Radiomic Feature-Based Nomogram: A Novel Technique to Predict EGFR-Activating Mutations for EGFR Tyrosin Kinase Inhibitor Therapy

Objectives To develop and validate a radiomic feature-based nomogram for preoperative discriminating the epidermal growth factor receptor (EGFR) activating mutation from wild-type EGFR in non-small cell lung cancer (NSCLC) patients. Material A group of 301 NSCLC patients were retrospectively reviewed. The EGFR mutation status was determined by ARMS PCR analysis. All patients underwent nonenhanced CT before surgery. Radiomic features were extracted (GE healthcare). The maximum relevance minimum redundancy (mRMR) and LASSO, were used to select features. We incorporated the independent clinical features into the radiomic feature model and formed a joint model (i.e., the radiomic feature-based nomogram). The performance of the joint model was compared with that of the other two models. Results In total, 396 radiomic features were extracted. A radiomic signature model comprising 9 selected features was established for discriminating patients with EGFR-activating mutations from wild-type EGFR. The radiomic score (Radscore) in the two groups was significantly different between patients with wild-type EGFR and EGFR-activating mutations (training cohort: P<0.0001; validation cohort: P=0.0061). Five clinical features were retained and contributed as the clinical feature model. Compared to the radiomic feature model alone, the nomogram incorporating the clinical features and Radscore exhibited improved sensitivity and discrimination for predicting EGFR-activating mutations (sensitivity: training cohort: 0.84, validation cohort: 0.76; AUC: training cohort: 0.81, validation cohort: 0.75). Decision curve analysis demonstrated that the nomogram was clinically useful and surpassed traditional clinical and radiomic features. Conclusions The joint model showed favorable performance in the individualized, noninvasive prediction of EGFR-activating mutations in NSCLC patients.


INTRODUCTION
With the development of molecular biology in cancer therapy, the treatment of NSCLC patients has become increasingly based not only on the patient's clinical characteristics and tumor morphology but also on individual mutational profiles (1). EGFR-activating mutations, including exon 19 deletion (DEL19) and exon 21 substitution (L858R), account for approximately 90% of all EGFR mutations in advanced NSCLC patients (2). For advanced NSCLC patients with EGFRactivating mutations, treatment with EGFR tyrosine kinase inhibitors (EGFR TKIs), such as gefitinib and afatinib, has become the standard of care (3,4). Accumulating evidence suggests that EGFR TKIs can significantly prolong progression-free survival (PFS) compared to standard chemotherapy in this genetically distinct subset of patients (5,6). Thus, the detection of EGFR-activating mutations at the time of initial diagnosis, before treatment, is critical.
Gene mutation testing can uncover pivotal information connected to underlying molecular biology. The most commonly used approach for obtaining specimens for a specific diagnosis and molecular testing is biopsy. However, the tissue acquired by invasive techniques may fail to represent the anatomic, functional, and physiological properties of cancer. Clinical studies have suggested that 10% to 20% of all NSCLC biopsies are inadequate for molecular analysis because of a lack of either sufficient tumor cells or amplifiable DNA (7). Moreover, intratumoral heterogeneity due to the diverse collection of cells harboring distinct molecular signatures will result in differential levels of sensitivity to treatment (8). Thus, an alternative approach for genetic testing is needed.
Computed tomography (CT) imaging presents a perspective of the entire tumor and its microenvironment, allowing prediction of the EGFR mutation status globally (9,10). Radiomics refers to the computerized extraction of a large number of quantitative radiomic features from radiologic images, and this method has unique potential to reveal tumorrelated information, such as pathological features, biomarker expression and genomic features, using machine learning algorithms (11,12). Radiomics provides quantitative and objective data collected from medical images to be utilized within clinical-decision support systems to improve diagnostic, prognostic, and predictive accuracy, especially in lung cancer (13)(14)(15). Developing such a quantitative imaging technique and testing its validity may offer a new non-invasive and convenient approach for the better management of therapeutic strategies, resulting in optimized clinical and economic benefits to the patient.
Herein, we examined the correlation between 396 radiomic features and EGFR-activating mutation subtypes in two independent cohorts comprising 301 NSCLC patients. Furthermore, we created a user-friendly nomogram by incorporating the radiomic signature with the clinical characteristics to predict the probability of an event based on the individual profile of each patient. Our results reveal that the combination of the repeatable, reproducible and low-cost CTderived radiomic signature and the clinical parameters can be used for evaluating the EGFR-activating mutation status. This may have important clinical influence, notably by allowing the better personalization of target therapy for NSCLC patients with EGFR-activating mutations.

