Computed Tomography‐Based Radiomics Signature: A Potential Indicator of Epidermal Growth Factor Receptor Mutation in Pulmonary Adenocarcinoma Appearing as a Subsolid Nodule

Persistent subsolid nodules are known to represent early‐stage lung adenocarcinoma, and epidermal growth factor receptor (EGFR) gene mutation is commonly involved in the occurrence, development, and metastasis of tumors. The aim of this study was to develop and validate predictive models to differentiate EGFR mutations for subsolid lung adenocarcinoma.

Abstract

Background.

Lung adenocarcinoma (LADC) with epidermal growth factor receptor (EGFR) mutation is considered a subgroup of lung cancer sensitive to EGFR‐targeted tyrosine kinase inhibitors. We aimed to develop and validate a computed tomography (CT)‐based radiomics signature for prediction of EGFR mutation status in LADC appearing as a subsolid nodule.

Materials and Methods.

A total of 467 eligible patients were divided into training and validation cohorts (n = 306 and 161, respectively). Radiomics features were extracted from unenhanced CT images by using Pyradiomics. A CT‐based radiomics signature for distinguishing EGFR mutation status was constructed using the random forest (RF) method in the training cohort and then tested in the validation cohort. A combination of the radiomics signature with a clinical factors model was also constructed using the RF method. The performance of the model was evaluated using the area under the curve (AUC) of a receiver operating characteristic curve.

Results.

In this study, 64.2% (300/467) of the patients showed EGFR mutations. L858R mutation of exon 21 was the most common mutation type (185/301). We identified a CT‐based radiomics signature that successfully discriminated between EGFR positive and EGFR negative in the training cohort (AUC = 0.831) and the validation cohort (AUC = 0.789). The radiomics signature combined with the clinical factors model was not superior to the simple radiomics signature in the two cohorts (p > .05).

Conclusion.

As a noninvasive method, the CT‐based radiomics signature can be used to predict the EGFR mutation status of LADC appearing as a subsolid nodule.

Implications for Practice.

Lung adenocarcinoma (LADC) with epidermal growth factor receptor (EGFR) mutation is considered a subgroup of lung cancer that is sensitive to EGFR‐targeted tyrosine kinase inhibitors. However, some patients with inoperable subsolid LADC are unable to undergo tissue sampling by biopsy for molecular analysis in clinical practice. A computed tomography‐based radiomics signature may serve as a noninvasive biomarker to predict the EGFR mutation status of subsolid LADCs when mutational profiling is not available or possible.

Introduction

With the widespread use of multislice spiral computed tomography (CT) and low‐dose CT screening for lung cancer, single and multiple subsolid nodules are being detected more frequently than before in clinical practice. Persistent subsolid nodules, appearing as pure ground‐glass nodule (pGGN) and part‐solid nodule, are known to frequently represent early‐stage lung adenocarcinoma (LADC). At present, patients with subsolid nodules are evaluated and managed primarily according to the Fleischner Society and the National Comprehensive Cancer Network (NCCN) guidelines [1], [2]. However, some cases of highly suspicious and persistent subsolid nodules cannot be accurately diagnosed by using a small biopsy specimen or cannot be operated owing to the patient's health condition or the presence of multiple subsolid nodules, raising questions over the possibility of performing targeted adjuvant therapy for these patients.

