Predicting The EGFR Mutation Status In Patients With Non-small Cell Lung Cancer Based On Deep Learning And Radiomics

Objective: The occurrence of non-small cell lung cancer(NSCLC)often means poor prognosis and reduced survival rate.However,with the wide application of targeted therapy,the carriers of epidermal growth factor receptor(EGFR)mutations,which account for the majority of NSCLC patients,will benefit from targeted drugs and their survival conditions will be effectively improved.With the rapid development of intelligent analysis of digital medical images,lung CT,as a routine screening tool for lung cancer,can predict the gene mutation status of patients.This study used lung CT and combined deep learning with radiomics to evaluate the EGFR mutation status of NSCLC patients non-invasively,assist clinical diagnosis and treatment decisions,and promote precision medicine.Methods: The total of 431 patients with NSCLC were collected retrospectively in this study,and the ROI was manually delineated by two experienced clinicians.After the normalization and standardization of CT image,the ROI was extracted,and the data set was divided into training set and testing set at a ratio of 8:2.Using the Pyradiomics package to extract the radiomics features,using the U test and m RMR algorithm for feature selection,and finally building the radiomics model based on the random forest algorithm.A 3D-Res Net-101 model was applied to build a deep learning model with transfer learning strategy and extract deep learning features.In addition,a feature fusion model was constructed by combining the output of the deep learning model with radiomics features.The best model was obtained by comparing the performance of different models on the testing set.Finally,the correlation analysis and cluster analysis were carried out on the radiomics features and deep learning features to clarify the relevant features of EGFR mutation statue and explore the characteristics and relationships of different features.Results: A total of 1648 radiomics features were extracted in this study,and 7 features were included in the radiomics model after feature screening.The AUC on the testing set reached 0.665,and after adding clinical features,the AUC reached 0.699.The AUC of the deep learning model on the testing set is 0.721,and the AUC after adding clinical features is 0.779.After deep learning combined with radiomics and clinical features,the AUC of the fusion model was 0.821,which was significantly higher than that of other models(p<0.05).In the DL+Rad+clinical model,two radiomics features most relevant to the EGFR mutation status of patients were included: log-sigma-4-0-mm-3D＿firstorder＿Median 、 wavelet-HLL＿glrlm＿Run Entropy and two clinical features:gender and smoking history.The result of feature analysis showed that the deep learning features are more closely related to the EGFR mutation statue than radiomics features,and some deep learning features were consistent with radiomics features.Conclusions: This study constructed the optimal prediction model of EGFR mutation statue in NSCLC patients through the combination of deep learning and radiomics,recognized the radiomic features and clinical features related to EGFR mutation statue,and achieves a relatively ideal prediction effect on the testing set.Compared with using single method,the combination of deep learning and radiomics can help improve the performance of EGFR mutation statue prediction model,so as to provide more accurate diagnosis and treatment decision-making assistance,and help NSCLC patients in accurate medical treatment.