The Potential of Radiomics Nomogram in Non-invasively Prediction of Epidermal Growth Factor Receptor Mutation Status and Subtypes in Lung Adenocarcinoma

Purpose: Up to 50% of Asian patients with NSCLC have EGFR gene mutations, indicating that selecting eligible patients for EGFR-TKIs treatments is clinically important. The aim of the study is to develop and validate radiomics-based nomograms, integrating radiomics, CT features and clinical characteristics, to non-invasively predict EGFR mutation status and subtypes. Materials and Methods: We included 637 patients with lung adenocarcinomas, who performed the EGFR mutations analysis in the current study. The whole dataset was randomly split into a training dataset (n = 322) and validation dataset (n = 315). A sub-dataset of EGFR-mutant lesions (EGFR mutation in exon 19 and in exon 21) was used to explore the capability of radiomic features for predicting EGFR mutation subtypes. Four hundred seventy-five radiomic features were extracted and a radiomics sore (R-score) was constructed by using the least absolute shrinkage and selection operator (LASSO) regression in the training dataset. A radiomics-based nomogram, incorporating clinical characteristics, CT features and R-score was developed in the training dataset and evaluated in the validation dataset. Results: The constructed R-scores achieved promising performance on predicting EGFR mutation status and subtypes, with AUCs of 0.694 and 0.708 in two validation datasets, respectively. Moreover, the constructed radiomics-based nomograms excelled the R-scores, clinical, CT features alone in terms of predicting EGFR mutation status and subtypes, with AUCs of 0.734 and 0.757 in two validation datasets, respectively. Conclusions: Radiomics-based nomogram, incorporating clinical characteristics, CT features and radiomic features, can non-invasively and efficiently predict the EGFR mutation status and thus potentially fulfill the ultimate purpose of precision medicine. The methodology is a possible promising strategy to predict EGFR mutation subtypes, providing the support of clinical treatment scenario.


INTRODUCTION
Lung cancer is the leading cause cancer-related death both in male and female (1). Non-small cell lung cancer (NSCLC) accounts for more than 80% of lung cancers, of which lung adenocarcinoma is the most common histological subtype (2). With the advances of genomics, molecular-targeted therapy like using tyrosine kinase inhibitors (TKIs), which targets the epidermal growth factor receptor (EGFR) mutations, is recommended as first-line system therapy before first-line therapy by National Comprehensive Cancer Network (NCCN) for patients with advanced EGFR-mutant NSCLC (2) and proved to substantially improve the progression-free survival (PFS) compared with conventional chemotherapy (3,4). Up to 50% of Asian patients with NSCLC have EGFR gene mutations (5), indicating that selecting eligible patients for EGFR-TKIs treatments is clinically important. In patients with NSCLC, the most commonly found EGFR mutations are deletions in exon 19 (45%) and in exon 21 (L858R in 40%) in patients with EGFR mutations (2). Both mutations are associated with sensitivity to the small molecule TKIs as well as erlotinib, gefitinib, afatinib, and osimertinib (2), however, with different survival outcomes in response to both EGFR-TKIs and chemotherapy (6). Therefore, identifying EGFR mutation subtypes, especially those responsive to TKI treatment, seems to be more critically and scientifically important than just predicting EGFR mutation status.
In this context, though more and more research has emerged on the non-invasive prediction of EGFR mutation status in recent years (7)(8)(9), no predictors are recommended for selecting patients in clinical decision-making. Moreover, substantial discrepancies are presented to date with regarding to some features, especially semantic features derived from medical images (10). Buoyed by the availability of big data and state-of-art data analysis strategy, such as radiomics and deep learning, decoding tumor phenotype to precisely predict genotype has becoming the point of attention (11). Several studies have investigated the potential ability of radiomics to non-invasively predicting EGFR mutation status and show promising results (12)(13)(14)(15). Few results are finally applied in clinical practice yet due to the complicated procedure (e.g., time consuming, poor reproducibility, remaining the operator-dependency that is not biases-free, and so on) of Abbreviations: EGFR, epidermal growth factor receptor; NSCLC, non-small cell lung carcinoma; PFS, progression-free survival; ALK, antigen anaplastic lymphoma kinase; ROS1, c-ros oncogene 1; TKIs, tyrosine kinase inhibitors. radiomic researches (16). In view of this, models that giving an individual numerical probability of a clinical event (e.g., nomogram) rather than a predicting accuracy, may be more suitable and convenient for clinical application.
In the current study, we aim to build radiomics-based nomograms, integrating radiomics, CT imaging features and clinical characteristics, to non-invasively predict EGFR mutation status and subtypes (exon 19 and 21mutation).

