This study was approved by the ethical committee of our institution, and the requirement for patient-informed consent was waived due to the retrospective nature of study. In total, 1095 patients admitted to our hospital from February 2012 to October 2019 were initially included. The inclusion criteria for target population were that the patients 1) were pathologically confirmed with LADC; 2) obtained EGFR-mutation testing results; 3) had completed clinical data, including age, gender, smoking history, and clinical cancer stage; and 4) had available chest contrast-enhanced CT data. Another 487 patients were excluded using the following criteria: 1) they received antitumor therapy prior to chest CT scans and EGFR gene detection; 2) they had multiple primary tumors; 3) their tumor had a boundary that could not be determined; 4) they had more than one EGFR-mutation subtype. There were 608 LADC patients finally included. Among them, 307 patients (50.5%) had an EGFR-mutation and 301 (49.5%) had a wild-type EGFR. Of the EGFR-mutant patients, 114 (37.1%) harbored a single 19del, 155 (50.5%) harbored a single L858R, and 38 (12.4%) harbored other rare mutations.

Molecular analyses were performed on tumor histologic or cytology samples. The EGFR-mutation statuses were detected using a real-time polymerase chain reaction-based amplification refractory mutation system using the Human EGFR Gene Mutations Detection Kit (Amoy Dx, Xiamen, China). Polymerase chain reaction included the 18 to 21 exons sequence and evaluated the 19del, L858R, T790M, 20 ins, G719X, S768I, and L861Q locus mutations.

All patients underwent chest contrast-enhanced CT scans using one of two CT systems (GE Healthcare, Milwaukee, WI, USA; Siemens Healthineers, Erlangen, Germany). All CT scans were performed at the end of inspiration, during a single breath-hold gap. The parameters were a 100–130 kVp tube voltage, 100–250 mA tube current, 5 mm/5 mm scanning slice thickness/interval, and a reconstruction thickness/interval of 0.625, 1 mm/0.625, 1 mm. After an unenhanced CT scanning, a non-ionic iodized contrast agent (300 mg iodide/mL) was injected through the antecubital vein with a double high-pressure injector at a dose of 1.5 mL/kg body weight (total volume 80–110 mL) at a flow rate of 3.0 mL/s. The contrast agent was followed by a 50 mL injection of saline solution. The arterial and delayed phase acquisition times were triggered at 30 and 120 s, respectively. Finally, the images were transferred to the picture archiving and communication system workstation system and exported to the DICOM format for image feature extraction.

Images were analyzed by two radiologists, blinded to the clinical data, with more than 10 years of experience in chest CT interpretation. A consensus on differences in opinions was reached through consultation. Clinical indicators including age, gender, smoking history, and clinical staging were collected. The following CT features were observed: tumor location (central, involving the segmental or more-proximal bronchi; peripheral, involving the subsegmental or more-distal bronchi), tumor size (the longest tumor diameter in the lung window setting), margin (spiculation, lobulation), density (subsolid, tumor with ground-glass opacity [GGO]; solid, tumor without GGO), internal characteristics (air bronchogram, air-filled bronchus within the tumor; air space, air attenuation within the tumor including cavity and pseudo-cavity; necrosis, focal area of low attenuation without enhancement; calcification), external characteristics (vascular convergence sign, convergence of vessels toward the tumor; pleural retraction, linear or tentiform structures connected between the tumor and pleura), and associated findings (pleural effusion; lymphadenopathy, the short diameter of lymph node >1 cm); multiple pulmonary metastases (number of metastases >10).

All CT images were imported to the Infer Scholar Center platform (https://www.infervision.com/, Infer Scholar). The region of interest (ROI) was manually outlined by two radiologists with more than 10 years of experience in chest CT interpretation using the Infer Scholar Center platform, which was defined as the maximum contour of tumor on axial CT image (20). Five samples were randomly selected from patients with negative EGFR-mutation, 19del-mutation, L858R-mutation, and other rare mutations (20 samples together), respectively, for the ROI segmentation. To assess interobserver repeatability, the ROI segmentation was performed in a blinded way by the two radiologists. To evaluate intra observer repeatability, observer 1 repeated the ROI segmentation 4 weeks after the first assessment. Thereafter, the intra-class correlation coefficients (ICCs) were calculated to evaluate the stability and reproducibility of feature extraction and these features with both inter- and intra-observer ICC values greater than 0.75 were included in this study. Radiomics feature extraction was performed with P-y-Radiomics (https://pyradiomics.readthedocs.io/en/latest/), a flexible open-source platform capable of extracting a large panel of engineered features from medical images (20). For each accurately segmented tumor, the P-y-Radiomics algorithms were used to automatically extract tumor region features. A total of 919 Radiomics features and 18 clinical and CT factoring features were initially extracted.

