Radiomic features are associated with EGFR mutation status in lung adenocarcinomas

Abstract

Background

This study retrospectively evaluated the capability of computed-tomography (CT) based radiomic features to predict EGFR mutation status in surgically-resected peripheral lung adenocarcinomas in an Asian cohort of patients.

Materials and Methods

298 patients with surgically resected peripheral lung adenocarcinomas were investigated in this institutional review board-approved retrospective study with waived consent. 219 quantitative 3D features were extracted from segmented volumes of each tumor, and 59 of these which were considered as independent features were included in the analysis. Clinical and pathological information were obtained from the institutional database.

Results

Mutant EGFR was significantly associated with female gender (p=0.0005); never smoker status (p<0.0001), lepidic predominant adenocarcinomas (p=0.017), and low or intermediate pathologic grade (p=0.0002). Statistically significant differences were found in 11 radiomic features between EGFR mutant and wild type groups on univariate analysis. Mutant EGFR status could be predicted by a set of five radiomic features that fall in three broad groups: CT attenuation energy, tumor main direction and texture defined by wavelets and Laws (AUC 0.647). Multiple logistic regression model showed that adding radiomic features to a clinical model resulted in a significant improvement of predicting power, as the AUC increased from 0.667 to 0.709 (p<0.0001).

Conclusions

CT based radiomic features of peripheral lung adenocarcinomas can capture useful information regarding tumor phenotype, and the model we built can be useful to predict the presence of EGFR mutations in peripheral lung adenocarcinoma in Asian patients when mutational profiling is not available or possible.

Introduction

Lung cancer accounts for 13% of new cancers diagnosed world-wide and is the leading cause of cancer related deaths ¹. 85% of all lung cancers are non-small cell lung carcinoma (NSCLC), with adenocarcinoma being the most common histologic subtype ². Over the last decade, molecular translational research advances have heralded major breakthroughs in the understanding, diagnosis and management of lung cancer, particularly the development of new target-based therapies directed against key signaling pathways involved in lung cancer growth and malignant progression ^{3, 4}. Small-molecule tyrosine kinase inhibitors (TKIs) that target the epidermal growth factor receptor (EGFR) were the first targeted drugs to enter clinical use for the treatment of NSCLC. Patients with EGFR mutations have a higher response rate to EGFR-TKIs (60%-80%) than those with EGFR–wild type or unknown mutation status (10%-20%) ⁵. Randomized trials have clearly shown that treatments with targeted TKIs, such as erlotinib, gefitinib, or afatinib are associated with longer progression-free survival (PFS) and higher objective radiographic response rates than standard first-line chemotherapy in patients with mutated EGFR lung cancer ^6-9. However, if gefitinib is administered in the case of non-EGFR mutated lung cancer, the patient will experience a shorter PFS compared to platinum-based chemotherapy ³, highlighting the importance of identifying this genetically unique subset of patients.

Cumulative epidemiology studies have identified several clinicopathological factors such as female, non-smokers, adenocarcinoma histology, and East Asian origin that associated with high prevalence of EGFR mutation ^{10, 11}. Unfortunately, there are no reliable clinical characteristics that allow for accurate prediction of EGFR mutation status. For some patients, biopsy samples may be the only tumor materials available for testing EGFR mutation status and they are often composed of variable ratios of tumor to normal cells,¹². Thus, mutant DNA alleles present at extremely low concentrations become difficult to detect, leading to a false negative result. Further, due to intratumoral heterogeneity, the portion of the tumor tested for EGFR mutation may also result as negative but may be truly positive ^13-15. In a recent study ¹⁶, two sequencing-based mutation detection approaches (dideoxy- and pyrosequencing) were validated against parallel sequencing in a clinical setting, and the results showed that dideoxy sequencing missed four responders and pyrosequencing missed two responders; meanwhile, precise quantification of mutant alleles revealed a low correlation of histopathological estimates of tumor content and frequency of mutant alleles, indicating that sequencing technologies with inferior sensitivity may fail to detect clinically relevant oncogene mutations in cancer patients. Therefore, receiving a negative mutation analysis result, one must consider whether the cell sample was truly representative for the EGFR mutation status of the lung tumor.