Patient Selection and Dataset Preparation
This retrospective study was approved by the institutional review board (No. 20170103), which waived the requirement for patients' informed consent referring to the CIOMS guideline. The flowchart of our study was showed in Figure 1.
A search of Picture Achieving and Communication System (PACS) and pathological system from January 2013 to June 2018 was performed by one author with the following inclusion criteria: (1) Available thin-slice chest CT (<1.5 mm) images before biopsies or surgical treatment; (2) Available detailed histological reports of adenocarcinoma; (3) Available detailed EGFR mutations testing reports; (4) No any prior treatment before EGFR mutations analysis.
Finally, 637 patients were included. Of the 637 lesions, 342 lesions tested positively for an EGFR mutation (EGFR Mut), where 295 lesions were classified as wild-type lung adenocarcinomas (EGFR WT). Note that only one malignant nodule was studied for each patient due to the availability of EGFR testing report. Among the 342 patients with EGFR Mut, 130 patients were detected an EGFR Mut in exon 19, whereas 190 patients were detected an EGFR Mut in exon 21. The clinical and histopathologic variables, including age, sex, smoking status, tumor size, tumor location, histological subtypes etc. were presented in Table 1.
Two tasks were investigated in the current study: task (a), differentiating EGFR Mut from EGFR WT; task (b), differentiating EGFR Mut in exon 19 from in exon 21. Each dataset in two tasks (n = 637 and n = 320) was split into 10 groups (1-10), each subset was randomly selected by choosing 10% of each of the 2 categories. In task (a), groups 1-5 were defined as training dataset, the rest groups were defined as validation dataset. In task (b), considering the insufficient training data, groups 1-6 were defined as training dataset, the rest groups were defined as validation dataset. Note that we only included EGFR mutations in exon 19 and 21 in task b to avoid the sparse training data and the disbalance of data distribution. Mover, constructing a model to predict these two exons in patients with EGFR mutation is clinically reasonable (see Introduction).

CT Acquisition and Imaging Interpretation
All included patients were performed with the following six scanners: GE Discovery CT750 HD, 64-slice LightSpeed VCT, Revolution CT (GE Medical Systems); Somatom Definition flash, Somatom Sensation-16, Somatom Force (Siemens Medical Solutions). The acquisition parameters were as follows: 120 kVp; 100-200 mAs; pitch, 0.75-1.5; and collimation, 1-1.5 mm, respectively. All imaging data were reconstructed by using a medium sharp reconstruction algorithm with a thickness of 0.75-1.5 mm. CT images were acquired in the supine position at full inspiration for all patients. Only plain CT images were used in the current study. Two radiologists (with 12 and 3 years of experience in chest CT interpretation) independently interpreted the CT images, blinded to clinical and histologic findings. Thirteen CT features ( Table 1) were evaluated. The definitions of these features were described in previous study (17,18). A re-evaluation for achieving a consensus was performed to solve the disagreement between two radiologists.

Segmentation and Radiomic Features Extraction
All nodules were manually delineated slice by slice using a medical image processing and navigation software 3D Slicer (version 4.8.0, Brigham and Women's Hospital) by one author (with 5 years of experience in chest CT interpretation), then the volume of interests (VOIs) were confirmed by another radiologist (with 12 years of experience in chest CT interpretation). Fifty randomly selected nodules were independently segmented by the two authors for feature reproductive analysis. Images and VOIs with NII format were exported for further analysis. Radiomic features from three categories, including 50 gray-level histogram features, 325 gray-level co-occurrence matrix (GLCM) features, and 100 gray-level run lengths matrix (GLRLM) features, were extracted using Matlab 2016b (MathWorks, Natick, USA). The details of extracted radiomic features were presented in Supplementary Data. Radiomic feature extraction methodology was described in our previous study (18).