Epidermal growth factor receptor (EGFR) gene mutation is one of the most common mutations involved in the occurrence, development, invasion, and metastasis of tumors. LADC with EGFR mutation is considered a subgroup of lung cancer that is sensitive to EGFR‐targeted tyrosine kinase inhibitors (TKIs) [3], [4], [5]. Although a molecular test is recommended in patients with advanced LADC, EGFR mutation testing may be performed in patients with early‐stage disease. Some studies have demonstrated the effective use of molecular evaluation of single and multiple subsolid lesions in the diagnosis and treatment of lung cancers, particularly those multiple subsolid nodules that typically have different driver gene mutations [6], [7], [8]. At present, the assessment of EGFR mutation status mainly depends on tumor tissue from surgery and biopsy in clinical practice. However, acquisition of adequate tissue samples through invasive biopsies for EGFR mutational analysis is not feasible at times, and a tissue sample is usually only a portion of a typically heterogeneous lesion and cannot completely represent the lesion. In addition, circulating cell‐free tumor DNA (ctDNA), which is released from tumor cells into the circulating blood, has been used in recent years to detect EGFR mutations in patients with non‐small cell lung cancer (NSCLC). However, the concordance rates between ctDNA and tumor tissues showed a big variation, ranging from 27.5% to 100% [9], and the technology of ctDNA for detecting EGFR mutations has not been promoted in developing countries and the cost of the technologies is high. Given these limitations, research into more convenient, economic, and less invasive techniques for detecting gene mutation status is urgently needed.

Because CT is recommended by the NCCN as the preferred imaging method for lung cancer in clinical practice, a few investigators have attempted to correlate the imaging features with EGFR mutations and showed that a ground‐glass nodule pattern, spiculated margin, and air bronchogram along with female sex and nonsmoking status are related to EGFR mutation [10], [11], [12]. However, these studies also showed limitations such as a relatively low accuracy and the use of semiqualitative radiological assessments that require substantially more work. Radiomics, which uses advanced image analysis algorithms in the artificial intelligence domain, makes it possible to reproducibly quantify the imaging phenotypes automatically by extracting a large number of image features and can reflect the voxel‐by‐voxel genetic information for complete, heterogeneous tumors [13], [14].

To the best of our knowledge, a few studies have attempted to develop CT‐based radiomics models that could predict EGFR mutations in lung cancer with relatively good performance [15], [16]. Nevertheless, there is little research on radiomics models for predicting EGFR mutations in subsolid LADCs. In order to provide a new noninvasive method for discriminating EGFR mutations in subsolid nodules, we hypothesized that radiomics features extracted from CT images could be used to predict the EGFR mutation status in cases of subsolid LADCs. Thus, the aim of this study was to develop and validate predictive models that can effectively differentiate EGFR mutations for subsolid LADCs.

Materials and Methods

Patient Selection

This retrospective study was approved by our institutional review board. A medical record review was performed in accordance with institutional ethics review board guidelines. We obtained the records of 2,932 consecutive patients who underwent surgical resection for primary lung cancer at our hospital from January 2016 to May 2018. CT images and clinical data of all cases were collected. The patient recruitment pathway is presented in supplemental online Figure 1 according to the inclusion and exclusion criteria.

The inclusion criteria were as follows: (a) classification of LADC histologic subtype according to the 2015 World Health Organization (WHO) classification of lung cancer; (b) examination of the mutation status of EGFR with a polymerase chain reaction‐based assay and confirmation through direct sequencing; (c) CT examination of the entire thorax using the same CT machine with the same algorithm (B30) and thickness (2 mm), within 2 weeks of surgery and before biopsy; (d) LADCs manifesting as persistent subsolid nodules on CT at 3–24 months of follow up; and (e) no previous chemotherapy, radiotherapy, or extrathoracic metastases before receiving CT examination.

The exclusion criteria were as follows: (a) presence of nodules that were proved to be of other histologic types, except adenocarcinoma; (b) presence of LADC that had not been examined using an EGFR mutation test; (c) CT imaging performed using different algorithms or thicknesses or on different CT machines; (d) LADCs as solid nodules on CT; and (e) cases wherein the radiomics features could not be accurately extracted for unknown reasons.

A total of 467 patients were included in this study on the basis of pathologic records showing subsolid LADCs with available EGFR mutation data and normalized CT scanning protocol findings. Baseline clinical and demographic data, including age, gender, smoking history, CT pattern, histopathological subtype, and EGFR mutation status, were derived from medical records. The CT pattern was evaluated according to the Fleischner Society statement [1]. A pGGN was defined as a focal nodular area of increased lung attenuation within visualized normal parenchymal structures, and a part‐solid nodule was defined as a combination of both ground‐glass and solid components, the latter obscuring the underlying lung architecture. Mutations in EGFR exons 18, 19, 20, and 21 were examined.

