Detection of EGFR mutations in early-stage lung adenocarcinoma by machine learning-based radiomics

Introduction

Molecular targeting therapy has revolutionized the landscape of the treatment of advanced-stage lung adenocarcinoma (1-5). Genetic mutations in the epidermal growth factor receptor (EGFR) domain are now established biomarkers to predict the treatment response using the EGFR tyrosine kinase inhibitor (TKI). The incidence of these mutations is approximately 50% in the Asian population and fifteen% in the Caucasian population (6). In the adjuvant settings, osimertinib, a third-generation EGFR-TKI, has recently been shown to prolong disease-free survival in patients with completely resected non-squamous non-small cell lung cancer (NSCLC) with EGFR mutations (7). Currently, the efficacy of neoadjuvant osimertinib is being tested to further improve the prognosis for resectable primary non-squamous NSCLC harboring EGFR mutations (8). EGFR mutation is also a key exclusion criterion in a clinical trial with neoadjuvant immunotherapy (Impower 30, NCT03456063). For inclusion in these clinical trials, invasive tissue biopsy is mandatory at this moment. Although computed tomography (CT)-guided biopsy can be effective in the case of peripheral lung tumors, tissue biopsy can be challenging in patients with early-stage operable lung cancers in clinical practice, where the sensitivity of the transbronchial tissue biopsy for peripheral tumors is around 57% (9) while that of the endoscopic ultrasound-guided needle aspiration of the mediastinum is around 89% (10).

Validated clinical predictors of EGFR mutations are reported (gender, smoking, ethnicity). A prediction model based on these three factors performed reasonably with area under the curve (AUC) of 0.75 (11), which could be further improved. Radiomics is defined as the quantitative analysis of multiple high-throughput data derived from medical imaging examinations. Previously, radiomic features, in combination with clinical or pathological data, can be used to develop models that predict tumor characteristics in the field of lung cancers (12).

We hypothesized that EGFR mutations could be detected by radiomics of primary early-stage resectable lung adenocarcinoma using preoperative CT. We present the following article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-22-2683/rc).

Methods

Patient selection

The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Institutional Review Board of the Cancer Institute Hospital (No. 2021-GA-1064) on June 28, 2021, and the patient consent requirement was waived because of its retrospective nature. This retrospective study included consecutive patients with clinical stage I/II lung adenocarcinoma who underwent curative-intent pulmonary resection from March 2016 to December 2016. Patients with pure ground-glass opacity or lobar atelectasis or those who received preoperative treatment were excluded. Patients who underwent only non-contrast CT or CT taken more than 2 months before surgery were also excluded. EGFR gene mutation status was analyzed from paraffin-embedded surgical specimens using a polymerase chain reaction (PCR) kit (cobas^®EGFR Mutation Test v2, Roche, Basel, Switzerland) targeting exons 18, 19, 20, and 21. The overall workflow of this study is shown in Figure 1.

Figure 1 Study design. A: radiomic features were extracted from the preoperative enhanced CT scan. B: an interclass correlation coefficient >0.8 between two physicians was considered reproducible features. Features showing an Spearman’s correlation coefficient ≥0.85 were excluded. C: machine learning-based models were created for prediction of EGFR mutations. D: AUC was calculated through five-fold cross-validation to evaluate predictive performance. EGFR, epidermal growth factor receptor; CT, computed tomography; AUC, area under the curve.

The segmentation

All patients underwent contrast-enhanced chest CT using Discovery CT750 HD (GE Healthcare, Chicago, IL, USA). The acquisition parameters were as follows: tube voltage, 120 kVp; tube current, 100–649 mA; exposure time, 400–699 ms; milliampere-seconds, 50–329.9 mAs, slice thickness, 1.25 mm (Table S1). In axial CT images under lung field view, tumor boundary was segmented into slices containing primary tumor using Eclipse software ver. 8.6 or ver. 10.0 (Varian Medical Systems, Palo Alto, CA, USA). We then created volume of interests (VOIs) as follows (Figure 1): the tumor (VOI within the tumor boundary), tumor rim (VOI within ±3 mm of the tumor boundary), and tumor exterior (VOI between +10 mm outside the tumor and the tumor boundary). The demarcation of the tumor rim and tumor exterior areas was calculated using the automated function of the software. If the created VOI went outside the lungs (i.e., the VOI partially included mediastinum or chest wall), the contour was manually modified to fit within the lung field.

