Predicting EGFR mutation status in lung adenocarcinoma: development and validation of a computed tomography-based radiomics signature

Abstract

Patients with epidermal growth factor receptor (EGFR) mutations in lung adenocarcinoma can benefit from targeted therapy. However, noninvasively determination of EGFR mutation status before targeted therapy remains a challenge. This study constructed a nomogram based on a combination of radiomics features with the clinical and radiological features to predict the EGFR mutation status. The least absolute shrinkage and selection operator (LASSO) and Wilcoxon test were used for feature selection. Decision tree (DT), logistic regression (LR), and support vector machine (SVM) classifiers were used for radiomics model building. Used the clinical and radiological features establish clinical-radiology (C-R) model. The C-R model with the best radiomics model to establish clinical-radiological-radiomics (C-R-R) model. The predictive performance of the model was evaluated by ROC and calibration curves, and the clinical usefulness was assessed by a decision curve analysis. The current study showed that twelve radiomics features were significantly associated with EGFR mutations. The best radiomics signature model was obtained using the SVM classifier. The C-R-R model had the best distinguishing ability for predicting the EGFR mutation status, with an AUC of 0.849 (95% CI, 0.805-0.893) and 0.835 (95% CI, 0.761-0.909) in the development and validation cohorts, respectively. Our study provides a non-invasive C-R-R model that combines CT-based radiomics features with clinical and radiological features, which can provide useful image-based biological information for targeted therapy candidates.

Introduction

An understanding of the pathological and molecular aspects of lung cancer has made great progress in recent years [1]. The discovery of lung cancer driver genes, especially the epidermal growth factor receptor (EGFR) gene, has increased the uptake of individualized targeted therapy [2,3]. Patients with EGFR mutations in lung adenocarcinoma respond well to EGFR tyrosine kinase inhibitors (TKI) [4-7]. By contrast, patients without these mutations are not eligible for treatment with the TKI at any disease stage [8]. These findings suggest that it is essential to determine patient EGFR mutation status ahead of treatment.

Currently, an analysis of the EGFR mutation status involves invasive procedures such as biopsy or surgical removal to acquire a specimen. However, these procedures are associated with limitations such as a small number of biopsy samples and the risk of sampling errors, among others [9,10]. In addition, tumors tend to be heterogeneous; thus, tissue samples obtained from a particular site might not be representative of the whole tumor [11]. Recently, a “liquid” biopsy has been proposed as an alternative approach; it involves the analysis of nucleic acids present in the peripheral blood and has high specificity for the detection of EGFR mutations. However, the sensitivity of this method is low, and the false negative rate is estimated at 30%; further, the recognized molecular variation may be unrelated to the tumor [12,13]. Overall, these findings suggest a need for a non-invasive and easy-to-use method to determine the EGFR mutation status.

Computed tomography (CT) is the preferred imaging method for lung cancer screening and diagnosis. Previous studies have shown that female sex, a non-smoking status, ground-glass opacity, air bronchogram findings, bubble-like lucencies, pleural retraction, or other CT signs are indicative of EGFR mutation [10,14-16]. However, these studies had a few limitations. For example, CT scan evaluation was performed subjectively by the observer and could not be quantified. In addition, the findings were inconsistent and their accuracy was relatively low.

Radiomics is a novel, non-invasive, and promising method, which uses advanced image algorithms in artificial intelligence to extract high-throughput features from medical images, quantify higher-dimensional features that cannot be observed in the human visual system, and apply useful image features to clinical decision-making [17,18]. Some studies have tried to predict EGFR mutation status with a CT-based radiomics model [11,19-25]. However, these studies were limited by a relatively small sample size, low accuracy, less radiomics characteristics, or the lack of validation data sets. Therefore, heterogeneity of the entire tumor cannot be effectively evaluated. In addition, clinical risk factors and radiological features were not included in the previous models despite evidence that including these features may improve the diagnostic performance of a model [11,20]. Therefore, we established a radiomics model based on the preoperative CT radiomics signature to predict the EGFR mutation status in this study. Further, we describe a user-friendly radiomics nomogram that combines the clinical risk factors, radiological features, and the radiomics signature.