A total of 306 patients (120 male and 186 female; mean age, 56.75 ± 11.46 years; range, 17–82 years) treated from January 2016 to July 2017 were identified as the training cohort, whereas 161 consecutive patients treated from August 2017 to May 2018 (74 male and 87 female; mean age, 57.40 ± 10.41 years; range, 11–80 years) made up an independent validation cohort.

Histopathological Evaluation and Molecular Analysis

All resected specimens were formalin‐fixed and stained with H&E in accordance with the routine regulations of our hospital. The histopathological subtype of the LADC was classified according to the 2015 WHO classification of lung cancer [17]. Paraffin specimens of tumor tissue were evaluated by pathologists to ensure they met the criterion of containing at least 50% tumor cells. Patients with insufficient or poor‐quality tissue for molecular analyses were excluded from this study. The EGFR mutation status was examined with a polymerase chain reaction‐based assay and confirmed through direct sequencing. All detections were performed according to the manufacturer's protocol.

CT Image Acquisition

All patients underwent unenhanced CT of the entire thorax in a multidetector CT system (Definition AS+ 128‐Slice; Siemens Healthcare, Germany). In our study, all CT images were acquired using normalized protocols. CT scan parameters were as follows: tube voltage, 120 kV; automatic tube current modulation, 35–90 mAs; pitch, 0.9; field of view, 180 mm × 180 mm; matrix, 512 × 512; reconstructed algorithm, B30 or I30; and reconstructed slice thickness and slice increment, both 2 mm. All images were exported in DICOM format for image feature extraction.

Lesion Segmentation and Radiomics Feature Extraction

For nodule segmentation, we employed a 3D U‐net model [18], which has been used in our previous research [19]. Most of the primary tumors were automatically segmented, and only a few poorly contoured tumors were manually revised by our radiologists (X.Y. and Y.G.) with more than 10 years of experience in chest CT interpretation. For each accurately segmented tumor, feature extraction of the tumor region was performed automatically using programmed algorithms in Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/), an open‐source Python package for the extraction of radiomics features from clinical images [20]. A total of 1,063 radiomics features were extracted from unenhanced CT images. Radiomics features were classified into the following six categories: (a) first‐order features; (b) shape features; (c) gray‐level co‐occurrence matrix features; (d) gray‐level run‐length matrix features; (e) gray‐level size zone matrix features; (f) neighboring gray tone difference matrix features; and (g) gray‐level dependence matrix features.

Radiomics Feature Selection

To reduce redundant radiomic features, we used the mean decrease impurity importance, derived from random forest (RF), to identify highly informative features with respect to mutational status and remove unimportant or irrelevant radiomics features. Each radiomics feature was given an importance score (all feature importance scores sum up to 1). The features with an importance score higher than 0.01 were selected for use. Radiomics extraction and analysis workflow are shown in Figure 1.

Figure 1.

Flow chart of the study. (A): Segmentation for the tumor region of interest (ROI) on all computed tomography slices. (B): Extraction of features from the ROI, such as tumor shape, texture, and wavelet features. (C): Prediction for EGFR mutation.

Abbreviations: AUC, area under the curve; EGFR, epidermal growth factor receptor; ROC, receiver operating characteristic.

Statistical Analysis

We trained an RF classifier on the training cohort. RF classification has been reported as a robust and accurate supervised machine learning method for radiomics analysis. We also built another RF classifier based on clinical features, which included age, gender, smoking history, CT pattern, and subtype. Values above the threshold were considered abnormal, and those below the threshold were considered normal. To check the additive effect on classification, we concatenated the radiomics features and clinical features. The combined model was built using the RF method with the combined features. The performance of the classifier was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. To further evaluate classifier performance, we set the probability threshold as event (mutation) ratios to calculate sensitivity, specificity, positive predictive value and negative predictive value.

