Prediction Of EGFR Gene Mutation Status In Lung Adenocarcinoma Based On Radionomics

Objective: To develop and evaluate radiomics model in classification of lung adenocarcinoma with/without epidermal growth factor receptor(EGFR)mutation by use of thin-section computer tomography(CT).Method: This pilot study included 284 patients with lung adenocarcinoma(141 with EGFR mutation and 143 without EGFR mutation).All patients had thin-section(1 mm axial images)diagnostic CT scans.The region of interest(ROI)for each tumor was volumetrically segmented by use of a semiautomatic approach(ITK-SNAP3.6.0),that is,the ROI is automatically generated by the computer using the threshold segmentation method first and the radiologist could change the contour if it needed.396 radiomics features were extracted by use of a radiomics software-Analysis-Kit(GE Healthcare,Life Science,China).The clinical information and imaging features used to develop the radiomics model include gender,age,smoking status,symptom,and tumor diameter,location,air bronachogram,pleural involvement,and vacuole/cavity.According to the proportion of 7:3,284 patients were randomly divided into train set(n=199)and test set(n=85).The Multi-information method was used to reduce the dimensionality of the radiomics features and the clinical features were screened,and then the Logistic Regression and Random Forest machine learning methods were used to develop the models of radiomics features,clinical features,combined radiomics features and clinical features.Finally,six models were built.The models were compared by receiver operating characteristic(ROC)curve area(AUC),accuracy,sensitivity,specificity,precision,F1 and Brier scores.The efficiency was evaluated using three indicators: ROC curve,confusion matrix and DCA curve.Results: EGFR mutant was significantly associated with female,smoking,and air bronachogram(P<0.05).Four radiomics features were found to be statistically significant,which were grayscale run-length matrices.The model efficiency using radiomics features combined with clinical features was higher than that using radiomic features,clinical features,and the Random Forest model showed the AUC of 0.841,while the Logistic Regression model showed a lower AUC of 0.729.Conclusion: A comprehensive model combining radiomics and clinical features using the Random Forest method can predict the EGFR mutation status of lung adenocarcinoma to a certain extent,and it is possible to improve the management of patient treatment plans in clinical practice.