Materials and methods

Patient selection

The institutional review board of our institution approved this retrospective study, and the need for informed consent was waived. We obtained the medical records of 780 patients with lung adenocarcinoma diagnosed by histopathology from January 2016 to May 2020, and collected their preoperative non-enhanced CT images and clinical data. The inclusion criteria were as follows: (1) patients with adenocarcinoma as the histological subtype according to the 2015 World Health Organization (WHO) lung cancer classification; (2) patients with thin-slice CT (1.25 mm) images and complete clinical data that can be used in the picture archiving and communication system (PACS); (3) age over 18 years; (4) patients who underwent a chest CT scan within 2 weeks before biopsy or surgery; (5) patients with no previous history of other malignant tumors; (6) and patients did not receive lung cancer-related treatment (such as chemotherapy, radiotherapy, or immunotherapy) prior to CT scanning. The exclusion criteria were as follows: (1) patients whose EGFR mutation status has not been tested; (2) patients with a histological subtype of lung cancer other than adenocarcinoma; (3) patients in whom the tumor boundary could not be easily delineated owing to massive pleural effusion or inflammation; (4) and patients whose CT image quality was poor. The flowchart of the inclusion and exclusion criteria is listed in the supplementary materials (Figure S1).

Based on the above criteria, a total of 420 patients [mean age ± standard deviation (SD), 57.43 ± 9.36 years; median age, 56.5 years; range age, 21-82 years, including 201 women (mean age ± SD, 56.30 ± 8.51 years; median age, 55.0 years; range age, 31-82 years) and 219 men (mean age ± SD, 58.46 ± 9.99 years; median age, 59.0 years; range age, 21-79 years)] were enrolled in this study. Patients were randomly divided into development and validation cohorts at a ratio of 7:3. Clinical variables included age, sex, smoking history [including non-smokers (never smokers) and smokers (previous and current smokers)], carcinoembryonic antigen (CEA) level, and 14 CT image features. The relationship between the clinical variables and radiological features of the two cohorts of patients and the EGFR mutation status is listed in the supplementary materials (Table S1).

Analysis of the EGFR mutation status

EGFR mutations were detected using the EGFR detection kit (Beijing SinoMD Gene Detection Technology Co., Ltd., China). The polymerase chain reaction (PCR)-based amplified refractory mutation system (ARMS) method was used to confirm the mutations in EGFR exons 18, 19, 20, and 21. All methods were performed according to the manufacturers’ instructions.

CT scanning protocol

CT scanning was performed using two spiral CT systems (Discovery CT750 HD, GE Healthcare, Waukesha, WI, USA; Philips iCT 256, Koninklijke Philips N.V.). The scanning parameters of the CT scanners were as follows: tube voltage, 120 kVp; tube current, 150-200 mA; tube rotation time, 0.5-1.0 seconds; collimator width, 40 mm; matrix, 512 × 512; axial image layer thickness, 5 mm; layer spacing, 5 mm; reconstruction layer thickness, 1.25 mm; and reconstruction layer interval, 1.25 mm. The scan range was from the tip to the bottom of the lung.

CT image analysis

Two radiologists with 5 years and 16 years of experience in thoracic tumor diagnosis, who were blinded to the clinical and histological data, independently evaluated the radiological features (Table S1) of all patients on the PACS. In case of disagreement, a consensus was reached after discussion. All results were analyzed in the lung window (width, 1500 HU; level, -500 HU) and mediastinum settings (width, 300 HU; level, 40 HU).