All model training and statistical tests were performed using Python 3.6. The “sklearn” package was used to implement RF model analysis and calculate AUCs. The ROC curves were plotted using the “matplotlib” package, and data distribution was plotted using the “seaborn” package. Clinical characteristics were analyzed between the EGFR‐positive and ‐negative (EGFR(+) and EGFR(−), respectively) groups by using Fisher's exact test or the chi‐square test for nominal categorical variables and the independent t test or Mann‐Whitney U test for continuous variables. AUCs of the models were compared using a two‐sided t test. A two‐sided p value of <.05 was considered significant.

Results

EGFR Mutation and CT Pattern of the Lesions

Of all the enrolled 467 patients with LADC, 300 (64.2%) showed EGFR mutations. Exon 21 missense (61.7%) and exon 19 missense (29.3%) were the most frequent types of mutations. Among these, the EGFR L858R mutation in exon 21 was detected in 185 patients (61.7%). No significant differences were detected between the mutation types in pGGNs and part‐solid nodules (χ² = 2.736, p = .603; Table 1).

Table 1. Epidermal growth factor receptor mutation type and lesion pattern on computed tomography.

Abbreviations: pGGN, pure ground‐glass nodule PSN, part‐solid nodule

Clinicopathologic Characteristics

Based on the results of EGFR mutation analysis, the patients were classified into two groups: EGFR(+) and EGFR(−). A total of 88 (18.8%) patients were former or present smokers at the time of their diagnosis. Patient baseline characteristics in the training and validation cohorts are listed in Table 2. The EGFR mutation rate was 61.1% (187/306) in the training cohort and 70.2% (113/161) in the validation cohort, with no significant differences between the two cohorts (p = .052). Within the two cohorts, the EGFR mutation groups showed no significant differences in gender and smoking status (p > .05), but they showed significant differences in age, tumor size, CT pattern, and subtype (p < 0.05). The EGFR(+) also appeared significantly more frequently in patients with part‐solid nodule.

Table 2. Characteristics of patients with EGFR mutations in the training and validation cohorts.

Abbreviations: AIS, adenocarcinoma in situ; APA, acinar predominant adenocarcinoma; CT, computed tomography; EGFR, epidermal growth factor receptor; IMA, invasive mucinous adenocarcinoma; LADC, lung adenocarcinoma; LPA, lepidic predominant adenocarcinoma; MIA, minimally invasive adenocarcinoma; MPA, micropapillary predominant adenocarcinoma; pGGN, pure ground‐glass nodule; PPA, papillary predominant adenocarcinoma; PSN, part‐solid nodule; SPA, solid predominant adenocarcinoma.

Selection of Radiomics Features

In total, 1,063 radiomics features were extracted from the region of interest. The RF algorithm was used for feature reduction and selection with EGFR mutations based on 306 patients in the training cohort (Fig. 2). We identified highly informative features associated with EGFR mutations. Forty‐three features with importance scores higher than 0.01 were selected for use in our model. The 43 features are shown in Figure 2, and the weighted coefficients of the selected features are listed in supplemental online Table 1.

Figure 2.

The heatmap shows the normalized mean difference in the selected 43 radiomics feature distributions among subsolid lung adenocarcinomas in the training cohort. Note that many features are significantly different between the EGFR(+) and EGFR(−) groups.

Abbreviation: EGFR, epidermal growth factor receptor.

Predictive Performance of the CT‐Based Radiomics Model, Clinical Model, and Combined Model

The prediction model for EGFR mutation based on radiomics features can differentiate EGFR mutation in subsolid LADCs. The ROC curves for the training and validation cohorts in the radiomics model are shown in Figure 3. The AUC for the training cohort of 306 patients was 0.831 (95% confidence interval [CI]: 0.786–0.876; Table 3). To validate the discrimination of the model, we also used a validation cohort consisting of 161 patients that yielded an AUC of 0.789 (95% CI: 0.713–0.865; Table 3). We also used the RF method to develop a clinical model based on clinical features such as age, gender, smoking status, and histological subtype. The AUC of the clinical model was 0.687 and 0.665 in the training and validation cohorts, respectively.

