Current progress and quality of radiomic studies for predicting EGFR mutation in patients with non-small cell lung cancer using PET/CT images: a systematic review

Abstract

Objectives:

To assess the methodological quality of radiomic studies based on positron emission tomography/computed tomography (PET/CT) images predicting epidermal growth factor receptor (EGFR) mutation status in patients with non-small cell lung cancer (NSCLC).

Methods:

We systematically searched for eligible studies in the PubMed and Web of Science datasets using the terms “radiomics”, “PET/CT”, “NSCLC”, and “EGFR”. The included studies were screened by two reviewers independently. The quality of the radiomic workflow of studies was assessed using the Radiomics Quality Score (RQS). Interclass correlation coefficient (ICC) was used to determine inter rater agreement for the RQS. An overview of the methodologies used in steps of the radiomics workflow and current results are presented.

Results:

Six studies were included with sample sizes of 973 ranging from 115 to 248 patients. Methodologies in the radiomic workflow varied greatly. The first-order statistics were the most reproducible features. The RQS scores varied from 13.9 to 47.2%. All studies were scored below 50% due to defects on multiple segmentations, phantom study on all scanners, imaging at multiple time points, cut-off analyses, calibration statistics, prospective study, potential clinical utility, and cost-effectiveness analysis. The ICC results for majority of RQS items were excellent. The ICC for summed RQS was 0.986 [95% confidence interval (CI): 0.898–0.998].

Conclusions:

The PET/CT-based radiomics signature could serve as a diagnostic indicator of EGFR mutation status in NSCLC patients. However, the current conclusions should be interpreted with care due to the suboptimal quality of the studies. Consensus for standardization of PET/CT-based radiomic workflow for EGFR mutation status in NSCLC patients is warranted to further improve research.

Advances in knowledge:

Radiomics can offer clinicians better insight into the prediction of EGFR mutation status in NSCLC patients, whereas the quality of relative studies should be improved before application to the clinical setting.

Introduction

Lung cancer is the leading cause of cancer-related deaths around the world, and more than 26% of cancer deaths would be caused by lung cancer. Non-small cell lung cancer (NSCLC) accounts for over 85% of lung cancer, of which, lung adenocarcinoma being the most common pathological type. ¹ In recent years, targeted therapy has attracted extensive attention in NSCLC due to the development of molecular biology. In particular, epidermal growth factor receptor (EGFR) targeting tyrosine kinase inhibitors (TKIs) has been widely used in the treatment of NSCLC. ^2,3 Targeting EGFR is a promising strategy to effectively improve the prognosis of patients with lung adenocarcinoma. Thus, it is of great importance to identify whether a patient has EGFR mutations before TKIs targeted therapy.

Currently, tumor tissue or cytological specimens are commonly used in clinical practice for gene testing, however, they are mostly invasive and unavailable to a number of NSCLC patients due to technical challenges, medical cost, patient health status, and tumor heterogeneity. ⁴ Accordingly, it is necessary to develop non-invasive and accurate methods to detect gene mutations. Some studies used ¹⁸F-FDG PET/CT to predict the EGFR mutation status, usually based on visual analysis or conventional quantitative indicators such as standard uptake value (SUV), leading to unsatisfactory predictive ability. ^5–7

Recent progress in EGFR mutation prediction demonstrate promising results in objectively interpreting PET/CT images using quantitative imaging analysis (QIA), such as an emerging field of radiomics. ^8,9 Radiomics refers to a high-throughput QIA method that extracts a huge number of quantitative features from standard-of-care medical images. ¹⁰ It is hypothesized that the features capture underlying information invisible to the naked eyes in which the intratumoral heterogeneity of the entire lung tumor can be detected and quantified. ^11,12 Recently, ¹⁸F-FDG PET/CT radiomics, which extracts textural features from PET/CT images, has demonstrated certain potential in predicting EGFR mutation. ^13–18

The radiomic workflow consists of five phases: data selection, medical imaging, feature extraction, exploratory analysis, and modeling. ¹⁹ The methodology used in all steps would affect the quality of a radiomic work. Various methodologies may lead to heterogeneous findings that make the comparison across studies difficult. To the best of our knowledge, the variation in methodologies applied in PET/CT-based radiomics for predicting EGFR mutation status in NSCLC patients has not yet been evaluated. This systematic review aims to assess the quality of the radiomic workflow using the Radiomics Quality Score (RQS) as proposed by Lambin et al ²⁰ and to report on the different methodologies used in all steps of radiomic workflow. The RQS assigns points to all steps in the radiomics workflow where the maximum number of points per item depends on the relevance to quality. In addition, we summarize and briefly evaluate the current results reported on the topic of PET/CT-based radiomics for the prediction of EGFR mutation in NSCLC patients. We intended to promote the quality of reporting of radiomics in NSCLC studies and increase the reliability of radiomics for the EGFR detection in the clinical setting.