There have been several reports regarding the relationship between Computed- Tomography (CT) features and EGFR mutation status in NSCLC ^17-20; however, the findings were not consistent with each other. According to a study conducted by Zhou et al. ¹⁸, there were no differences in morphological CT features between EGFR mutation and wild-type tumors. In contrast, Rizzo et al. ¹⁷reported that EGFR mutation was significantly associated with air bronchogram, pleural retraction, small lesion size, and absence of fibrosis. Recent technological advances in medical imaging allow high throughput extraction of quantitative imaging features. Radiomics is the process of converting images to mineable data through computational approaches. These data can be used to develop decision support systems to accurately estimate patient risk and improve individualized treatment ^{21, 22}. Several studies have showed that such features extracted from CT images of lung cancers can be useful to distinguish radiation-induced fibrosis from tumor recurrence ²³, differentiate the presence of K-ras mutation from pan-wild type NSCLC ²⁴, provide independent predictive indicator of response to first-line chemotherapy ²⁵, and identify patients with locally advanced lung adenocarcinoma at risk of developing distant metastasis ²⁶. CT imaging is routinely used in lung cancer, and we thus hypothesize that if CT-based radiomic features associated with EGFR mutation status can be determined, they could provide a useful clinical predictor in patients with unresectable lung cancer or those wherein biopsy is unable to be performed. Imaging-based risk models may also provide additional information for clinicians on whether re-biopsy is needed for those patients with negative EGFR mutation result. Therefore, in this retrospective study, we performed a radiomic analysis to identify image biomarkers of harboring EGFR mutation in peripheral lung adenocarcinomas in a Chinese cohort of patients.

Materials and Methods

This retrospective study was approved by the institutional review board. Informed consent was waived.

Study Population

A consecutive search of the surgical database at our institution between December 2012 and March 2014 identified 397 patients with primary lung adenocarcinoma who fulfilled the following inclusion criteria: (a) pathology reports with diagnosis of lung adenocarcinoma; (b) preoperative thin-section CT images at Picture Archiving and Communication System (PACS) and the location of the lesion was peripheral (tumor involving subsegmental bronchus or smaller airway); (c) available test results for EGFR mutation status; (d) available clinical data. Thereafter, 99 patients were excluded due to the following reasons: receiving preoperative treatment, such as radiotherapy or chemotherapy; the duration between CT examination and subsequence surgery exceeded one month; and cases with lung cancer that is difficult to contour the tumor margin on CT images.

Clinical and pathological data collected for analysis included gender, age at diagnosis, smoking status, pathologic TNM stage, and histologic lung adenocarcinoma subtypes. Smoking status was categorized into two groups, never smokers and smokers which included former or current smokers. Tumors were staged pathologically according to the seventh edition of the American Joint Committee on Cancer Staging Manual ²⁷. Tumors were diagnosed as adenocarcinoma and then categorized according to the 2011 IASLC/ATS/ERS classification system ²⁸.

CT Examination

Chest CT examinations were conducted using Somatom Sensation 64 (Siemens Medical Solutions, Forchheim, Germany), Light speed 16 (GE Medical Systems, Milwaukee, WI), or Discovery CT750 HD scanner (GE Medical Systems, Milwaukee, WI). The parameters used were as follows: 120 kVp with tube current adjusted automatically, pitch was 0.969, reconstruction thickness was 1.5 mm, and reconstruction interval was 1.5 mm for the 64-detector scanner; tube voltage was 120 kVp, tube current was 150-200 mA, pitch was 0.969; reconstruction thickness was 1.25 mm, and reconstruction interval was 1.25 mm for the 16-detector scanner and Discovery CT750 HD scanner.