Radiomics feature selection

Tumor segmentation and feature extraction

Tumor segmentation is a key step in radiomics feature extraction and model building. At present, manual segmentation is the most accurate and recognized segmentation method [19]. Preoperative thin-layer CT images were uploaded to ITK-SNAP 3.8 (http:/www.itksnap.org) in the medical digital imaging and communication (DICOM) format for three-dimensional (3D) manual segmentation of the regions of interest (ROI) [26]. Bin width was set to 25, that is, every 25 gray units corresponded to one gray level, for a total of 80 (2000/25) gray levels. To minimize between-observer differences [27], a radiologist (G.J.Z) with 5 years of experience in chest diagnostics manually delineated the ROI on the axial image of the CT lung window (width, 1500 HU; level, -500 HU), and then confirmed it in the coronal and sagittal positions. If the ROI was found to be inaccurate, further manual adjustment was required. Finally, the segmented region outlined on each slice was merged to generate the volume of interest (VOI) [28]. Each VOI was verified by another radiologist (J.Z) with 16 years of experience. Both readers were blinded to the clinical and histological data of all patients.

For each accurately segmented VOI, the open source Python software package PyRadiomics (https://pyradiomics.readthedocs.io/en/Latest/) was used to automatically extract the radiomics features in the VOI. A total of 1468 radiomics features were extracted from each VOI. The features were divided into three main categories: (1) first-order features; (2) shape features; (3) and texture features [including gray-level co-occurrence matrix features (GLCM); gray-level run-length matrix features (GLRLM); gray-level size zone matrix features (GLSZM); neighboring gray-tone difference matrix features (NGTDM); and gray-level dependence matrix features (GLDM)].

Feature selection

Radiomics feature extraction and analysis workflow is shown in Figure 1. Feature selection is important to improve the generalization ability and optimize the model [19]. A large number of radiomics features may lead to overfitting, reducing model classification ability. The radiomics features used to build the model account for only a small part of the total. Therefore, in order to reduce the redundancy between the features, we first standardized all radiomics features in the development cohort using the z-score method and applied the same way on the validation cohort. Subsequently, Wilcoxon rank sum test was used to retain features with a P-value of < 0.0000341 (0.05/1468, the significance level of the test level α = 0.05 divided by the number of radiomics features for Bonferroni correction) [29,30]. Least absolute shrinkage and selection operator (LASSO) regression was applied to select the most relevant predictive features. The radiomics features selection process is described in Figures S2 and S3.

Figure 1.

Radiomics features extraction and analysis workflow. A. Original CT images of a patient with lung adenocarcinoma. B. Segmentation of the tumor volume of interest (VOI) on all CT slices by experienced radiologists. C. Feature extraction from the VOI, including tumor shape, intensity, texture, and wavelet features. D. Clinical, radiological, and radiomics feature analysis. E. Model building.

The selection of clinical and radiological features was based mainly on their correlation with the EGFR mutation status [14,15]. Firstly, univariate analysis was used to select the clinical and radiological features that were significantly different from the EGFR mutation status in the development cohort; next, multiple logistic regression analysis was further used to select the most relevant variables.

To evaluate the reproducibility and robustness of the feature extraction process, 3 months later, 40 patients were randomly selected from the development cohort, and the radiologist (G.Z.J) segmented the data again to construct a re-segmentation set. In addition, 40 patients were randomly selected from each CT scanner to establish different CT scanner sets to calculate intra-/interclass correlation coefficients (ICCs); values > 0.80 indicated good consistency.

Radiomics signature building

After eliminating the redundant features, we input the final selected radiomics features into the classifier to construct a radiomics signature for biological evaluation. In this study, three classifiers were evaluated, namely logistic regression (LR), decision tree (DT), and support vector machine (SVM). For DT, the size of the tree is controlled by complex parameters (CP). The CP value of the DT model was selected with the minimum ‘Bxerror’ (the mean value of 10-fold cross-validation error) [31]. SVM classifier uses radial basis function as the kernel function and employs 10-fold cross-validation to select the best-performing model during development.