Methods and materials

Literature search

For this systematic review, a structured search using the PubMed and Web of Science datasets from the conception to October 4, 2020, were performed. Two researchers with 10 and 15 years’ experience in lung cancer diagnosis independently conducted the literature search to identify the eligible studies. The search strategies are described in Supplementary Material 1.

Inclusion criteria and study selection

Studies were included if they met the following criteria: (1) The cohort consists of pathology-confirmed NSCLC patients who underwent pre-treatment ¹⁸F-FDG PET/CT scans before surgery or biopsy; (2) patients received no anti tumor prior to PET/CT; (3) patients had gene testing for EGFR mutation; (4) PET/CT-based radiomic analysis was used to predict the EGFR mutation status; and 5) studies were reported in English with available full-text. We excluded reviews, meta-analysis, case reports, conference abstracts, letters to the editor, comments, technical reports, and duplicate studies. Full details of screening and eligibility assessment are provided in Figure 1.

Figure 1.

Flow diagram of the study selection process

Two independent investigators first screened the titles and abstracts to determine whether the citation met the eligibility criteria. They screened the full text for potentially relevant radiomic studies when both agreed that a citation met the eligibility criteria. Disagreements between the investigators were resolved by consensus and, if necessary, consultation with a third investigator. For each included study, reference lists were searched manually for additional eligible studies.

Data extraction

The following data were extracted (if available): the first author, year of publication, study design, sample size, tumor stage, the proportion of EGFR-mutant patients, PET/CT imaging parameters, image processing, features extraction platform, number of features, the region of interest (ROI) segmentation method, feature selection method, modeling method, and prediction performance in terms of area under the curve (AUC).

Assessment of the quality of radiomic studies

To quantify the radiomic workflow quality and its reporting, the included studies were evaluated in consensus by the two initial reviewers using the RQS. This checklist is comprised of 16 components in the radiomic workflow, with a minimum score of −8 to 0 defined as 0% and a maximum score of 36 defined as 100%. ²⁰ The selected studies were assessed for (1) Image protocol quality; (2) Multiple segmentations; (3) Phantom study on all scanners; (4) Imaging at multiple points; (5) Feature reduction or adjustment for multiple testing; (6) Multivariable analysis with non-radiomic features; (7) Detect and discuss biological correlates; (8) Cut-off analyses; (9) Discrimination statistics; (10) Calibration statistics; (11) Prospective study registered in a trial database; (12) Validation; (13) Comparison to “gold standard”; (14) Potential clinical utility; (15) Cost-effectiveness analysis; and (16) Open science and data. To reflect the importance of prospective studies, seven points are assigned when these are performed and published. Another important step is external validation on at least three datasets, which yielded five points, but if no external validation is performed at all, five points are deducted. Most items can receive zero, one, or two point(s). The detailed checklist is described in Supplementary Table 1.

Statistical analysis

Statistical analysis was performed with SPSS v.23.0 software (IBM Corporation, Armonk, NY). The summed RQS rating per study was calculated and the average rating of all raters is reported. The interclass correlation coefficient (ICC) was determined to describe inter rater agreement for the RQS.

Results

Study selection

A total of 251 records were identified through the searches in PubMed (21 records) and Web of Science (230 records). After removing duplicates, the titles and abstracts of records were screened, resulting in 230 records eligible for inclusion. The full text of these studies was read and the selection criteria were applied. Finally, a total of six studies were included for this systematic review.

Characteristics of the included studies

A total of 973 NSCLC patients were retrospectively included in the six studies presented in this review (Table 1). All studies were recently published, between 2019 and 2020. Sample size ranged from 115 to 248, with a largest validation cohort of 73 patients. Most studies were performed at a single center, only two were conducted in two centers. All studies reported the detection method of EGFR mutation with the exception of Koyasu et al ¹⁸ . The percentages of EGFR mutation ranged from 22.5 to 62.6%.

Table 1.