Detection of EGFR mutations

For the gene mutation analysis, tumor specimens were obtained by surgical resection. We performed EGFR mutation analyses of four tyrosine kinase domain (exons 18-21), which are frequently mutated in lung cancer, as previously described ²⁹. EGFR mutations were determined by amplification refractory mutation system (ARMS) real-time technology using Human EGFR Gene Mutations Detection Kit (Beijing ACCB Biotech Ltd).

Tumor segmentation

We used Definiens Developer XD© (Munich, Germany) as the image analysis platform to perform tumor segmentation and feature extraction. A lung tumor analysis (LuTA) tool within Definiens Cognition Network Technology was used. Lesions were volumetrically segmented using semi-automatic approach by two radiologists with more than 6 and 3 years of experience in CT imaging of thoracic malignancies (Figure 1). The semi-automatic segmentation workflow which contained the following four steps named pre-processing, semi-automated correction of the pulmonary boundary, click and grow, and manual refinement and generation of lesion statistics were described in detail in previous studies ^30-32. Here, the single click ensemble segmentation (SCES) algorithm ³¹ which is an advanced version of the previous click & grow algorithm and reduces sensitivity towards the location of the initial seed-point was used. The SCES makes use of the original algorithm by choosing different seed points automatically within a specified area of the lesion and performing region growing with each generated seed point. Each of the two radiologists reviewed the segmented images in consensus, and any discrepancies were resolved by discussion until consensus was reached.

Fig. 1.

Representative CT image with tumor segmentation using semi-automated algorithm for peripheral adenocarcinoma: (a) A lobulated lung tumor in the right middle lobe was chosen for segmentation. (b) One radiologist segmented the boundary of the tumor which was shown in green outline. (c) 3D view of the lung and segmented tumor.

Radiomic feature extraction

We extracted a total of 219 features from each of the 3D objects. These features were divided into eight categories, including tumor size, shape, location, air space, pixel intensity histogram, run length & co-occurrence, laws texture, and wavelets. A description of all features is provided (Supplementary Table 1), and detailed description of texture features could be found in the previous study ³³. Our features cover a variety of descriptors from size, location, attachment of the lesion of interest, to CT pixel distribution and texture as appearing on CT image. We characterize texture in observed CT images using Laws feature descriptor and in the decomposed domain by using wavelets. Texture feature are known carry information that are not always typically observed by human eye. Metrics described by Wavelet transformation, Laws features are considered to describe subtle characteristics in the image and have been shown to be useful in image classifications ³⁴.

Statistical Analysis

All statistical analyses were performed using SAS software (version 9.4, Cary NC) and MATLAB (Natick, MA). Power transformation was considered to apply parametric analytical tools. The correlation between features was investigated in order to address collinearity issue. The highly correlated features (correlation > 0.9) were regarded as dependent features which were not considered in this analysis. Thus, 59 of 219 features were considered as mutually independent features for prediction of EGFR mutation. Fisher’s exact test and the Kruskal-Wallis test were used for categorical and continuous variables between two groups, respectively. Multiple logistic regression analysis was performed. The final model was selected using the backward elimination method. Variables with p-value of < 0.25 in univariate model were entered in the initial model, and then a variable with p-value of > 0.15 was eliminated at each step. The elimination procedure was terminated when the p-value of all variables in the model is < 0.15. The accuracy and error of the predictive models (i.e., the area under the curve (AUC) and 95% confidence interval (CI)) were estimated by the Bootstrap method with 1000 Bootstrap samples. Receiver operating characteristic (ROC) curves for each model were constructed and the AUC was calculated with EGFR mutation status determined by amplification refractory mutation system-PCR as outcome. Further, various predictive models were developed by the support vector machine (SVM) and the principal component analysis (PCA) and were compared with the logistic regression model. In this study, the regression model was selected as the AUC of the model which was higher than any other models (results were omitted). A two-sided p-value of < 0.05 was regarded as statistically significant.

