Detection of EGFR mutations in early-stage lung adenocarcinoma by machine learning-based radiomics.

Abstract

We hypothesized that epidermal growth factor receptor (EGFR) mutations could be detected in early-stage lung adenocarcinoma using radiomics. This retrospective study included consecutive patients with clinical stage I/II lung adenocarcinoma who underwent curative-intent pulmonary resection from March-December 2016. Using preoperative enhanced chest computed tomography, 3,951 radiomic features were extracted in total from the tumor (area within the tumor boundary), tumor rim (area within ±3 mm of the tumor boundary), and tumor exterior (area between +10 mm outside the tumor and tumor boundary). A machine learning-based radiomics model was constructed to detect EGFR mutations. The combined model incorporated both radiomic and clinical features (gender and smoking history). The performance was validated with five-fold cross-validation and evaluated using the mean area under the curve (AUC). Of 99 patients (mean age, 66±11 years; female, 66.6%; clinical stage I/II, 89.9%/10.1%), EGFR mutations in the surgical specimen were detected in 46 (46.5%). A median of 4 (range, 2 to 8) radiomic features was selected for each validation session. The mean AUCs in the radiomics and combined models were 0.75 and 0.83, respectively. The two top-ranked features in the combined model were the radiomic features extracted from the tumor exterior and the tumor, indicating a higher impact of radiomic features over relevant clinical features. Radiomic features, including those in the peri-tumoral area, may help detect EGFR mutations in lung adenocarcinomas in preoperative settings. This non-invasive image-based technology could help guide future precision neoadjuvant therapy.