The radiomics feature extraction

Prior to feature extraction, all CT images were resampled to an isotropic grid of 1×1×1 mm³ using B-splines to consistently calculate the three-dimensional features. The Laplacian of Gaussian (LoG) filter (σ =0.5, 1.0, 1.5, 2.0, and 2.5 mm) and wavelet filter that computes eight decompositions (HHH, HHL, HLH, HLL, LHH, LHL, LLH, LLL) per level were used for transforming CT images. A total of 3,951 radiomic features, including 108 original features (first-order statistics, n=18; shape, n=15; texture, n=75) and 1,209 filtered features (LoG, n=465; wavelet, n=744), were extracted from each VOI (n=1,317×3) using PyRadiomics version 3.0.1 (13). The texture features were categorized as gray-level co-occurrence matrix (n=24), gray-level run length matrix (n=16), gray-level size zone matrix (n=16), gray-level dependence matrix (n=14), and neighboring gray-tone difference matrix (n=5). The bin width was set to 25 Hounsfield units (14).

The feature selection and model building

To assess the inter-observer reproducibility for VOI segmentations, segmentation of the tumor was performed using 30 randomly selected patients by two physicians with experience in lung cancer diagnostics (KH and RS) (15,16). Then, the interclass correlation coefficient (ICC) of >0.8 calculated between two physicians was considered a reproducible feature. To eliminate redundant features, Spearman’s correlation coefficient (SCC) among the features of all possible two combinations was calculated, and features showing an SCC of ≥0.85 were excluded.

Prior to model building, each feature was standardized using Z-scores. The machine learning-based radiomics model [using the Boruta algorithm (17-19) (BorutaPy version 0.3) for feature selection and Random Forest (Scikit-learn version 1.0.2) for prediction] was constructed to detect EGFR mutation. Suitable hyperparameters for Random Forest were determined by Optuna version 2.10.0 (20), while those for Boruta were empirically determined. The tuned hyperparameters by Optuna were max_depth, max_leaf_nodes, min_samples_leaf, and n_estimators. The meanings of these parameters are described in Kursa et al. (17). The number of maximum iterations to select the optimal features for Boruta was set to 500. The combined model incorporated both radiomic as well as clinical features (gender and smoking history). Gender and smoking history were selected as these are clinically validated predictor of EGFR mutation (21). As this study was performed in monotonous population (Asians), ethnicity was not included. Other clinical variables in this study were not included to avoid overfitting. The current study is based on TRIPOD Type 1b (22).

Assessment of performance of the models and statistical analyses

The performance was validated with non-nested five-fold cross-validation, and the mean area under the receiver operating characteristic (ROC) curve was used for evaluating the prediction models. Accuracy, recall, precision, specificity, and F1 score were also calculated. Comparison of the clinical characteristics was performed for continuous and categorical variables using t-test or Mann-Whitney test, and Fisher’s exact test, respectively. A comparison of AUCs was performed using Delong’s test. Statistical analyses were performed using R software and GraphPad Prism version 8 (GraphPad Software Inc., La Jolla, CA, USA). All P values were two-sided, and P<0.05 was considered statistically significant.

Results

Clinical characteristics of patients

The patient characteristics are summarized in Table 1. This study enrolled 99 patients (mean age, 66±11 years; female, 66.6%; clinical stage I/II 89.9%/10.1%). There was no lung cancer that invaded mediastinum or chest wall. EGFR mutations in the surgical specimen were detected in 46 (46.5%) patients. Compared to wild type, patients with EGFR mutation included significantly more female (female: EGFR wild group, 54.7% vs. EGFR mutation group, 78.3%; P=0.02) and less smoking history (current or previous) (smoking history: EGFR wild group, 60.4% vs. EGFR mutation group, 19.6%; P=0.01) as expected. Further, a significant difference in the histologic subtype (P=0.005) between the groups existed. In the EGFR wild group, the most common histologic subtype was the papillary type (39.6%), followed by the solid subtype (30.2%). In the EGFR mutant group, the most common histologic subtype was the papillary type (71.7%), followed by the adenocarcinoma in situ (AIS)/minimally invasive adenocarcinoma (MIA) pattern (13.0%). There were no significant differences in mode of surgery and nodal status between the groups.