The receiver operating characteristic (ROC) curve was used to calculate the area under the curve (AUC), sensitivity, specificity, and accuracy to evaluate the performance of different radiomics models, and the radiomics model with the highest AUC was taken as the best model.

Used the clinical and radiological features independently related to EGFR mutation status after multivariate analysis to establish a clinical-radiology (C-R) model. Then, combined the C-R model with the best radiomics model to establish a clinical-radiological-radiomics (C-R-R) model and calculated the diagnostic efficiency of the C-R-R model.

The goodness of fit of the model was evaluated by the calibration curve and the Hosmer-Lemeshow test [32]. For the development and validation cohorts, a decision curve analysis (DCA) was used to calculate the net benefits of each model under different threshold probabilities to evaluate the clinical usefulness of the model.

Statistical analyses

All statistical analyses were performed with R 3.6.3 (http://www.rproject.org) and IBM SPSS Statistics for Windows 22.0 (IBM Corporation, USA). Radiomics feature extraction were performed on Python 3.6.3 (https://www.python.org) with PyRadiomics tool kit. The chi-square or Fisher’s exact tests were used to evaluate on category variables as appropriate. Independent sample t-test, and Mann-Whitney U test were used to evaluate on continue variables as appropriate. Receiver operating characteristic (ROC) curve analysis was performed to compare the results of models in both sets. Youden’s index was used to determine the optimal cutoff point in the ROC analysis and the corresponding accuracy, sensitivity and specificity were also calculated. The between-model differences in AUC values were compared by DeLong test. Two-sided P-values of less than 0.05 were considered of a statistically significant finding.

Results

Clinical and radiological features of patients

A total of 420 patients were enrolled in this study, of which 294 and 126 were in the development and validation cohorts, respectively. The rate of EGFR mutation in total, development and validation groups was 50.5% (212/420), 50.34% (148/294), and 50.8% (64/126), respectively. There was no significant difference in the EGFR mutation rate between the development and validation cohorts (P > 0.05). In addition, there was no significant difference between other clinical and radiological features between the two cohorts (all P > 0.05). The balance of data between the two cohorts demonstrates that the grouping of patients in this study was reasonable (Table S1).

The relationship between clinical and radiological features and the EGFR mutation status in the development cohort is shown in Table S2. Univariate analysis revealed that 12 features were significantly different between the EGFR mutant and wild-type groups. Multivariate analysis revealed that smoking history (odds ratio [OR], 0.373; 95% confidence interval [CI], 0.182-0.765; P = 0.007), bubble-like lucency (OR, 3.669; 95% CI, 1.975-6.816; P < 0.001), pleural attachment (OR, 0.296; 95% CI, 0.148-0.594; P = 0.001), and pleural retraction (OR, 2.207; 95% CI, 1.188-4.100; P = 0.012) were correlated independently with the EGFR mutation status.

Radiomics feature selection and signature building

A total of 1468 radiomics features were successfully extracted from each patient’s VOI. In order to establish radiological markers, we first performed univariate analysis in the development cohort. According to Bonferroni correction, 57 radiomics features were selected, and then simplified to 12 potential predictors by LASSO regression (Table S3). In the development cohort, ‘wavelet.LLH_firstorder_Kurtosis’ and ‘wavelet.LLL_glszm_GLNN’ were significantly lower while the other 10 radiomics features were significantly higher in the EGFR mutant group than in the wild-type group (Figure 2). In the development cohort, the 12 radiomic features selected above were used to construct three different radiomics models for predicting EGFR mutations using LR, DT, and SVM classifiers. The correlation analysis of 12 radiomics features with one clinical variable and three radiological features is shown in Figure 3.

Figure 2.

Boxplot showing 12 radiomics features that are significantly different between the EGFR mutant and wild groups in the development cohort.

Figure 3.