Dataset
Our study was approved by the institutional review board of Lishui Hospital of Zhejiang University. Because of its retrospective nature, requirement for informed consent was waived. Patients who were diagnosed with pathologically confirmed NSCLC from June 30, 2015, to January 18, 2018, were enrolled. A total of 590 were included according to the following inclusion criteria (1): CT imaging performed within one month before surgery (2); histological diagnosis of NSCLC (3); EGFR mutations (EGFR EXON18 G719X、EGFR EXON19 19-Del、EGFR EXON20 T790M、EGFR EXON20 20-Ins、 EGFR EXON20 S768I、EGFR EXON21 L858R、EGFR EXON21 L861Q) detected by amplification refractory mutation system-Scorpion real-time PCR (ARMS-PCR); and (4) clinical data were available. Thereafter, 289 patients were excluded according to the following exclusion criteria (1): preoperative treatment at the time of the initial diagnosis (n=96) (2); tissue sample obtained by biopsy rather than surgery (n=138); and (3) histological diagnosis of SCLC (n=55). Eventually, a total of 301 patients were enrolled in our study; 210 patients and 91 patients were allocated to the training and validation cohorts, respectively with a ratio of 7:3 (16).

CT Image Acquisition and Interpretation
Patients underwent preoperative unenhanced CT scanning using a 64-channel Philips Brilliance CT system (Philips Medical Systems). Details regarding the acquisition parameters were set as follows: tube current, 200 mA; tube voltage, 120 kV; slice thickness, 0.9 mm; collimation width, 40 mm (64 × 0.625 mm); reconstruction interval with iDose3 hybrid iterative reconstruction algorithm, 0.45 mm; scan field of view (SFOV), 15-20 cm; pitch, 1.2; rotation time, 350 ms; and pixel matrix size, 1024×1024. The images were processed in the Extended Brilliance Workspace (EBW, Philips). Multi-planar reconstruction was used for image reconstruction with a thickness of 5 mm.

Tumor Segmentation and Radiomic Feature Extraction
CT images of selected patients were exported from the picture archiving and communication system (PACS) according to the inclusion and exclusion criteria. ITK-SNAP software (version 3.4.0, www.itk-snap.org) was used for three-dimensional semiautomatic segmentation (18). All images were automatically segmented and adjusted by a radiologist with 18 years of experience (Z.W., reader 1), who repeated the same procedure within 2 weeks. The interobserver reproducibility of each segmentation was evaluated by another radiologist with 20 years of clinical experience (J.J., reader 2).

Inter-and Intraobserver Reproducibility
The inter-and intraobserver reproducibility of semantic image features, tumor segmentation and feature extraction were evaluated by intraclass correlation coefficients (ICCs). Two radiologists specialized in chest CT interpretation initially analyzed the images obtained from 30 randomly selected patients within 2 weeks in a blinded fashion. ICCs greater than 0.75 were considered as good consistency, and the remaining image segmentation was performed by reader 1.