Among the EGFR mutant patients, 26 (54.2%) presented exon 21 L858R, 19 (39.5%) had exon 19 deletions, and 1 (2.1%) exhibited exon 21 L861Q as an uncommon mutation. Of 26 patients with exon 21 L858R mutation, two (4.2%) had de novo exon 20 T790M as well.

Feature selection and diagnostic performance

After assessing the ICC, 1,317 features extracted from VOI of the tumor were reduced to 1,267 features (in total, 3,955 features were reduced to 3,901 features). Further, 3,901 features were reduced to 453 robust radiomic features using SCC analysis (Table S2). Of 453 features that cleared the ICC and SCC, a median of 4 (range, 2 to 8) radiomic features from each validation session were selected using the Boruta algorithm.

The predictive performance of radiomics and combined models are summarized in Table S3. In the training sets, the mean AUC of the radiomics and combined models were 0.78 (range, 0.64–0.94) and 0.74 (range, 0.70–0.80), respectively. In the validation set, the combined model showed acceptable predictive performance with a mean AUC of 0.83 (range, 0.74–0.95), which tended to be higher than that of the radiomics model (mean AUC, 0.75; range, 0.65–0.82). However, Delong’s test revealed that the difference in AUC did not reach statistical significance (P=0.052). The ROC of each validation set in the radiomics and combined model are presented in Figure 2.

Figure 2 Receiver operating curves of each validation session in each model. The mean AUCs of five cross-validation sessions in the radiomics and combined models were 0.75 (range, 0.65–0.82) and 0.83 (range, 0.74–0.95), respectively. AUC, area under the curve.

Importance of features selected using machine learning models

A total of 15 radiomic features from the radiomics model and 17 features, including two clinical features, from the combined model were selected through a five-fold cross-validation session. The selected radiomic and clinical features and the coefficients in each cross-validation session are listed in Table S4.

Figure 3 represents the ranking of each selected radiomic feature based on the average coefficient through five sessions in the radiomics model. Figure 4 shows the rankings of radiomic and clinical features in the combined model. Each feature was converted to a percentage value relative to the average coefficient of the top-ranked feature in each model. It was noted that the average coefficient of two of the radiomic features (wavelet-LHL_firstorder_90Percentile from the tumor exterior and wavelet-HHH_gldm_DependenceVariance from the tumor) were superior to those of the relevant clinical features (gender and smoking history) in the combined model, suggesting that the radiomic features might have a higher impact on detecting EGFR mutations over the clinical features.

Figure 3 Ranking of features selected using Boruta algorithm in the radiomics model (each number represents a percentage of coefficient of the selected feature with reference to the top-ranked feature). HHH, HLL, HHL, LHH, LHL, HLH: wavelet filter.

Figure 4 Ranking of features selected using Boruta algorithm in the combined model (each number represents a percentage of the coefficient of the selected feature with reference to the top-ranked feature). HHH, HHL, LHH, LHL, HLH: wavelet filter.

Significance of top-ranked radiomic features

Representative images showing the top three and bottom three cases in the two features that were ranked above clinical features in the combined model are demonstrated in Figure 5. The first order describes the distribution of voxel intensities within the image region. A voxel intensity of 90 percentile in the tumor exterior was selected in this feature. Dependence variance measures the variance in dependence size in the image (within the tumor in this feature). No apparent recognizable pattern was observed, particularly in these images, indicating that radiomics could detect subtle changes unrecognizable to human eyes.

Figure 5 Left side: representative images of top three ranked cases in two radiomic features which was ranked above relevant clinical features in the combined model. Allow indicates cases with higher-ranked feature within EGFR mutation+. Right side: cases with bottom three ranked cases in each radiomic feature. Allow indicates cases with lower-ranked feature within EGFR mutation−. LHL, HHH: wavelet filter. EGFR, epidermal growth factor receptor.