Results

The patients’ demographic and clinicopathological data are presented in Table 1. All the enrolled 298 patients were surgically treated: lobectomy in 280 patients, pneumonectomy in 4, wedge resection in 6, and segmentectomy in 8 patients. Overall, there were 126 men and 172 women with a median age of 60 years (range: 30 – 80 years). The pathologic stage distribution was as follows: IA in 117 patients (39.26%), IB in 63 patients (21.14%), IIA in 22 patients (7.38%), IIB in 6 patients (2.01%), IIIA in 72 patients (24.16%), IIIB in 2 patients (0.67%), and IV in 16 patients (5.37%). Most of the tumors were early stage (stage I or II) (69.80%). All cases were lung adenocarcinomas and the most common histologic subtype among invasive adenocarcinomas was acinar predominant subtype (42.95%), followed by lepidic predominant subtype (23.83%). EGFR mutation results (Supplementary Table 2) were satisfactorily demonstrated in all patents, with 137 (45.97%) patients out of the full cohort of 298 cases were identified as EGFR mutant and 161 (54.03%) as EGFR wild-type.

Table 1.

Clinicopathological characteristics of patients

Variables		Overall dataset (n=298)	EGFR mutant (n=137)	EGFR wild-type (n=161)	p-value
Median Age (range)		60 (30 – 80)	59 (37 – 80)	60 (30 – 79)	0.51
Gender	Male	126	43 (31.39%)	83 (51.55%)	0.001
	Female	172	94 (68.61%)	78 (48.45%)
Smoking status	Smokers	136	43 (31.39%)	93 (57.76%)	<0.0001
	Never smokers	162	94 (68.61%)	68 (42.24%)
Pathologic grade†	High	74	20 (14.60%)	54 (33.75%)	0.0002
	Low/intermediate	223	117 (85.40%)	106 (66.25%)
Histologic subtype*	Lepidic predominant adenocarcinomas	74	43 (31.39%)	31 (19.25%)	0.022
	others	224	94 (68.61%)	130 (80.75%)
Stage	I or II	208	93 (67.88%)	115 (71.43%)	0.53
	III or IV	90	44 (32.12%)	46 (28.57%)

^†Adenocarcinoma in situ (AIS) and minimally invasive adenocarcinoma (MIA) were classified as low grade; lepidic, acinar, and papillary as intermediate grade; micropapillary, solid and invasive mucinous adenocarcinoma as high grade. One case with enteric was eliminated in this analysis.

^*Histologic subtype was categorized as lepidic predominant adenocarcinomas (AIS, MIA, and lepidic predominant invasive adenocarcinoma) and other subtypes of dominant histologic findings (acinar, papillary, micropapillary, and solid predominant as well as variants of invasive adenocarcinoma).

There were significant differences in gender, smoking status, pathologic grade, and histologic subtype between EGFR wile-type group and EGFR mutant group (Table 1). Concerning gender, there were significantly more female patients with mutant, compared to wild-type EGFR in lung adenocarcinomas (OR=2.33, 95% CI: 1.45 – 3.74, p=0.001). Smokers with mutant EGFR were significant lower than smokers with wild-type (OR=2.99, 95% CI: 1.85 – 4.82, p<0.0001). EGFR mutations were also significantly more frequent in patients with low or intermediate pathologic grade (OR=2.98, 95% CI: 1.67 – 5.30, p=0.0002), and patients with lepidic-predominant adenocarcinomas (OR=1.92, 95% CI: 1.13 – 3.27, p=0.022). There were no differences in stage distribution or median age between EGFR mutant and wild-type groups (p=0.53; p=0.51). Univariate analyses revealed that four clinical factors may be associated with EGFR mutation status in peripheral lung adenocarcinoma: gender, smoking status, histologic subtype, and pathologic grade were all significant predictors of harboring and EGFR mutation (Table 2).

Table 2.