Radiomic Feature-Based Prediction Model Construction
We built the radiomic signature model based on selected features from the training cohort. Z-score was applied to feature normalization before feature selection. Two feature selection methods, maximum relevance minimum redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO), were used to select the features. First, mRMR was performed to eliminate redundant and irrelevant features. LASSO was used to select the most useful features by penalty parameter tuning and 10-fold cross-validation based on the minimum criteria. LASSO includes choosing the regular parameter l to determine the number of features. After the number of features was determined, the most predictive subset of features was chosen, and the corresponding coefficients were evaluated. The coefficients for most radiomic features were reduced to zero, and any remaining radiomic features with non-zero coefficients were selected. Next, we built a model with selected radiomic features. A radiomic score (Radscore) was computed for each patient through a linear combination of selected features weighted by their respective coefficients. The final formula for the Radscore was as follows: "Radscore = -0.152*Small Area Emphasis + -0.097*Long Run High Grey Level Emphasis_ angle0_offset4 + 0.035*Cluster Prominence _All Direction_ offset7_SD + 0.082*Inverse Difference Moment_All Direction _offset4_SD + 0*Low Grey Level Run Emphasis_All Direction_offset4_SD + -0.064*Long Run Low Grey Level Emphasis_All Direction_offset7_SD + 0.275*Correlation_ angle0_offset7 + 0.211*std Deviation + -0.068*GLCM Energy_All Direction_offset4_SD + -0.018". Furthermore, the Radscore was compared between the wild-type EGFR and EGFR-activating mutations in both the training and validation cohorts.
Logistic regression with L1 regularization was performed to select the independent clinical predictors in the training cohort. Prediction models combining radiomic features and clinical variables were established. We built a radiomic nomogram based on the multivariate logistic regression model in the training cohort, and receiver operating characteristic (ROC) curves were developed to evaluate the discriminatory ability of the nomogram. The calibration curve of the nomogram was used to assess how closely the nomogram predicted EGFR-activating mutations relative to the actual probability (20,21). The Hosmer-Lemeshow test was used to evaluate the goodness-offit of the calibration curve (22). In addition, decision curve analysis (DCA) was used to determine the clinical usefulness of the prediction model by quantifying the net benefits at different threshold probabilities. DCA estimates the net benefit of a model through the difference between the true-positive and falsepositive rates, weighted by the odds of the selected threshold probability of risk (23).

Statistical Analysis
Statistical analysis was performed using R software (version 3.3) for quantitative feature analysis. The characteristic features of patients with EGFR-activating mutations and wild-type EGFR were compared by Student's t-test for normally distributed data; otherwise, the Mann-Whitney U test was used. Multivariate binary logistic regression was performed with the "rms" package. A nomogram was established by incorporating significant characteristic features and radiomic features. ROC curves were plotted to evaluate the diagnostic efficiency of the model. The area under the ROC curve (AUC) was then calculated. The nomogram was constructed and the calibration plots were created using the "rms" package. A p-value <0.05 was considered significant.