Discussion

In this study, we demonstrated the feasibility of predicting EGFR mutations in early-stage operable adenocarcinomas using radiomic features with a mean AUC of 0.75 in the validation settings. Although this was not statistically different, mean AUC improved up to 0.83 by combining with established clinical predictive factors of EGFR mutations (gender and smoking history). To quantify the contribution of these clinical variables more clearly, we further created the Clinical model and the all-combined model which includes all clinical variables (age, gender, smoking history, and clinical stage) and radiomic features (Table S5). Consequently, the mean AUC of the clinical model was identical to that of the radiomics model (0.77 vs. 0.75). However, the predictive performance in the all-combined model was limited (i.e., the mean AUC of 0.82), implying the selection of gender and smoking history for the Combined model is reasonable. We utilized a machine learning model to select important features and generate a predictive model out of a large number of features extracted from radiological images. Classically, the existence of air bronchogram or ground glass opacity is a known semantic CT feature associated with EGFR mutations (23). However, subtle changes in the portion of ground glass opacity may not be distinguishable to the human eye. Radiomics can translate these tiny changes into the probability of mutation. Other unknown semantic or unrecognizable features, as suggested in Figure 3, were also integrated into the computation.

Previous studies have shown the possibility of EGFR mutation status in NSCLC utilizing CT images (24-29). However, only few of them comprised the surgical population. In this study, we focused on the preoperative CT images from early-stage NSCLC, aiming for potential usage in neoadjuvant settings. Given the ongoing research and the rapid evolution in clinical practice of advanced-stage NSCLC, as well as the adjuvant setting driven by EGFR-TKI (7,30) and immunotherapy (31), neoadjuvant precision medicine can plausibly become a reality with preoperative treatment guided by actionable biomarkers. EGFR mutation should play a pivotal role in this setting, considering its prevalence, the existence of a series of effective EGFR-TKIs (1-5), and its association with a low response rate to immunotherapy (32). However, biopsies for relatively small peripheral targets in early-stage operable patients can be challenging. Occasionally, the quantity or quality of samples is not enough for DNA testing. Moreover, even in the liquid biopsy (ctDNA), there is the possibility of false-negative result in NSCLC (33), thereby considering the tissue-based analysis for EGFR mutations as the reference standard (34). Our radiomics-based non-invasive prediction of EGFR mutations could help guide the choice of preoperative treatment in this setting as additional information to the existing methods. Potentially, integration with other non-invasive predictive measures such as liquid biopsy (35) can decrease the need for biopsies in the future.

Several potential benefits of radiomic biomarkers exist in the neoadjuvant setting. First, this could avoid biopsy-related complications. Although transbronchial and trans-thoracic biopsies are considered established procedures, rare but serious complications hampering curative operations can still occur (36,37). Second, our technology could save time by skipping biopsy and waiting time for results. Waiting time before surgery can impact long-term outcomes (38); hence, saving this time could benefit patients from not just avoiding complications but also potentially improving the oncological outcomes.

Our findings suggested the importance of the peritumoral region to detect EGFR mutations. There are several explanations. First, the peritumoral region may have detected the skewed field of the lung parenchyma or broncho-vasculature structures, possibly due to the related mutation status of the tumor. Also, cancerous islets, known to be found adjacent to tumors (39), may be subtle changes to the human eye field but could have been detected as a radiomic feature. Second, manual demarcation may have missed some tumor edge which could have been picked as the peritumoral region. The top-ranked radiomic feature in the combined model and the second-ranked feature in the radiomics model were from the tumor exterior region, suggesting the importance of the signal outside the visible tumor margin for predicting the mutation profile.