Univariate analysis for clinicopathological factors that predict EGFR mutation status

Clinical Features		Odds Ratio			p-value
		Point	95% CI
			Lower	Upper
Stage	I or II	reference
	III or IV	1.18	0.72	1.94	0.51
Gender	Female	reference
	Male	0.43	0.27	0.69	0.0005
Histologic subtype	Others	reference
	Lepidic predominant adenocarcinomas	1.92	1.13	3.27	0.017
Pathologic grade	Low or Intermediate	reference
	High	0.34	0.19	0.60	0.0002
Smoking status	Never smokers	reference
	Smokers	0.34	0.21	0.54	<0.0001
Age (per 5 years old increase)		0.97	0.85	1.11	0.68

CI – confidence interval

We then investigated the association of radiomic features with EGFR mutation status. Giving that this analysis produced far more features that were considered dependent, only a prioritized subset of features were selected for further analysis to avoid over fitting. Prioritization methods are described in ref. 29. Fifty-nine independent features were selected finally (Supplementary Table 3). Among these 59 independent features, 11 were independent predictors of harboring EGFR mutation (Table 3, Supplementary Table 4) on univariate analysis, including one feature describing tumor shape (F26), two features describing tumor location (F17, F27), one feature describing airspace (F19), two pixel intensity histogram based features (F185, F186), two runlength & co-occurrence based features (F47, F51), two laws texture features (F90, F111), and one wavelet texture feature (F190).

Table 3.

Univariate analysis for Radiomic features that predict EGFR mutation status

Radiomic Features		Odds Ratio			p-value
		Point	95% CI
			Lower	Upper
F19	10a_3D_Relative_Volume_AirSpaces	1.16	1.04	1.3	0.007
F186	Histogram ENERGY Layer 1	0.49	0.29	0.82	0.007
F90	3D Laws features L5 L5 L5 Layer 1	0.82	0.7	0.96	0.015
F51	AvgLRE	0.11	0.02	0.72	0.021
F47	AvgCoocurrence-mean	0.98	0.97	1	0.025
F190	3D Wavelet decomposition. P2 L2 C9 Layer 1	1.26	1.03	1.55	0.025
F185	Histogram SD Layer 1	1.88	1.07	3.31	0.027
F26	Elliptic Fit	0.11	0.02	0.79	0.028
F17	9f-3D-Min-Dist-COG-to-Border	0.59	0.36	0.96	0.032
F111	3D Laws features R5 E5 R5 Layer 1	0.71	0.51	0.99	0.041
F27	Main direction	1.21	1	1.45	0.048

With multiple logistic regression analyses, clinical features of smoking status and pathologic grade proved to be independent predictors of EGFR mutation, and the AUC of ROC was 0.667 (95% CI: 0.604 - 0.721). model produced from radiomic features alone showed moderate predictive power (AUC=0.647; 95% CI: 0.576 - 0.701) for identifying EGFR mutant status. There was a significant difference between AUCs of the logistic regression model incorporating only clinical features, and that incorporating only radiomic features (p<0.0001). When clinical and radiomic features were combined, the AUC was increased to 0.709 (95% CI: 0.654 - 0.766) (Table 4). The model generated with combined clinical and radiomic features was superior to both the model generated with clinical features alone (p<0.0001) and the model created with radiomic features alone (p<0.0001) (Figure 2).

Table 4.

Multiple logistic regression analysis of clinicopathological parameters and radiomic features predicting the presence of EGFR mutation in peripheral lung adenocarcinomas