Radiomic Signature Construction, Validation, and Evaluation
A total of 396 radiomic features were extracted from unenhanced CT images. The intraobserver ICCs ranged from 0.80 to 0.89, and the interobserver ICCs ranged from 0.76 to 0.90, indicating satisfactory intra-and interobserver feature extraction reproducibility. In all, 20 features were retained after the mRMR algorithm was applied. Then, LASSO was performed, including selection of the regular parameter l (log l=0.03), to determine the number of features (Figures 1A, B). After the number of features was determined, the most predictive subset of 9 features was chosen (Supplementary Table 1), and the corresponding coefficients were evaluated ( Figure 1C) and used to build a prediction model. The Radscore showed a significant difference between NSCLC patients with wild-type EGFR and EGFR-activating mutations in the training (P<0.0001) and validation cohorts (P=0.0061). Patients with EGFR-activating mutations generally showed a higher Radscore ( Figure 2).
As shown in Figure 3, the radiomic feature only model achieved an AUC of 0.70 in the training cohort and 0.67 in the validation cohort. We incorporated the clinical indicators with P values less than 0.01 and the radiomic features into the logistic    Table 2). The joint model yielded an AUC of 0.81 (95% CI, 0.75-0.87) with a sensitivity of 84% in the training cohort ( Figure 3A) and an AUC of 0.75 (95% CI, 0.65-0.86) with a sensitivity of 76% in the validation cohort ( Figure 3B), which showed an improved performance over the radiomic signature in both the training and validation cohorts. Table 2 lists the predictive performance of the joint model, using the AUC, accuracy, sensitivity and specificity as the main measurements. The joint model outperformed the radiomic feature model and the clinical characteristics-based model in terms of sensitivity in the training and validation cohorts.
Subsequently, a nomogram integrating smoking status, spiculation, air bronchogram, CEA, SCCA and Radscore was constructed, as presented in Figure 4A. The calibration curve of the nomogram for the prediction of EGFR-activating mutations demonstrated favorable agreement between estimation with the radiomic nomogram and actual observations. The p value obtained via the Hosmer-Lemeshow test for the predictive ability of the nomogram was 0.57 in the training cohort ( Figure 4B) and 0.24 in the validation cohort ( Figure 4C).
DCA for the prediction model showed that the joint nomogram had the highest net benefit compared with the clinical and radiomic feature models across the majority of the range of reasonable threshold probabilities ( Figure 4D). The decision curve showed that if the threshold probability of a patient was within the range from 10% to 65%, using the joint nomogram developed in our study to predict EGFR-activating mutations added more benefit than the treat-all-patients scheme or the treat-no-patients scheme.