There are several limitations to this study. First, this is a single institutional retrospective study with relatively small sample size. Although we performed cross-validation according to TRIPOD guidelines (22), our findings need further validation with a large sample set. Moreover, because it is well-known that an EGFR mutation profile can differ across ethnicities (40), our model may perform differently in other regions or ethnicities. Therefore, the models should be further validated in a multi-national and multi-ethnicity study. Second, although we focused only on adenocarcinomas considering the clinically proven utility of EGFR-TKIs against the disease and relative rarity and uncertainty of the predictive value of the treatment efficacy of squamous cell lung carcinoma EGFR mutation (in fact recommendation of the EGFR mutation testing for squamous cell carcinoma is somewhat mixed among society guidelines), we are aware that adenocarcinoma can be confirmed only after invasive biopsy at this moment. Nevertheless, adenocarcinoma is the most relevant type of NSCLC, and this prediction model perhaps could be combined with the radiomics prediction model of the histological type of NSCLC. Besides, sometimes biopsy of a tumor only provides histology diagnosis but not enough sample for molecular testing result in daily practice. Our image biomarker may be useful in that situation. Third, we utilized manual demarcation of the tumor boundary to extract radiomic features that can be a source of errors. Although we performed ICC to diminish inter-observer variability, this could have affected the results as radiomics is sensitive to slight changes. Incorporating automated detection technology could reduce the inter-observer variation. Fourth, we chose non-nested cross validation for evaluating the prediction performance, thus optimal hyperparameters for Random Forest were not tuned using independent validation sets. Finally, we did not consider other machine-learning techniques other than Random Forest for prediction. Jia et al. used Random Forest classifier to predict EGFR mutations using radiomic features and clinical variables (29), and demonstrated that relatively higher predictive performance was observed (AUC =0.828), which is consistent with our results (AUC =0.83).

Conclusions

Radiomic features may help detect EGFR mutations in lung adenocarcinomas in preoperative settings. Important signals even outside the recognizable tumor boundary may exist. This non-invasive image-based technology in detecting actionable mutations could help guide future precision neoadjuvant therapy.

Acknowledgments

This paper was presented as a poster in the 102nd annual meeting of the American Association for Thoracic Surgery, Boston, MA, May 14–17, 2022.

Funding: This study was funded by the Japanese Foundation for Multidisciplinary Treatment of Cancer and the divisional funding.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-22-2683/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-22-2683/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-22-2683/prf

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-22-2683/coif). The authors have no conflicts of interest to declare.

