeshow test showed good agreement between the predicted and observed values in either the training（χ²=3.568,P=0.894）or the validation set（χ²=11.196,P=0.191）.Conclusion:Radiomics based on PET/CT has good predictive value for EGFR mutation in non-small cell lung cancer,and has a guiding significance for the clinical decision-making of targeted therapy.Part Three Prediction of EGFR mutation subtypes in non-small cell lung cancer by ¹⁸F-FDG PET/CT radiomicsObjective:To investigate the value of pre-therapy ¹⁸F-FDG PET/CT radiomic features in differentiating EGFR mutation subtypes（Exon19deletions,Exon21 L858R missense）.Methods:We included 172 NSCLC patients with definite diagnosis of EGFR mutation in our department from January 2015 to November 2019.All patients underwent pre-therapy ¹⁸F-FDG PET/CT examination,with clear pathological diagnosis and EGFR mutation detection results,including 75patients with Exon 19 deletions and 97 patients with Exon 21 L858R missense mutation.We randomly divided the patients into two sets in the ratio of 7:3,with 121 cases in the training set and 51 cases in the validation set.The LIFEx package（version 4.00,http://www.lifexsoft.org）was used to extract radiomic features of PET/CT images of lesions in the same VOI,and 47 PET-based features and 45 CT-based features were extracted.Then,the LASSO algorithm were used for feature screening in the training set.Three machine learning models were constructed based on the selected optimal feature subsets:Logistic Regression（LR）,Random Forest（RF）,and Support Vector Machine（SVM）.The model parameters were determined by 10-fold cross-validation in the training set.The receiver operating characteristic（ROC）curves of the three models were drawn in the two sets,and the AUC,sensitivity,specificity and accuracy were calculated to evaluate the performance of the three models.Finally,Decision Curve Analysis（DCA）was used to evaluate the clinical value of the three models.Results:A total of 9 radiomics features（6 PET features and 3 CT features）were screened by Lasso regression algorithm.The three machine learning models have similar predictive performance.In the training and validation set,the random forest model（AUC:0.79,0.77;Accuracy:74.38%,68.63%）,The support vector machine model（AUC:0.76,0.75;Accuracy:70.25%,72.55%）,The logistic regression model（AUC:0.77,0.75;Accuracy:73.55%,82.35%）.The results of Delong test showed that there were no statistical significance in the AUC values of the three models in both the training and validation sets（RF vs SVM,P=0.406,0.836;RF vs LR,P=0.723,0.881;SVM vs LR,P=0.736,0.965）.However,from the perspective of the accuracy of the models,the logistic regression model appeared to be more stable in both the two sets.Decision curve analysis（DCA）showed that the curves of the three machine learning models were all higher than the extreme curves,and the range of selectable thresholds were all larger,which had good net benefits and clinical value in predicting EGFR mutation subtypes.Conclusion:Radiomic features based on ¹⁸F-FDG PET/CT images have a good performance in predicting the major EGFR mutation subtypes,providing a non-invasive method for the identification of exon 19 deletions and exon 21 L858R missense mutations,which has certain guiding significance for clinical decision-making and the formulation of individualized treatment plan.