DISCUSSION
We undertook this study to develop and validate a joint modelbased nomogram for the preoperative individualized prediction of EGFR-activating mutations in NSCLC patients. The nomogram integrated 5 clinical features, i.e., smoking status, spiculation, air bronchogram, CEA, and SCCA, and 9 radiomic features. Our findings suggest that NSCLC patients could be classified as having EGFR-activating mutations or wild-type EGFR according to our nomogram, indicating that the nomogram could be used as a novel and user-friendly instrument for the better management of NSCLC patients. Moreover, this study provides a visualized explanation to help clinicians understand the prediction outcomes in terms of CT data. Diagnosis of the EGFR mutational status on an individual basis is vital for defining personalized treatment strategies. EGFR mutation including the sensitivity (EGFR Del19 and L858R) and resistance mutation (EGFR T790M) to TKIs. Recently, researchers have been seeking novel approaches to replace or complement conventional molecular analysis in routine CT examinations. Wang et al. proposed an end-to-end deep learning model to predict the EGFR mutation status by preoperational CT scanning, with an AUC of 0.85 in a primary cohort (24). However, the developed model can only be used to distinguish patients with wild-type EGFR and EGFR mutations and cannot identify whether mutations are EGFR activating or drug resistant mutations. In addition, although the deep learning method is labor-saving since it does not require precise nodule segmentation (25), the accuracy of segmentation is controversial. Liu et al. collected 289 patients with surgically resected peripheral lung adenocarcinomas and extracted 219 radiomic features to predict the EGFR mutation status, with an AUC of 0.709 (26). The prediction model in our study, with an AUC of 0.81 in the training cohort, is more reliable and can be used for discriminating wild-type EGFR and EGFR-activating mutations to guide targeted therapy.
Although smoking has been well established as the major cause of lung cancer, EGFR mutations have proved to be the most common genetic alteration in never-smoking NSCLC patients. A meta-analysis performed by Ren et al. revealed that non-smoking was associated with a significantly higher EGFR mutation rate. The frequency of EGFR mutations ranged from 22.7% to as high as 72.1% in never-smokers (27). Our results are in line with those of a previous study in that the presence of EGFR mutations was closely associated with the never-smoking status in NSCLC patients (28).
The relevance of CT features to the EGFR mutation status has also been reported recently. Spiculated margins, subsolid density, and non-smoking were confirmed to be significantly associated with EGFR-activating mutations (29). Zhou et al. found that spiculated margins, pleural retraction, and air bronchogram were more frequent in the EGFR mutation group than in the wild-type group, but there was no significant difference between these groups (30). On the other hand, air bronchogram was reported as an indicator of EGFR mutations in NSCLC (31). This result is consistent with Liu's findings, which revealed a significant correlation between a small lesion size and air bronchogram with EGFR mutations in lung adenocarcinoma (32).
Serum tumor markers, such as CEA, SCCA, CYFRA 21-1, NSE, and ProGRP, are considered to be predictive or prognostic in NSCLC, and some of these markers have been shown to be correlated with EGFR mutations (33). CEA is widely known as a serum tumor marker of NSCLC (34,35). It has also been uncovered that the serum CEA level in Chinese patients is not only positively associated with EGFR mutation but also negatively associated with the efficacy of TKI therapy (36). These findings raise the question of whether there is any correlation between the serum CEA level and EGFR mutations. In our study, the CEA level (below or above 5 ng/mL) served as an independent marker for predicting EGFR-activating mutations in NSCLC patients. Consistent with a previous report, an elevated serum CEA level predicted the presence of EGFR mutations in pulmonary adenocarcinoma (37). The low frequency of an elevated SCCA level has been reported in EGFR-mutated NSCLC, but no further evidence has been presented regarding the relation between SCCA and EGFR-activating mutations (38,39). In our study, patients with a normal SCCA level showed higher scores, suggesting that this factor may contribute to the increased possibility of EGFRactivating mutations.
With the radiomic approach, we identified that 9 radiomic features from 4 different feature categories (GLCM, histogram, RLM, GLZSM) were significantly associated with EGFRactivating mutations and could serve as indicators for the prediction of EGFR-activating mutations. The AUC of the  radiomic feature model was lower than that of the joint model (P=0.0005), suggesting that the radiomic features helped improve the performance of the joint model, as indicated by the higher AUC. These findings suggest that models integrating radiomic features with clinical features are more effective. DCA demonstrated that the joint nomogram was superior to both the clinical feature model and the radiomic model across the majority of the range of reasonable threshold probabilities, which also indicates that the radiomic signature added value to the traditional clinical features used for individualized EGFRactivating mutation estimation. Therefore, a non-smoking patient presenting with an abnormal serum CEA level, a normal SCCA level, spiculation, air bronchogram and a high Radscore might be more likely to have EGFR-activating mutations.
This study has several limitations. First, this was a retrospective study and thus may have selection bias. Second, tumor segmentation was performed by a semi-automatic process, which was time consuming for the radiologists. However, the results are more robust, especially for tumors with unclear margins. Third, different CT scanning devices with different acquisition protocols were used. Thus, multicenter validation need to be performed to prove nomogram reliability.
In conclusion, we established a CT image-based model combining radiomic features and clinical variables for the prediction of EGFR-activating mutations before initial treatment in patients with NSCLC. The radiomic feature-based nomogram can serve as an alternative approach to determine better candidates for first-generation EGFR TKI therapy.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by institutional review board of Lishui Hospital of Zhejiang University. The patients/participants provided their written informed consent to participate in this study.

AUTHOR CONTRIBUTIONS
QW and JJ designed the research. MX, WZ, JT and WM helped collect patient information. JuH, HW, and CLu performed experiments. JiH, CZ, PP, and LZ analyzed data. SF, MC, CLu, and YB prepared figures and tables. QW and JuH wrote the paper. HY and JJ conceived the project and supervised and coordinated all aspects of the work. All authors contributed to the article and approved the submitted version.

FUNDING
This study was supported by Zhejiang Medical and Health Science Project (2019RC320 to QW, 2020KY1080 to HW, 2018KY197 to LZ, 2018KY932 to SF, 2018KY183 to YB), Natural Science Foundation of Zhejiang Province (LY18H160059 to JSH, LYY19H310004 to YB), The Public Welfare Project of Zhejiang Province (LGF18H160035 to HW), and the Science and Technology Project of Lishui City (2020ZDYF09 to QW).