Ethical Statement:The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Institutional Review Board of the Cancer Institute Hospital (No. 2021-GA-1064) on June 28, 2021, and the patient consent requirement was waived because of its retrospective nature.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Mok TS, Wu YL, Thongprasert S, et al. Gefitinib or carboplatin-paclitaxel in pulmonary adenocarcinoma. N Engl J Med 2009;361:947-57. [Crossref] [PubMed]
2. Zhou C, Wu YL, Chen G, et al. Erlotinib versus chemotherapy as first-line treatment for patients with advanced EGFR mutation-positive non-small-cell lung cancer (OPTIMAL, CTONG-0802): a multicentre, open-label, randomised, phase 3 study. Lancet Oncol 2011;12:735-42. [Crossref] [PubMed]
3. Ramalingam SS, Vansteenkiste J, Planchard D, et al. Overall Survival with Osimertinib in Untreated, EGFR-Mutated Advanced NSCLC. N Engl J Med 2020;382:41-50. [Crossref] [PubMed]
4. Sequist LV, Yang JC, Yamamoto N, et al. Phase III study of afatinib or cisplatin plus pemetrexed in patients with metastatic lung adenocarcinoma with EGFR mutations. J Clin Oncol 2013;31:3327-34. [Crossref] [PubMed]
5. Wu YL, Cheng Y, Zhou X, et al. Dacomitinib versus gefitinib as first-line treatment for patients with EGFR-mutation-positive non-small-cell lung cancer (ARCHER 1050): a randomised, open-label, phase 3 trial. Lancet Oncol 2017;18:1454-66. [Crossref] [PubMed]
6. Han B, Tjulandin S, Hagiwara K, et al. EGFR mutation prevalence in Asia-Pacific and Russian patients with advanced NSCLC of adenocarcinoma and non-adenocarcinoma histology: The IGNITE study. Lung Cancer 2017;113:37-44. [Crossref] [PubMed]
7. Wu YL, Tsuboi M, He J, et al. Osimertinib in Resected EGFR-Mutated Non-Small-Cell Lung Cancer. N Engl J Med 2020;383:1711-23. [Crossref] [PubMed]
8. Tsuboi M, Weder W, Escriu C, et al. Neoadjuvant osimertinib with/without chemotherapy versus chemotherapy alone for EGFR-mutated resectable non-small-cell lung cancer: NeoADAURA. Future Oncol 2021;17:4045-55. [Crossref] [PubMed]
9. Rivera MP, Mehta AC, Wahidi MM. Establishing the diagnosis of lung cancer: Diagnosis and management of lung cancer, 3rd ed: American College of Chest Physicians evidence-based clinical practice guidelines. Chest 2013;143:e142S-65S.
10. Silvestri GA, Gonzalez AV, Jantz MA, et al. Methods for staging non-small cell lung cancer: Diagnosis and management of lung cancer, 3rd ed: American College of Chest Physicians evidence-based clinical practice guidelines. Chest 2013;143:e211S-50S.
11. Aye PS, Tin Tin S, McKeage MJ, et al. Development and validation of a predictive model for estimating EGFR mutation probabilities in patients with non-squamous non-small cell lung cancer in New Zealand. BMC Cancer 2020;20:658. [Crossref] [PubMed]
12. Lee G, Lee HY, Park H, et al. Radiomics and its emerging role in lung cancer research, imaging biomarkers and clinical management: State of the art. Eur J Radiol 2017;86:297-307. [Crossref] [PubMed]
13. van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
14. Kakino R, Nakamura M, Mitsuyoshi T, et al. Application and limitation of radiomics approach to prognostic prediction for lung stereotactic body radiotherapy using breath-hold CT images with random survival forest: A multi-institutional study. Med Phys 2020;47:4634-43. [Crossref] [PubMed]
15. Nakanishi R, Akiyoshi T, Toda S, et al. Radiomics Approach Outperforms Diameter Criteria for Predicting Pathological Lateral Lymph Node Metastasis After Neoadjuvant (Chemo)Radiotherapy in Advanced Low Rectal Cancer. Ann Surg Oncol 2020;27:4273-83. [Crossref] [PubMed]
16. Murakami Y, Soyano T, Kozuka T, et al. Dose-Based Radiomic Analysis (Dosiomics) for Intensity Modulated Radiation Therapy in Patients With Prostate Cancer: Correlation Between Planned Dose Distribution and Biochemical Failure. Int J Radiat Oncol Biol Phys 2022;112:247-59. [Crossref] [PubMed]
17. Kursa MB, Rudnicki WR. Feature Selection with the Boruta Package. J Stat Softw 2010;36:1-13. [Crossref]
18. Otani S, Himoto Y, Nishio M, et al. Radiomic machine learning for pretreatment assessment of prognostic risk factors for endometrial cancer and its effects on radiologists' decisions of deep myometrial invasion. Magn Reson Imaging 2022;85:161-7. [Crossref] [PubMed]