Features Selection and Radiomic Score Construction
The least absolute shrinkage and selection operator (LASSO) method, which is an accepted algorithm for feature selection in high-dimensional variables (19), was applied to select the features that were most distinguishable and build a logistic regression model in training dataset. Then a radiomic score (R-score) was calculated for each lesion using features selected by LASSO and weighted by the respective coefficients.

Nomogram Construction and Validation
Univariate analysis was firstly to identify the potential predictors among clinical characteristics, CT features and R-score. Factors that associated with EGFR mutation status and subtypes were then included to multivariate analysis to identify the independent predictors. Furthermore, the identified independent factors were selected to construct the final nomogram in the training dataset.

Histologic Evaluation and EGFR Mutation Analysis
The included lung adenocarcinomas were categorized according to the 2011 IASLC/ATS/ERS classification system (20) (drug target-associated). Molecular analysis of mutation status of EGFR exons 18-21 was examined using a PCRbased amplification-refractory mutation system (ARMS) by AmoyDx company.

Statistical Analysis
Statistical analysis was performed using R software (version 3.4.2; http://www.Rproject.org). The mean values and standard deviations were expressed for continuous variables (age, lesions size), and frequency or percentage for categorical variables. The Wilcoxon rank sum test and the x 2 -test were used to compare medians and proportions between two groups, respectively. Predictive performance was evaluated by the area under the curve (AUC) of the receiver operator characteristic (ROC). DeLong test was used to evaluate the difference of the ROCs between various models (21). Hosmer-Lemeshow test was used to evaluate the goodness-of-fit of the constructed nomogram. ICC analysis was performed using "irr" package, features with an ICC> 075 were consider as robust features. Lasso logistic regression was done using the "glmnet" package. Multivariate logistic regression, nomograms and calibration plots were done with the "rms" package. A P<0.05 indicated a significant difference.

Associations Between Clinical, CT Features and EGFR Mutation Status and Subtypes
Associations between clinical, CT features and EGFR mutation status and subtypes were presented in Tables 2, 3. The incidence of harboring EGFR mutation was significantly higher in female than male in two datasets (P = 0.002, P = 0.013, respectively). Patients with EGFR Mut had a higher age in the current study.
Smoking was not a significant factor to differentiate EGFR Mut lesions from EGFR WT lesions. In terms of radiographic features, 9 features, including size, margin, shape, pleural retraction, bronchiole change, lobulation, speculation, peripheral emphysema, peripheral fibrosis were significantly associated with EGFR mutation status. Patients with EGFR mutations in 21 exon had a higher age than those with EGFR mutations in 19 exon in two datasets (P = 0.013, P = 0.003, respectively). No other clinical and CT features were identified as potential factors to predict the EGFR mutation subtypes.
In terms of the task of predicting the EGFR mutation subtypes (task b), 32 features were finally left after performing the LASSO analysis (Supplementary Data). A R-score * was also calculated for each lesion by using the formula presented in the Supplementary Data. Lung adenocarcinomas with EGFR Mut in exon 19 had a lower R-score * than ones with EGFR Mut in exon 21 in training dataset (−0.39 ± 1.35 vs. 0.27 ± 1.12, P   = 0.000) and in validation dataset (−0.70 ± 1.25 vs. 0.23 ± 1.12, P = 0.000) ( Table 3). The proposed R-score * demonstrated a good performance in differentiating EGFR mutation subtypes with AUCs of 0.684, 0.708 in two datasets (Figures 4A,B). The Hosmer-Lemeshow test for R-score * yielded a non-significant statistic in the training and validation set (P = 0.295, P = 0.242, respectively), which suggested that there was no departure from a perfect fit.