Model	Features		p-value	Odds Ratio			AUC
				Point	95% CI		Point	95% CI
					Lower	Upper		Lower	Upper
Clinical features	Pathologic grade (ref= Lower or Intermediate)	High	0.002	0.39	0.22	0.70	0.667	0.604	0.721
	Smoking status (ref = Never smokers)	Yes	<.0001	0.38	0.23	0.61
Radiomic features	Main direction (F27)		0.115	1.17	0.96	1.42	0.647	0.576	0.701
	3D Laws features L5 L5 S5 Layer 1 (F92)		0.010	0.52	0.31	0.85
	Histogram ENERGY Layer 1 (F186)		0.025	0.50	0.27	0.92
	3D Wavelet decomposition. P2 L2 C9 Layer 1 (F190)		0.048	1.29	1.00	1.67
	3D Wavelet decomposition. P1 L2 C5 Layer 1 (F216)		0.137	1.16	0.96	1.40
Clinical + Radiomic features	Pathologic grade (ref= Lower or Intermediate)	High	0.016	0.47	0.25	0.87	0.709	0.654	0.766
	Smoking status (ref = Never smokers)	Yes	0.001	0.41	0.25	0.67
	Main direction (F27)		0.20	1.14	0.93	1.40
	3D Laws features L5 L5 S5 Layer 1 (F92)		0.056	0.60	0.36	1.01
	Histogram ENERGY Layer 1 (F186)		0.11	0.60	0.32	1.13
	3D Wavelet decomposition. P2 L2 C9 Layer 1 (F190)		0.13	1.22	0.94	1.59
	3D Wavelet decomposition. P1 L2 C5 Layer 1 (F216)		0.24	1.13	0.92	1.38

AUC – area under curve

Fig. 2.

Receiver operating characteristic (ROC) curves for the prediction of EGFR mutation using logistic regression model that included clinical factors alone (green line), a model that use radiomic features alone (red line), and a model that combined clinical factors and radiomic features (blue line). The highest area under the curve (AUC) was achieved for the combination of clinical factors and radiomic features (AUC=0.709).

Discussion

There are various methods to detect EGFR mutations, such as direct sequencing of polymerase chain reaction (PCR)-amplified genomic DNA, high-resolution melting analysis, fragment analysis, restriction fragment length polymorphism and the amplification refractory mutation system ³⁵; however, these molecular methods are generally costly, and sometimes in a low percentage of tumor cells, we would have most likely missed the mutations and re-biopsy had to be recommended. In this study, we sought to apply radiomic features to peripheral lung adenocarcinomas to determine if we could noninvasively discriminate EGFR-mutant from EGFR- wild type case on a routine practice without adding additional cost.

We found that 137 of the tumors harbored EGFR mutation, which corresponded to 45.97% of the 298 tumor samples, this is in keeping with previous reports on Asian patients ³⁶; in addition, exon 19 and exon 21, the most common mutation types of the EGFR gene, showed nearly the same percentage in these patients (46.72%, 48.91% respectively). Dual mutations were detected in one patient (0.73%) in this study, confirming that the existence of both EGFR TKI resistant mutation (exon 20) and sensitive mutation (exon 21) is rare. Several studies indicated that EGFR mutation was strongly related with never smokers, female, adenocarcinoma, and pathologic stage ^{29, 36}. Consistent with most of these findings, the rate of EGFR mutation in female and never smoker patients with lung adenocarcinomas was considerably higher compared to male and smokers in this study. Although EGFR mutation was detected more frequently in early stage patients (67.88%) compared to advanced stage (32.12%), the difference was not significant. The results of the multiple logistic regression analysis revealed that pathologic grade and smoking status were independent predictors for EGFR mutation and the AUC was 0.667. Besides clinical information, more detailed factors will likely be required to identify those at high probability of harboring EGFR mutations.