Chord diagram of the correlation between a clinical variable and three radiological and 12 radiomics features. Correlation analysis between selected radiomics features and clinical and radiological features in the development (A) and validation (B) cohorts. Pearson’s correlation analysis showed that each link is significant correlation (P < 0.05). The width of the link indicates the relative strength. For example, pleural attachment was significantly correlated with LLL_glszm_GLNN and LLL_firstorder_Kurtosis in the development and validation cohorts.

Radiomics signature predictive performance and validation

The results of the radiomics signature and C-R model in the two cohorts are shown in Table 1 and Figure 4. The AUC values of SVM in the two cohorts [0.810 (95% CI, 0.761-0.859) and 0.796 (95% CI, 0.717-0.876), respectively] were higher than those of the other two classifiers (LR and DT). The AUC values of LR in the two cohorts were 0.783 (95% CI, 0.731-0.836) and 0.778 (95% CI, 0.695-0.862), and those of DT were 0.768 (95% CI, 0.716-0.819) and 0.761 (95% CI, 0.679-0.844). The AUC values of the C-R model in the two cohorts were 0.776 (95% CI, 0.724-0.828) and 0.739 (95% CI, 0.652-0.826), respectively.

Table 1.

Performance of the radiomics signature, C-R and C-R-R models

	AUC	Accuracy	Sensitivity	Specificity
Development
LR	0.783 (0.731-0.836)	0.728 (0.673-0.778)	0.774 (0.664-0.849)	0.682 (0.514-0.743)
DT	0.768 (0.716-0.819)	0.731 (0.677-0.781)	0.637 (0.496-0.726)	0.824 (0.695-0.877)
SVM	0.810 (0.761-0.859)	0.748 (0.695-0.797)	0.726 (0.568-0.801)	0.770 (0.648-0.831)
C-R	0.776 (0.724-0.828)	0.711 (0.655-0.762)	0.720 (0.675-0.794)	0.608 (0.455-0.693)
C-R-R	0.849 (0.805-0.893)	0.786 (0.734-0.831)	0.808 (0.685-0.870)	0.764 (0.642-0.845)
Validation
LR	0.778 (0.695-0.862)	0.738 (0.652-0.812)	0.773 (0.560-0.924)	0.700 (0.533-0.834)
DT	0.761 (0.679-0.844)	0.762 (0.678-0.833)	0.652 (0.368-0.731)	0.883 (0.624-0.953)
SVM	0.796 (0.717-0.876)	0.746 (0.661-0.819)	0.742 (0.499-0.909)	0.750 (0.583-0.850)
C-R	0.739 (0.652-0.826)	0.714 (0.627-0.791)	0.712 (0.448-0.803)	0.717 (0.467-0.817)
C-R-R	0.835 (0.761-0.909)	0.802 (0.721-0.867)	0.773 (0.485-0.879)	0.833 (0.617-0.917)

AUC, area under the curve; C-R, clinical-radiological; C-R-R, clinical-radiological-radiomics; DT, decision tree; LR, logistic regression; SVM, support vector machine.

Figure 4.

Receiver operating characteristic (ROC) curves from LG, DT classifier, SVM classifiers, C-R model, and C-R-R model used to predict EGFR mutations in development (A) and validation (B) cohorts. C-R, clinical-radiological; C-R-R, clinical-radiological-radiomics; DT, decision tree; LR, logistic regression; SVM, support vector machine.

The C-R-R model developed by combining the radiomics signature derived from the higher diagnostic efficiency of SVM and the C-R model produced higher AUC values in the development and validation cohorts: 0.849 (95% CI, 0.805-0.893) and 0.835 (95% CI, 0.761-0.909), respectively. In addition, the C-R-R model had the best discriminative ability in both cohorts, with sensitivities of 0.808 (95% CI, 0.685-0.870) and 0.773 (95% CI, 0.485-0.879) and specificities of 0.764 (95% CI, 0.642-0.845) and 0.833 (95% CI, 0.617-0.917). In the development cohort, the scores of the radiomics, C-R model, and C-R-R model were significantly higher in the EGFR mutant group than in the wild-type group. This result was confirmed in the validation cohort (Figure 5).