Development and Validation of the Radiomics Nomograms for Predicting the EGFR Mutation Status and Subtypes
There was no multicollinearity between the significant factors identified by univariate analysis and R-score. After performing the multivariate analysis, sex, peripheral emphysema and R-score were identified as independent prognostic factors of harboring EGFR mutation ( Table 4) and subsequently incorporated to develop the radiomics-based nomogram ( Figure 3C). The constructed nomogram obtained a significantly incremental performance for predicting EGFR mutation status compared with that of sex and peripheral emphysema ( Table 5). The Hosmer-Lemeshow test for the nomogram yielded a non-significant statistic in the training and validation set (P = 0.313, P = 0.816, respectively), which suggested that there was no departure from a perfect fit. Though the nomogram significantly outperformed Rscore in training dataset, the statistically difference was not found in validation dataset ( Table 5).
In terms of task b, age and R-score * were identified as independent prognostic factors of predicting EGFR mutation subtypes ( Table 4) and subsequently incorporated to develop the radiomics-based nomogram * (Figure 4C). The Hosmer-Lemeshow test for the nomogram * yielded a non-significant statistic in the training and validation set (P = 0.760, P = 0.413, respectively), indicating that there was no departure from a perfect fit. Not surprisingly, the constructed nomogram * significantly outperformed age and potentially outperformed Rscore * in predicting EGFR mutation subtypes ( Table 5).