For model 1, the least absolute shrinkage (LASSO) and selection operator algorithm were used to select the optimal predictive features and a five-fold cross-validation was used to select the best machine learning algorithm. Finally, we used the Gradient Boost Tree by combing radiomics, clinical, and CT morphological features to build model 1. For models 2 and 3, instead of feature selection, 919 Radiomics features and 18 clinical and CT features were initially included for obtaining a better performance. And then, we used light GBM algorithm to conduct feature screening and classification of model modeling. We first trained light GBM model and conducted feature importance ranking using Permutation Importance method, and selected important features through supervised learning for modeling prediction. The permutation feature importance is defined to be the decrease in a model score when a single feature value is randomly shuffled (21). Finally, we used 202 features whose importance score is greater than 0 to build model 2, including 29 first-order features, 4 shape 2D features, 159 advanced textural features (43 GLCM, 32 GLDM, 38 GLRLM, 35 GLSZM, 11 NGTDM) as well as 10 clinical and CT morphological features, and we used 358 features whose importance score is greater than 0 to build model 3, including 64 first-order features, 1 shape 2D features, 282 advanced textural features (106 GLCM, 54 GLDM, 56 GLRLM, 42 GLSZM, 24 NGTDM) as well as 11 clinical and CT morphological features.

In all models, patients were randomly split into training and testing cohorts, 80% and 20%, respectively. The detailed split-sample settings for each model are shown in Figure 1 . Model 1 was used to identify positive and negative EGFR mutations, model 2 was used to distinguish 19del from non-19del mutations, and model 3 was used to determine L858R from non-L858R mutations in LADC patients. The workflow is shown in Figure 2 . The receiver operating characteristic curve (ROC) and the area under the curve (AUC) of training and validation sets as well as the accuracy, sensitivity, and specificity in validation sets were used to evaluate the performance of three models.

Statistical analyses were performed by using SPSS statistics (version 25; IBM, Armonk, NY, USA). The clinical and CT features between patients with positive and negative EGFR mutations, between patients with 19del and non-19del mutations, and between patients with L858R from non-L858R mutations were compared, respectively. Furthermore, for testing whether the background factors between cohorts were balanced, the clinical and CT features of patients in training and validation sets in each model were compared respectively. For continuous variables of clinical and CT morphological features, two independent samples Student’s t test was performed; for categorical variables, Chi-square test was used for comparisons between groups. A two-tailed p-value of < 0.05 was considered statistically significant.

Among the 608 patients with LADC, 272 patients were women and 336 were men with an average age of 61.7 ± 10.4 (range: 30–85) years. For clinical staging, 190 patients (31.2%) were in stages I-II and 418 (68.8%) in stages III-IV. Compared to patients without EGFR-mutation, female, nonsmokers, tumor size<3cm, subsolid density, air bronchogram, air space, spiculation, pleural retraction, vascular convergence sign, and multiple pulmonary metastases were more common in those with EGFR-mutation (all p <0.05). Compared to patients with non 19del-mutation, younger age, female, nonsmokers, tumor size <3cm, subsolid density, air bronchogram, pleural retraction, and vascular convergence sign were more frequent in those with 19del-mutation (all p <0.05). Compared to patients with non L858R-mutation, female, nonsmokers, pleural retraction, vascular convergence sign, and without necrosis were more common in those with L858R-mutation (all p <0.05) ( Table 1 ). The clinical data and CT morphological features of patients in training and validation cohorts for model 1 to 3 were shown in Tables 2 – 4 , respectively. No significant differences were observed in clinical and CT morphological features between both cohorts in each model (all p > 0.05).

Model 1 was built with 137 radiomics features, including 25 first-order features, two shape 2D features, 110 advanced textural features (30 GLCM, 35 GLSZM, 20 GLRZM, 13 NGTDM, and 12 GLDM) as well as 18 clinical and CT morphological features. The AUCs for predicting EGFR-mutation positive cases were 0.969 and 0.886 in the training and validation cohorts, respectively. The accuracy, sensitivity, and specificity of the validation cohort were 0.810, 0.902, and 0.717, respectively ( Figure 3 ).

Model 2 was built with all 202 features and the top 10 ones sorted by their importance scores included 5 radiomics features (5 advanced texture features), 3 clinical features (female, no-smokers, younger age), and 2 CT morphological features (tumor size<3cm, subsolid density). The AUC values for predicting 19del-mutation were 0.999 and 0.847 in the training and validation cohorts, respectively. The accuracy, sensitivity, and specificity in the validation cohort were 0.852, 0.739, and 0.879, respectively ( Figure 4 ). The radiomics, clinical, and CT morphology features used for establishing model 2 and their importance scores are detailed in Supplementary Table 1 .

Model 3 was built with all 358 features and the top 10 ones sorted by their importance scores included 6 radiomics features (2 first-order features, 4 advanced texture features), 2 clinical (no-smokers, female) and 2 CT morphological features (pleural retraction, vascular convergence sign). The AUCs for predicting L858R-mutation were 0.984 and 0.806 in the training and validation cohorts, respectively. The accuracy, sensitivity, and specificity in the validation cohort were 0.713, 0.879, and 0.652, respectively ( Figure 5 ). The radiomics, clinical, and CT morphology features used for establishing model 3 and their importance scores are detailed in Supplementary Table 2 .