Figure 5.

Boxplot showing the comparison of radiomics scores from the SVM classifier (A, D), C-R model (B, E), and C-R-R model (C, F) in the development (A-C) and validation (D-F) cohorts. C-R, clinical-radiological; C-R-R, clinical-radiological-radiomics; SVM, support vector machine.

Based on the above results, the performance of the C-R-R model in the development cohort was significantly better than that of the SVM classifier (AUC, 0.81; P = 0.008), LR classifier (AUC, 0.783; P < 0.001), DT classifier (AUC, 0.768; P < 0.001), and C-R model (AUC, 0.776; P < 0.001). Similarly, the C-R-R model had the best performance in the validation cohort, which was significantly different from that of the C-R model (P = 0.004) but not from that of the three classifiers (P > 0.05). In short, the C-R-R model had a higher predictive value for EGFR mutations. In both cohorts, the AUC value, accuracy, and sensitivity of the C-R-R model in predicting the EGFR mutation status were higher than other classifiers and the C-R model, and the effect of the LR classifier was better than that of the DT classifier and the C-R model.

Clinical application of the C-R-R model

A nomogram was constructed based on the radiomics model with the highest AUC value in combination with one clinical variable and three radiological features. Among the five models, the radiomics nomogram had the best discriminating ability (Figures 4 and 6A).

Figure 6.

Building and performance of the C-R-R model. (A) The nomogram developed based on the C-R-R model. (B, C) Calibration curves of different classifiers and models generated from the development (B) and validation (C) cohorts. The goodness of fit of the predicted probability from different classifiers and models with the actual outcomes of EGFR mutations was evaluated. The x-axis represents the EGFR mutation probability calculated by different classifiers and models, and the y-axis represents the actual EGFR mutation probability. The diagonal line represents the ideal estimate of the ideal model. (D, E) Decision curves of different classifiers and models generated from the development (D) and validation (E) cohorts. The x-axis shows the threshold probability, and the y-axis measures the net income. C-R-R, clinical-radiological-radiomics.

The calibration curve revealed that the probability of predicting EGFR mutations by the nomogram was in good agreement with the actual probability (Figure 6B, 6C). The Hosmer-Lemeshow test was applied to the SVM classifier, C-R model, and C-R-R model with the best discrimination ability in the development and validation cohorts. The values of the SVM classifier in the two cohorts were 0.625 and 0.251, those of the C-R model were 0.646 and 0.111, and those of the C-R-R model were 0.426 and 0.313, respectively, indicating that the C-R-R model did not deviate from the perfect fit in both cohorts (Figure S4).

The decision curve was used to evaluate the clinical performance of different classifiers and models from the perspective of clinical application, thereby reflecting the clinical applicability of different classifiers and models (Figure 6D, 6E). The area under the decision curve of the C-R-R model was larger than that of other models, indicates that the model had the best clinical utility. When the threshold probability was in the range of 9%-83%, the C-R-R model provides net benefits.

Discussion

In this study, we constructed a C-R-R model based on the combination of a CT radiomics signature with clinical and radiological features to predict the EGFR mutation status of lung adenocarcinoma using a relatively large data set. We used three classifiers (LR, DT, and SVM) to calculate the diagnostic performance of radiomic signatures, and finally selected the SVM classifier with the highest AUC value. The C-R-R model was developed based on data from 294 patients. To further verify its diagnostic performance, we evaluated the model in a validation cohort of 129 patients. In the development and validation cohorts, the AUC values of the C-R-R model were high (0.849 and 0.835, respectively). Therefore, this model may help determine the EGFR mutation status of lung adenocarcinoma to guide personalized targeted therapy.