Characteristics of included studies

First author	Zhang et al 16	Shiri et al 15	Liu et al 14	Koyasu et al 18	Yang et al 13	Li et al 17
Publication date	2019	2020	2020	2019	2020	2019
Study design	Retrospective	Retrospective	Retrospective	Retrospective	Retrospective	Retrospective
Sample size	248	150	148	138	174	115
Single center or multicenter	Single center	Unclear	Two centers	Two centers	Single center	Single center
PET/CT manufacturer	PHILIPS	GE and Siemens	Siemens	Unclear	Siemens	GE
Image processing	Unclear	Yes	Yes	No	Yes	No
ROI segmentation method	Semi-automatically	PET, manually CT, semi-automatically	Manually	Semi-automatically	Semi-automatically	Semi-automatically
Feature classes	FOS, Shape, GLCM, GLZLM, GLRLM, NGLDM	FOS, GLCM, GLRLM, Shape, GLSZM, GLDM, NGTDM	FOS, Shape, GLCM, GLRLM, GLSZM, GLDM, NGTDM	Histogram, GLCM, LBP	FOS, Shape, GLCM, GLDM, GLRLM, GLSZM, NGTDM	morphological features, grayscale statistic features, GLCM, GGCM, GLDS
Feature extraction software	LIFEx package	Python with package Pyradiomics	Python 3.6.2 with package Pyradiomics	Unclear	Radiomics prototype (Radiomics, Frontier, Siemens)	MATLAB
Feature selection method	LASSO	SKB, VT, SP, SM, VT-SM, VT-KB	VIF, RF, LR	Unclear	LASSO	LASSO
No. of total features per patient	92 (47 from PET and 45 from CT)	109 from PET	1570 (1470 from CT and 100 from PET)	261 (173 from CT and 88 from PET)	3344 (1672 from CT and 1672 from PET)	38
No. of selected features	9	3	10	6	11	9
Modeling algorithm	LR	SVM, KNN, DT, QDA, MLP, SGD, LR, NB, GNB, RF, AB, BAG	XGBoost	RF and XGBoost	RF	Ensembles for boosting machine learning scheme
No. EGFR mutation (%)	53.6	22.5	50.6	27.5	62.6	56.0
Statistical findings: radiomic model performance	The model showed a significant ability to discriminate between EGFR mutation and EGFR wild type, with AUC of 0.79 in the training set, and 0.85 in the validation set, compared with 0.75 and 0.69 for the clinical model. When clinical variables and radiomics signature were combined, the AUC increased to 0.86 (95% CI: 0.80–0.91) in the training set and 0.87 (95% CI: 0.79–0.95) in the validation set, thus showing better performance in the prediction of EGFR mutations.	Multivariate machine learning-based AUC performances were significantly improved to 0.82 for EGFR (LOG preprocessed image of PET with sigma three with variance threshold feature selector and stochastic gradient descent classifier (q-value = 4.86E-05). .	The corresponding radiomic predictors achieved the prediction accuracies of 0.77 and 0.92 in terms of AUC, respectively. Combing these two predictors, the overall model for predicting EGFR mutation positivity was also constructed, and the AUC was 0.87.	In the classification of EGFR mutation status, the AUC values were: RF, single type: 0.625; XGBoost, single type: 0.617; RF, multiple types: 0.577; XGBoost, multiple types: 0.659.	The mutant/wild-type model was identified in the training (AUC, 0.77) and validation (AUC, 0.71) groups. The 19/21 mutation site model had an AUC of 0.82 and 0.73 in the training and validation groups, respectively.	A radiomic signature based on both PET and CT radiomic features outperformed individual radiomic features, the PET or CT radiomic signature, and the conventional PET parameters including the maximum standardized uptake value (SUVmax), SUVmean, SUVpeak, metabolic tumor volume (MTV), and total lesion glycolysis (TLG), in discriminating between mutant-type of EGFR and wild-type of EGFR- cases with an AUC of 0.805, an accuracy of 80.798%, a sensitivity of 0.826 and a specificity of 0.783.
Conclusions	The PET/CT-based radiomic features showed good performance in predicting EGFR mutation in NSCLC, providing a useful method for the choice of targeted therapy in a clinical setting.	Our work demonstrated that non-invasive and reliable radiomics analysis can be successfully used to predict EGFR and KRAS mutation status in NSCLC patients. We demonstrated that radiomic features extracted from different image-feature sets could be used for EGFR and KRAS mutation status prediction in NSCLC patients and showed improved predictive power relative to conventional image-derived metrics.	We established predictive models based on radiomic analysis of 18F-FDG PET/CT images. And it achieved a satisfying prediction power in the identification of EGFR mutation status as well as the certain EGFR mutation subtypes in lung cancer.	The radiomics approach to PET/CT images, together with XGBoost and Bayesian optimization, is useful for classifying histological subtypes and EGFR mutation status in lung cancer.	Radiomic features based on (18F)-FDG PET/CT combined with clinicopathological factors could reflect genetic differences and predict EGFR mutation type.	Tumor imaging phenotypes that are driven by somatic mutations may be predicted by radiomics based on PET/CT images.