CT imaging is used widely in oncologic practice for lung tumor characterization. Usually, we interpret the image based on visual assessment; there are features, however, within each image that may not be perceived by the naked eye and requires computer-aided techniques. Previous authors have demonstrated the potential of quantitative CT based texture analysis in differentiation of K-ras mutation from pan-wildtype NSCLC ²⁴. To the best of our knowledge, only one study ³⁷that published recently has explored the association between CT gray-level texture features and EGFR mutations in a relatively small sample size (25 patients with EGFR mutation and 20 patients with EGFR wild-type). In this study, we present comprehensive radiomic analysis using semi-automatic segmentation in 298 peripheral lung adenocarcinomas. 219 radiomic features were extracted to assess the ability to predict EGFR mutation status. Considering that using too many features in the classification algorithm can lead to over-fitting, in which noise or irrelevant features may exert undue influence on classification decisions, only those features which are not associated with other features were selected for further analysis. With this approach, we found that 11 radiomic features from seven different feature categories were significantly associated with EGFR mutations. Texture features, Wavelet features, Laws features, and along with pixel statistical have seen a resurgent use in medical images, especially CT and MRI images ^{24, 38}. In our study we formed sets of five one dimensional Laws filters, each designed to describe different structures in the image (E: Edges, S: Spots, R: Ripple, W: Waves, L: Lowpass). The wavelet transformation was limited to two levels (L) of decomposition on each of the nine faces (C) on the 3D tumor, with two types of metrics (P), namely energy and entropy. Our statistical model finds Laws features L5L5S5 and Wavelet P2L2C9 & P1L2C5 to be one of the five predictors of EGFR mutation status. Several previous studies ^{17, 39}showed that EGFR mutation was associated with small tumor size. In our analysis, size based features including longest diameter and short axis were not significant predictor for EGFR mutation. We believe that this difference can be explain in part by the fact that most clinically assessed tumor diameters are manually drawn on the central slice of CT images and limited to one dimension of the tumor, while whole tumor volume was taken into account in the size based radiomic features.

With logistic regression analysis, we identified that radiomic features could be served as an imaging surrogate for EGFR mutation, although the AUC of radiomic features alone was not as high as the model created with clinical features, they provide complementary information, as the combination results in an improved ROC curve (AUC, 0.709), and we were able to significantly increase the predictive performance of EGFR mutation. These findings suggested that the combination of radiomic data and demographic information in a system model is more effective.

There are some limitations in this study. One limitation is that this study is retrospective and limited to only Eastern Asian populations; care should be taken before generalizing our findings to other populations. Second, radiomic features were derived from semi-automatic segmentation by radiologists, which can be influence by observers’ subjective trend. However, the results of automatic boundary extraction method were not satisfactory for all the lesions, particularly in the case of tumors with GGO components, as their margins are usually unclear form the adjacent normal lung parenchyma. Furthermore, since atelectasis is common in patients with central lung cancers, and differentiation between them is rather difficult as both appear as solid density on CT, our analysis did not include central lung cancers. A prospective multi-Institutional study with a large patient cohort would be required to confirm our observations.

Conclusions

In summary, this study revealed associations between CT based radiomic features and EGFR mutation status in peripheral lung adenocarcinomas, and therefore non-invasive radiomic phenotype analysis has the potential to improve the differentiation of EGFR mutate from wild type when used in addition to clinical predictors.

Clinical Practice Points.

• The presence of activating EGFR mutations has been shown to be prognostic for a more favorable outcome to TKI therapy in lung adenocarcinomas.

• Several reports have described the relationship between mutation status of EGFR and traditional radiological features. Radiomic-based approach allows high throughput extraction of quantitative parameters from CT images which beyond what is visually perceived by the human eyes. Thus, we hypothesized that radiomic analysis of routinely performed preoperative CT images could provide imaging biomarkers for EGFR mutations.

• In this study we demonstrated that eleven CT based radiomic features have significant association with EGFR mutations. Adding radiomic features to clinical model could improve the predicting power of EGFR mutations, thus helping in formulating a better clinical decision without adding additional cost.

Abbreviations

NSCLCs

non-small cell lung cancers

TKIs

tyrosine kinase inhibitors

EGFR

epidermal growth factor receptor

PFS

progression-free survival

CT

computed tomography

PACS

picture archiving and communication system

IASLC/ATS/ERS

International Association for the Study of Lung Cancer, American Thoracic Society, and European Respiratory Society

ARMS

amplification refractory mutation system

SCES

single click ensemble segmentation

CI

confidence interval

ROC

receiver operating characteristic

AUC

area under the curve

PCR

polymerase chain reaction

AIS

adenocarcinoma in situ

MIA

minimally invasive adenocarcinoma