The EGFR mutation rate in this study was 50.5% (212/420), which was consistent with that in previous studies [10,14,15,33]. In the development and validation cohorts, the rate was 50.3% (148/294) and 50.8% (64/126), respectively. Our study found that sex and smoking history were significantly related to the EGFR mutation status: women and non-smokers were more likely to have EGFR mutations. In addition, our findings showed that 10 radiological features were significantly related to the EGFR mutation status (P < 0.05). For example, bubble-like lucency was more common in patients with EGFR mutations, while pleural attachment was more common in wild-type patients, which was consistent with previous studies [10,14]. In multivariate regression analysis, smoking history, bubble-like lucency, pleural attachment, and pleural retraction were independently correlated with EGFR mutation status (P < 0.05), while sex and seven other radiological features were not significantly different between the two groups (P > 0.05).

To construct the radiomics signature, we screened 12 independent features highly correlated with EGFR mutations from 1468 candidate features, which were stable in the validation cohort. Since most of the selected radiomics features (11/12) were extracted from the filtered image, among which seven features were obtained through wavelet transform, so the texture and high-dimensional features were more strongly correlated with the EGFR mutation status. Wavelet transform is currently one of the commonly used methods for signal processing, such as noise elimination, and data smoothing and filtering. It can smooth the image and improve the ability to acquire features related to tumor heterogeneity [31]. Texture features cannot be recognized by the human visual system, nor can they be understood as specific meanings [34,35]. Our results showed that the radiomics features, including original_firstorder_Skewness, exponential_glrlm_SRE, and wavelet.LLL_glszm_ZE, were higher in the EGFR mutation group and were significantly related to EGFR mutations. Among them, exponential_glrlm_SRE was related to run length and fine textural texture. A larger value represented a shorter run length and a finer textural texture [36,37], while wavelet.LLL_glszm_ZE was related to texture heterogeneity; a higher value represented more extensive heterogeneity in texture patterns [38]. However, the features of wavelet.LLL_firstorder_Kurtosis and wavelet.LLL_glszm_GLNN were more extensive in the wild-type group, and were significantly correlated with wild-type EGFR. Among them, wavelet.LLL_firstorder_Kurtosis was related to the “peakedness” in the image ROI. A higher kurtosis value implies that the quality of the distribution was concentrated in the tail [39], and wavelet.LLL_glszm_GLNN was related to the gray value in the image. The lower the value, the greater the similarity of the intensity value [38]. This indicated that the finer the texture, the greater the heterogeneity within the tumor and the greater the variability of gray values, alongside the higher risk of EGFR mutations. These findings are consistent with those of a previous study [11]. Moreover, it also shows that these features are closely related to radiological characteristics. For example, in the development and validation cohorts, pleural attachment was significantly correlated with features such as LLL_glszm_GLNN and LLL_firstorder_Kurtosis, pleural retraction was significantly correlated with features such as LLL_glszm_ZE and LLL_glszm_GLV, bubbe-like was significantly correlated with features such as LLL_glcm_SE and grad_gldm_DE, and smoking was significantly correlated with features such as exp_gldm_SDE and fo_Skewness. The correlation between these features indicated that although texture-based features are not observable to the naked eye, specific combinations of several texture features can to some extent be explained by certain radiological features.

Medical images are routinely collected in the hospital during the diagnosis and treatment of patients with cancer, and the genotypes of tumors can be reflected to a certain extent by evaluating the characteristics of these images. For example, some studies have found that radiological features can predict the state of EGFR mutation status [10,14,15]; however, these studies have selected a large number of radiological features related to EGFR mutations, which will undoubtedly increase the clinical workload; and using a large number of features is associated with an increased risk of error. In addition, some studies have investigated the association between radiomics features and tumor genotypes [11,19-25]. However, these studies usually have a small sample size or low diagnostic efficiency, and more importantly, some studies have lack validation, which is an important part of radiomics analysis [18,40]. Our study selected the independent predictors most associated with EGFR mutation status through multivariate analysis of clinical and radiological features. The radiological features included in the model were mainly because certain features were related to the EGFR mutation status. However, when manually delineating the tumor ROI, a small part of these radiological features, such as pleural depression and fine spiculation, were too small to be included in the ROI. Therefore, the omission of these features may lead to the reduction of feature information, which cannot fully reflect the heterogeneity of the entire tumor. Our research showed that the AUC of the model containing only the radiomics signature was 0.81, while the AUC of the model containing clinical variables and radiological features (C-R model) was only 0.776, and the AUC of the fusion model (C-R-R model) was increased to 0.849. This demonstrated that the inclusion of clinical variables and radiological features in the model can improve the diagnostic performance of the model. In addition, the diagnostic performance of the radiomics model was better than that of the C-R model.