DISCUSSION
Non-invasively and preoperatively predicting the EGFR mutation status, a field attracted continuous efforts of researchers, can overcome the disadvantages of molecular assays (e.g., high cost, intratumoral heterogeneity, poor sample quality) well and furtherly help clinicians to select the eligible patients for targeted therapy. Moreover, attempting to predict the EGFR mutation subtypes, especially those are sensitivity to small TKIs, may provide important information for making finer treatment scenario. In the current study, we developed and validated two radiomics based nomograms, incorporating clinical characteristics, CT features and radiomics, to predict the EGFR mutation status and subtypes with promising performance (AUC = 0.734, AUC = 0.757, respectively).
It is well-documented that the EGFR Mut rate are significantly higher in female than male (22), which is also confirmed in our study (P = 0.002 and P = 0.013 in two datasets, respectively). Note that smoking status, another potential clinical factor verified by previous studies (8,14,17), failed to show significant association with EGFR mutation status in the current study. Note that smoking status is still a contentious risk factor (23) and not recommended to use as the criteria for selecting eligible patients (2). Another reason why the difference between EGFR Mut and WT in terms of smoking status is diminished may be that the incidence of lung adenocarcinomas in female was higher than male (20,24). Moreover, female may be less likely to be current/former smoker than male. Wu et al. (25) had reported that younger patients (<50 years old) with lung adenocarcinoma had lower EGFR mutation rate, which was not verified in multivariate logistic analysis in the current study. The discrepancy may be contributed to the data bias (ours vs. Wu's: 20.9 vs. 15.5%).
Many prior studies have investigated the role of morphological features in predicting EGFR mutation status. The results remain controversial. On one hand, most radiographic features are non-quantitative subjective features and susceptible to the discrepancy of evaluation caused by different knowledge structure of observers. On the other hand, obvious radiographical features are more frequently presented in advanced tumors instead of early-stage tumors. For example, stage I lung cancers account for 60-70% of detected lung cancers in screening programs (26). In our study, 60.9% of included patients were stage I lung cancers. In this context, the differentiating performance of these semantic features may be compromised.
Peripheral emphysema was the only independent risk factors for predicting EGFR mutation status in the current study.
Another disadvantage of semantic features is that they only reflect few tumor information in biological level. By contrast, radiomics method can encode a more comprehensive level of feature abstraction and thus potentially provide better prediction performance. Previous studies have revealed the potential associations between these engineered features and EGFR mutation status (12,23,27) and proved that the performance of models can benefit from the integration of radiomics and clinical features (14,23). These results were also confirmed in our study. Despite the promising results, the complicated process of feature extraction and the inconvenience of formula-based model limit its application in clinical context. Hence, an easy-touse way for radiomics method is urgently needed. Incorporating multiple radiomic features into a radiomics score can tactfully make multi-marker analyses less complicated to use (18,28), which is similar to the construction of multi-factor panels. In this study, the LASSO regression model was used to select and compress the radiomic features and thus construct the Rscore. The constructed R-score outperformed the clinical and CT features in predicting EGFR mutation status ( Table 5). Another strategy to make radiomics method friendly use is presenting the results in a more intuitive way, such as the nomogram (29,30). Nomogram is a statistic model that can provide an individual numerical probability of a clinical event by integrating multiple variables (31,32). To comprehensively investigate the predictive performance of non-invasively available variables, including clinical factors, imaging semantic features, imaging radiomic features, we adopted the two above-mentioned strategies to construct a radiomics-based nomogram for predicting the EGFR mutation status. The constructed nomogram in the current study was conveniently used to individualized predict the probability of harboring EGFR mutation by calculation the respective points of three variables (sex, peripheral emphysema, and R-score), with a promising AUC of 0.734. Another promising technique (i.e., liquid biopsy) was consider as an alternative to test EGFR mutations. However, the high false-negative rate (30%) need to be further resolved (33).
In patients with NSCLC, the most commonly found EGFR mutations are deletions in exon 19 (Exon19del in 45%) and in exon 21 (L858R in 40%) (2). Both mutations result in activation of the tyrosine kinase domain, and both are associated with sensitivity to the small molecule TKIs. Nevertheless, patients with exon 19 deletion are associated with longer PFS compared to those with L858R mutation after first-line EGFR-TKIs (34,35). In addition, the incidence of T790M mutation, which is associated with acquired resistance to reversible EGFR-TKIs (36), might be different between exon 19 deletions and L858R mutations (6), resulting in different treatment scenarios. As a result, furtherly predicting the specific EGFR mutation subtypes may be clinically important. Inspired by the satisfied results of predicting EGFR mutation status, the potential relations between  clinical factors, imaging semantic features, imaging radiomic features, and EGFR mutation subtypes were also investigated in our study. Age was the only independent factors that can be used to differentiate different EGFR mutation subtypes.
No CT features were identified as potential predictors, which is consistent with previous studies (37,38), indicating the difficult of predicting EGFR mutation subtypes through semantic features (relatively low-level). This conclusion encouraged us to investigate whether the radiomic features (relatively high-level) might be competent on this task. Surprisingly, the constructed Rscore * was identified as an independent predictor and obtained a good performance with an AUC of 0.708. A radiomicsbased nomogram * , incorporating age and R-score * , achieved a better performance (AUC = 0.757), indicating the efficiency of radiomic features in different medical tasks.
We may conclude that radiomics outperformed the clinical and semantic features in both tasks (see Table 5). The diagnostic benefits may due to the possession of more presentative and discriminative information of radiomics, which could reflect the tumor spatial heterogeneity, tumor microenvironment, as well as tumor gene patterns. Yet, completely interpreting the association between radiomics and the complex biological processes (EGFR mutation status in the current study) remains an intractable challenge. The interpretability warrants further investigation. Please note that one of the implications of our study was to introduce the nomogram, which was easily used in clinical practice and may facilitate the clinical transformation of radiomics researches.
There are several limitations should be noted. First, although the imaging normalization and reproductive analysis can mitigate the influence of radiomic feature variabilities, it cannot make the current study completely immune to the potential confounding variability caused by different CT scanning parameters (39,40). Paradoxically, studies with homogeneous images must sacrifice the amount of data. How to balance it well is difficult for each radiomics-based research. Second, this was a single-center study and lacked of external validation, indicating the potential data selection bias. Conducting a multi-center study and validating the constructed model in an independent external dataset may not only improve the generalization and robustness of a model efficiently, the models can also substantially benefit from diversified data from different regions, races and countries. Third, the current study narrowly focused on EGFR mutation status and subtypes. Constructing a nomogram which could cover all clinical relevant EGFR mutation and even other genetic mutations (e.g., ROS-1, ALK) may be interesting and worth investigating in the future work.

CONCLUSION
Radiomics-based nomogram, incorporating clinical characteristics, CT features and radiomic features, can noninvasively and efficiently predict the EGFR mutation status and thus potentially fulfill the ultimate purpose of precision medicine. The methodology is a possible strategy to predict EGFR mutation subtypes, providing the support of clinical treatment scenario.

DATA AVAILABILITY STATEMENT
The datasets generated for this study are available on request to the corresponding author.

ETHICS STATEMENT
This retrospective study was approved by the Institutional Review Board of our institution (No. 20170103), which waived the requirement for patients' informed consent referring to the CIOMS guideline.