Figure 3.

ROC curves among different models. (A): Receiver operating characteristic (ROC) curves show differences between radiomics features, clinical variables, and combination of these two aspects in the training cohort. (B): ROC curves show differences between radiomics model, clinical model, and combined model of these two aspects in the validation cohort.

Abbreviation: AUC, area under the curve.

Table 3. Predictive performance of the radiomics model, clinical model, and combined model.

Abbreviations: AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value; SEN, sensitivity; SPE, specificity.

To illustrate the potential ability of the models to predict EGFR mutations, we combined the radiomics features and clinical features to develop a combined model, and compared the radiomics model, clinical model, and combined model (Table 3). The AUCs of the radiomics and combined models were superior to that of the clinical model (t test; p < .001). However, the simple radiomics model and combined model showed no significant differences in prediction ability (t test; p = .345 and p = .829 for the training and validation cohorts, respectively).

Discussion

In this study, we investigated whether CT‐based radiomics features could be used to predict EGFR mutation status in subsolid LADCs. The proposed CT‐based radiomics signature that we constructed using the RF algorithm was associated with EGFR mutations. It showed relatively high sensitivity and positive predictive value in the prediction of EGFR mutation, although it showed a relatively low negative predictive value in the training and validation cohorts. The performance of the combined model was consistent with those simple radiomics signature. Our study is the first to evaluate EGFR mutation status in subsolid LADCs using a radiomics signature in a large cohort, and the findings indicate that the CT radiomics signature may be useful to predict EGFR status of patients with LADC and has the potential to aid in the determination of therapeutic strategies of inoperable persistent single or multiple subsolid nodules.

Previous studies [10], [11], [12] have attempted to investigate the relationship between image characteristics and EGFR mutations in lung cancer. However, these studies generally had small sample sizes and they used subjective observer‐dependent imaging descriptors. A meta‐analysis [21] demonstrated that the most frequently used CT morphological features were part‐solid nodule pattern. CT‐based quantitative radiomics signatures have been used in predicting the gene phenotype in lung cancer. The study by Zhang et al. [15] showed that radiomics biomarkers constructed by a least absolute shrinkage and selection operator are better predictors of EGFR mutation status than conventional semantic CT image features or clinical variables, and the AUCs of the training and validation cohorts were 0.86 and 0.87, respectively. However, this study may be biased because of a relatively small sample of 180 patients, and this study focused on NSCLC, including adenocarcinoma (ADC) and squamous cell carcinoma (SCC). Another review [22] reported that the incidence of EGFR mutations in lung SCC was less than 5%. A previous study [23] reported differences in radiomics features between ADC and SCC and showed that a radiomics signature can be used as a diagnostic factor to discriminate ADC from SCC. Therefore, we think that these factors may cause a bias in the prediction results for lung cancer with EGFR mutations.

Rios Velazquez et al. [16] have investigated the associations between the radiomics feature phenotype and EGFR mutation in LADCs, and the results showed that a radiomics signature related to radiographic heterogeneity could successfully discriminate between EGFR(+) and EGFR(−) status in LADCs (AUC = 0.69). The combination of this signature with a clinical model of EGFR status (AUC = 0.70) significantly improved the prediction accuracy (AUC = 0.75). However, their CT images were acquired using scanners manufactured by different companies, with a range of image reconstruction algorithms, different slice thicknesses, and different dosages. A study by Clay et al. [24] showed that a quantitative CT analysis of a nodule and surrounding lung may noninvasively predict the presence of EGFR mutations (AUC = 0.77) in pulmonary nodules of the adenocarcinoma spectrum. This study had a small sample, including only 15 cases of EGFR(+) and 103 cases of EGFR(−). Persistent subsolid nodules are the most common manifestations of early lung adenocarcinoma, and some previous studies [25], [26], [27], [28] also showed that LADCs with EGFR mutations showed significantly higher frequencies of subsolid nodules than those without EGFR mutations. Therefore, we focused on the relationship between CT‐based radiomics features and EGFR mutations in LADCs with subsolid nodules in a relatively large cohort.

