Predictive Value Of A Model Constructed Based On 18F-FDG PET/CT Radiomics Features And Clinical Factors For EGFR Mutation Status In Lung Adenocarcinoma

Objectives:This study develops a practical model to predict EGFR mutations in lung adenocarcinoma based on ¹⁸F-deoxyglucose（FDG）PET/CT radiomic features and clinical risk factors,and determines the diagnostic accuracy of the model for EGFR mutation status.Methods:155 patients with lung adenocarcinoma were retrospectively analyzed.All patients underwent ¹⁸F-FDG PET/CT at Affiliated Cancer Hospital of Xinjiang Medical University from July 2018 to January 2022and had EGFR gene test results.All patient were randomly divided into training and testing sets according to the ratio of 7:3.All clinical risk factors of EGFR mutation were performed univariate and multivariate logistic regression for to establish clinical risk factor models.Region of interest（VOI）were outlined on pre-processed PET and CT images,respectively,to extract radiomic features.Feature dimension reduction and feature selection were performed using Student’t test,Mann-Whitney U test,Pearson correlation test and Lasso.The EGFR mutation of lung adenocarcinoma was predicted by three models of machine learning,Support vector machine（SVM）,random forest classifier（RF）and logistic regression（LR）models based on radiomic features.The best model was selected by Receiver operating characteristic（ROC）curve and was combined with clinical risk factors to construct a composite model.Results:There were 69 cases of wild type EGFR and 86 cases of mutant EGFR in 155 cases of lung adenocarcinoma.Multivariate logistic regression showed that smoking status was a predictor of clinical risks.After feature extraction,feature screening and dimension reduction,15 CT and 3 PET image features were retained.Three different machine learning methods（SVM,RF,LR）predicted the EGFR mutation state in the radiomic features,and the AUC values of the training set were 0.769,0.857,0.783 respectively.In the test set,they were 0.777,0.775 and 0.764 respectively.After Delong’s test,RF model showed better predictive performance.A prediction model of EGFR mutation was constructed by integrating clinical risk factors and radiomic features.Training set:AUC:0.882;Testing set:0.822.Conclusion:Some differences in EGFR mutation prediction by different machine learning models for radiomic features,with the RF model showing better predictive performance;The model constructed based on the¹⁸F-FDG PET/CT radiomic features and clinical risk factors has a greater application in predicting EGFR mutations in lung adenocarcinoma.