19. Kursa MB. Robustness of Random Forest-based gene selection methods. BMC Bioinformatics 2014;15:8. [Crossref] [PubMed]
20. AkibaTSanoSYanaseTA Next-generation Hyperparameter Optimization Framework. arXiv:1907.10902v1. 2019.
21. Zhang YL, Yuan JQ, Wang KF, et al. The prevalence of EGFR mutation in patients with non-small cell lung cancer: a systematic review and meta-analysis. Oncotarget 2016;7:78985-93. [Crossref] [PubMed]
22. Collins GS, Reitsma JB, Altman DG, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ 2015;350:g7594. [Crossref] [PubMed]
23. Cheng Z, Shan F, Yang Y, et al. CT characteristics of non-small cell lung cancer with epidermal growth factor receptor mutation: a systematic review and meta-analysis. BMC Med Imaging 2017;17:5. [Crossref] [PubMed]
24. Huang Y, Liu Z, He L, et al. Radiomics Signature: A Potential Biomarker for the Prediction of Disease-Free Survival in Early-Stage (I or II) Non-Small Cell Lung Cancer. Radiology 2016;281:947-57. [Crossref] [PubMed]
25. Rossi G, Barabino E, Fedeli A, et al. Radiomic Detection of EGFR Mutations in NSCLC. Cancer Res 2021;81:724-31. [Crossref] [PubMed]
26. Wang S, Shi J, Ye Z, et al. Predicting EGFR mutation status in lung adenocarcinoma on computed tomography image using deep learning. Eur Respir J 2019;53:1800986. [Crossref] [PubMed]
27. Liu Y, Kim J, Balagurunathan Y, et al. Radiomic Features Are Associated With EGFR Mutation Status in Lung Adenocarcinomas. Clin Lung Cancer 2016;17:441-448.e6. [Crossref] [PubMed]
28. Hong D, Xu K, Zhang L, et al. Radiomics Signature as a Predictive Factor for EGFR Mutations in Advanced Lung Adenocarcinoma. Front Oncol 2020;10:28. [Crossref] [PubMed]
29. Jia TY, Xiong JF, Li XY, et al. Identifying EGFR mutations in lung adenocarcinoma by noninvasive imaging using radiomics features and random forest modeling. Eur Radiol 2019;29:4742-50. [Crossref] [PubMed]
30. Zhong WZ, Wang Q, Mao WM, et al. Gefitinib Versus Vinorelbine Plus Cisplatin as Adjuvant Treatment for Stage II-IIIA (N1-N2) EGFR-Mutant NSCLC: Final Overall Survival Analysis of CTONG1104 Phase III Trial. J Clin Oncol 2021;39:713-22. [Crossref] [PubMed]
31. Felip E, Altorki N, Zhou C, et al. Adjuvant atezolizumab after adjuvant chemotherapy in resected stage IB-IIIA non-small-cell lung cancer (IMpower010): a randomised, multicentre, open-label, phase 3 trial. Lancet 2021;398:1344-57. [Crossref] [PubMed]
32. Gainor JF, Shaw AT, Sequist LV, et al. EGFR Mutations and ALK Rearrangements Are Associated with Low Response Rates to PD-1 Pathway Blockade in Non-Small Cell Lung Cancer: A Retrospective Analysis. Clin Cancer Res 2016;22:4585-93. [Crossref] [PubMed]
33. Rolfo C, Mack PC, Scagliotti GV, et al. Liquid Biopsy for Advanced Non-Small Cell Lung Cancer (NSCLC): A Statement Paper from the IASLC. J Thorac Oncol 2018;13:1248-68. [Crossref] [PubMed]
34. Malapelle U, Sirera R, Jantus-Lewintre E, et al. Profile of the Roche cobas® EGFR mutation test v2 for non-small cell lung cancer. Expert Rev Mol Diagn 2017;17:209-15. [Crossref] [PubMed]
35. Merker JD, Oxnard GR, Compton C, et al. Circulating Tumor DNA Analysis in Patients With Cancer: American Society of Clinical Oncology and College of American Pathologists Joint Review. J Clin Oncol 2018;36:1631-41. [Crossref] [PubMed]
36. Asano F, Aoe M, Ohsaki Y, et al. Deaths and complications associated with respiratory endoscopy: a survey by the Japan Society for Respiratory Endoscopy in 2010. Respirology 2012;17:478-85. [Crossref] [PubMed]
37. Tomiyama N, Yasuhara Y, Nakajima Y, et al. CT-guided needle biopsy of lung lesions: a survey of severe complication based on 9783 biopsies in Japan. Eur J Radiol 2006;59:60-4. [Crossref] [PubMed]
38. Ponholzer F, Kroepfl V, Ng C, et al. Delay to surgical treatment in lung cancer patients and its impact on survival in a video-assisted thoracoscopic lobectomy cohort. Sci Rep 2021;11:4914. [Crossref] [PubMed]
39. Koukourakis MI, Giatromanolaki A, Thorpe PE, et al. Vascular endothelial growth factor/KDR activated microvessel density versus CD31 standard microvessel density in non-small cell lung cancer. Cancer Res 2000;60:3088-95. [PubMed]
40. Shi Y, Au JS, Thongprasert S, et al. A prospective, molecular epidemiology study of EGFR mutations in Asian patients with advanced non-small-cell lung cancer of adenocarcinoma histology (PIONEER). J Thorac Oncol 2014;9:154-62. [Crossref] [PubMed]