To our knowledge, this is the first report on prediction of the EGFR mutation status using a relatively large data set based on CT radiomics signature combined with clinical and radiological features. Three different classifiers were applied to evaluate the performance of the radiomic model, and finally the SVM classifier with the highest diagnostic performance was selected. Yang et al. [11] have reported that a model based on the CT radiomics signature can be used to predict the EGFR mutation status of patients with lung adenocarcinoma, with an AUC of 0.826 in the training cohort and only 0.779 in the validation cohort. Velazquez and colleagues [20] used the CT radiomics signature combined with clinical variables to predict the EGFR mutation status, and the AUC obtained only 0.75, and lack of verification, limiting its clinical applicability. In this study, the proposed C-R-R model shows good prediction performance in the development cohort (AUC, 0.849), and is well calibrated and stable (AUC, 0.835) in the verification cohort. Moreover, the DCA confirmed the clinical usefulness of the C-R-R model.

TKI therapy can provide significant clinical benefits to patients with EGFR mutations. Compared to treatment with platinum-based chemotherapy alone, approximately 70% of patients receiving TKI therapy experience symptom alleviation, improved quality of life, and prolonged progression-free survival [3,41,42]. Therefore, determining the EGFR mutation status is a prerequisite for receiving targeted therapy. Although image analysis cannot replace histological examination, it can provide additional information and help identify high-risk patients with EGFR mutations [3,10]. For example, when radiomics predicts a high possibility of EGFR mutation in patients with false-negative histological results, the tumors should be re-sampled for biopsy; otherwise, these patients will not benefit from targeted therapy [10]. Similarly, in patients with multiple tumors, radiomics can be used to select the most suspicious tumor for biopsy [43]. Therefore, when analysis based on histological examination is not feasible, imaging-based biomarkers can potentially be used in clinical practice.

There are several limitations to this study. First, like any other retrospective study, the results of this study may be affected by selection bias, and we encourage researchers from more centers to use different CT scanners and parameters to prospectively verify our findings. Second, this study was limited to the analysis of lung adenocarcinoma and did not involve other pathological subtypes because most EGFR mutations are present in adenocarcinoma. Finally, the 3D delineation of all tumors in this study was performed manually by a radiologist; in future studies, tumors may be effectively segmented automatically.

In conclusion, preoperative prediction of the EGFR mutation status will help guide individualized targeted therapy. On radiomics analysis, 12 radiomic features were highly correlated with EGFR mutations, which was confirmed in the validation cohort. A C-R-R model was constructed based on CT radiomic features combined with the clinical variables and radiological features. The model showed excellent diagnostic performance and high sensitivity in predicting EGFR mutation status and could provide useful image-based biological information for patients eligible for targeted therapy.

Disclosure of conflict of interest

None.

Abbreviations

AUC

area under the curve

C-R

clinical-radiological

C-R-R

clinical-radiological-radiomics

CT

computed tomography

DT

decision tree

EGFR

epidermal growth factor receptor

LR

logistic regression

LASSO

least absolute shrinkage and selection operator

PACS

picture archiving and communication systems

ROI

region of interest

TKI

tyrosine kinase inhibitors

SVM

support vector machine

VOI

volume of interest