Detailed identification of epidermal growth factor receptor mutations in lung adenocarcinoma: Combining radiomics with machine learning

Purpose To investigate the use of radiomics in the in-depth identification of epidermal growth factor receptor (EGFR) mutation status in patients with lung adenocarcinoma. Methods Computed tomography images of 438 patients with lung adenocarcinoma were collected in two different institutions, and 496 radiomic features were extracted. In the training set, lasso logistic regression was used to establish radiomic signatures. Combining radiomic index and clinical features, five machine learning methods, and a tenfold cross-validation strategy were used to establish combined models for EGFR(+) vs EGFR(-), and 19Del vs L858R, groups. The predictive power of the models was then evaluated using an independent external validation cohort. Results In the EGFR(+) vs EGFR(-) and 19Del vs L858R groups, radiomic signatures consisting of 12 and 7 radiomic features were established, respectively; the area under the curves (AUCs) of the lasso logistic regression model on the validation set was 0.76 and 0.71, respectively. After inclusion of the clinical features, the maximum AUC of combined models on the validation set was 0.79 and 0.74, respectively. Logistic regression analysis showed good performance in the two groups, with AUCs of 0.79 and 0.71 on the validation set. Additionally, the AUC of combined models in the EGFR(+) vs EGFR(-) group was higher than that of the 19Del vs L858R group. Conclusions Our study shows the potential of radiomics to predict EGFR mutation status. There are imaging phenotypic differences between EGFR(+) and EGFR(-), and between 19Del and L858R; these can be used to allow patients with lung adenocarcinoma to choose more appropriate and personalized treatment options.