We constructed the radiomics models by using the RF algorithm, and this radiomics model showed good performance in predicting EGFR mutations with subsolid nodules (AUCs of 0.831 and 0.789 for the training and validation cohorts, respectively). Our result indicated relatively better performance than a few previously published studies [16], [29]. This result is logical because our study adopted a normalized scanning protocol (the same reconstruction algorithms and slice thicknesses). Clinical factors have been widely acknowledged as an important indicator of EGFR mutation status. However, there were no significant differences in gender and smoking status between EGFR(−) and EGFR(+) groups, and in contrast, the average age of the EGFR(−) group was lower than that of the EGFR(+) group. These differences may be due to case selection. We added clinical risk factors to the simple radiomics model to develop combined model. However, there were no significant differences in the radiomics model and combined model; this finding is different from those reported in previous studies [15], [16]. The difference may be attributable to the fact that the clinical feature distribution between EGFR (+) and EGFR (−) groups in our study was quite similar, making it difficult to distinguish two classes based on clinical features. In contrast, our radiomics model outperformed the clinical model, perhaps because the radiomics model involved the use of the most informative features. Thus, combining clinical features would not add any useful information to the radiomics model.

Although a recent study [30] showed that CT radiomics could stratify the prognosis of stage IV EGFR‐mutant NSCLC with EGFR‐TKI therapy, our study mainly focused on early‐stage resected LADC appearing as subsolid nodules, wherein the majority of patients had not undergone target EGFR‐TKI therapy after surgery. Hence, we did not study the relationship between radiomics and effectiveness of EGFR‐TKI therapy. In future research, we hope to improve the prediction accuracy of radiomics models by deep‐learning and external validation in order to apply the radiomics model to inoperable persistent solitary subsolid nodule or multiple subsolid nodules, and guide the clinician to decide on the use of targeted EGFR‐TKI therapy.

This study has several limitations. First, this was a retrospective study and was performed in a single large tertiary referral center in China, and there may have been a bias in patient selection. Second, the prevalence of EGFR mutation is known to be 20%–30% in Western countries and 50%–65% in East Asian countries. This study represents the prevalence of EGFR mutation (64.2%) in only Chinese patients; therefore, the results also lack universality, and we encourage more researchers to further study and validate them. Third, normalized CT scanning protocol limits the clinical applicability because radiomics features are affected by CT scanner parameters such as reconstruction kernel or section thickness. Some researchers [31], [32] have proposed methods to solve this issue, such as standard data collection, evaluation criteria, and image compensation. We encourage other researchers to validate our radiomics models using similar CT scanning protocols, such as using standard soft tissue algorithm and 2‐mm slice thickness in more institutions or hospitals. Fourth, we did not intend to build models dedicated to pGGN and part‐solid nodules for predicting gene mutation. For subgroup analysis, larger samples of subsolid LADC are needed to improve the accuracy of predictions.

Conclusion

In conclusion, we demonstrated an association between radiomics features and EGFR mutation in subsolid LADC. The radiomics model built in this study can be useful in differentiating EGFR mutation status with LADC appearing as inoperable persistent solitary or multiple subsolid nodules, and may guide clinicians to select appropriate treatments when mutational profiling is not available or possible.

See http://www.TheOncologist.com for supplemental material available online.

Author Contributions

Concept/design: Xinguan Yang, Yubao Guan

Provision of study material or patients: Nanshan Zhong, Yubao Guan

Collection and/or assembly of data: Xinguan Yang, Xiao Dong, Zhuoran Gu

Data analysis and interpretation: Xinguan Yang, Jiao Wang, Weiwei Li, Zhuoran Gu

Manuscript writing: Xinguan Yang, Zhuoran Gu

Final approval of manuscript: Xinguan Yang, Xiao Dong, Jiao Wang, Weiwei Li, Zhuoran Gu, Dashan Gao, Nanshan Zhong, Yubao Guan

Disclosures

The authors indicated no financial relationships.