AB, Adaptive boosting; AUC, area under the curve; BAG, Bagging; CI, Confidence interval; DT, Decision tree; EGFR, Epidermal growth factor receptor;FOS, First-order statistics; GLCM, Gray-level co-occurrence matrix; GLDM, Gray-level dependence matrix; GLRLM, Gray-level run length matrix; GLSZM, Gray-Level Size Zone Matrix; GNB, Gaussian naive Bayes; KNN, K-nearest neighbors; KRAS, Kirsten rat sarcoma viral oncogene; LASSO, Least absolute shrinkage and selection operator; LR, Logistic regression; MLP, Multilayer perceptron; NB, Naive Bayes; NGTDM, Neighboring gray tone difference matrix; NSCLC, Non-small cell lung cancer; PET/CT, Positron emission tomography/Computed tomography; QDA, Quadratic discriminant analysis; RF, Random forest; ROI, Region of interest; SGD, Stochastic gradient descent; SKB, Select K best; SM, Select from model; SP, Select percentile; SVM, Support vector machine; VIF, Variance inflation factor; VT, Variance threshold; VT-KB, Variance threshold and select K best; VT-SM, Variance threshold and select from model.

Quality analysis of the included studies

The RQS heatmap (Figure 2) presents the score of each item and the total score for each study. The mean RQS of all studies evaluated by two raters was 12.7 (35.2%) points, ranging from 5 (13.9%) to 17 (47.2%) points. All studies were scored below 50% due to defects on multiple segmentations, phantom study on all scanners, imaging at multiple time points, cut-off analyses, calibration statistics, prospective study registered in a trial database, potential clinical utility, and cost-effectiveness analysis. The ICC results for the assessment of major RQS items were good (Table 2). The ICC for summed RQS between two raters was 0.986 (95% CI: 0.898–0.998), demonstrating good agreement (Table 2).

Figure 2.

Heat map of radiomics quality score (RQS) per component of all included studies.

Table 2.

Average rating and interrater agreement per component of QRS

No.	RQS scoring item	Range	Average	ICC (95% CI)
1	Image protocol quality	0 to 2	1.17	1.00 (1.00–1.00)
2	Multiple segmentations	0 to 1	0	1.00 (1.00–1.00)
3	Phantom study on all scanners	0 to 1	0	1.00 (1.00–1.00)
4	Imaging at multiple time points	0 to 1	0	1.00 (1.00–1.00)
5	Feature reduction or adjustment for multiple testing	−3 to 3	2.50	1.00 (1.00–1.00)
6	Multivariable analysis with non-radiomics features	0 to 1	0.58	0.83 (-0.23–0.98)
7	Detect and discuss biological correlates	0 to 1	0.50	0.33 (-3.76–0.91)
8	Cut-off analyses	0 to 1	0	1.00 (1.00–1.00)
9	Discrimination statistics	0 to 2	1.67	0.60 (-1.86–−0.94)
10	Calibration statistics	0 to 2	0.33	1.00 (1.00–1.00)
11	Prospective study registered in a trial database	0 to 7	0	1.00 (1.00–1.00)
12	Validation	−5 to 5	1.67	1.00 (1.00–1.00)
13	Comparison to “gold standard”	0 to 2	2.00	1.00 (1.00–1.00)
14	Potential clinical utility	0 to 2	0	1.00 (1.00–1.00)
15	Cost-effectiveness analysis	0 to 1	0	1.00 (1.00–1.00)
16	Open science and data	0 to 4	2.25	0.76 (-0.70–0.97)
	Total points: −8 to 0 = 0%, 36 = 100%	0–100%

CI, Confidence interval; ICC, Interclass correlation coefficient; RQS, Radiomics quality score.

Radiomics workflow

1. Image acquisition and pre-processing: One study established the imaging acquisition protocol following the Image Biomarker Standardization Initiative (IBSI) reporting guidelines. ¹⁶ The authors of all studies performed their analyses on the pre-treatment PET/CT images. None of the studies analyzed the differences between PET/CT exams at least two time points. Various PET/CT manufacturers were used, including GE, PHILIPS, Siemens, or both. Half of the studies reported on image pre-processing performed by interpolation (sitkBSpline algorithm, B-spline of order three interpolation), Wavelet transform +Laplacian of Gaussian (Fine, Median, Coarse)+Bin discretization, resample +discretization + rescale to reduce the bias caused by different scanners and imaging protocols. None of the studies performed feature normalization.

2. Volume of interest segmentation: All studies performed three dimensional (3D) segmentation, four performed semi-automatically, one manually, and one both semi-automatically (CT images) and manually (PET images). For semi-automatic segmentation by methods including region growing or drawing a line along the boundary of the tumor with a Radiomics prototype, the region of interest (or tumor) used a threshold of 40–42% of the maximum standardized uptake value (SUVmax). The manual segmentations were conducted with OSIRIX^® or ITK-SNAP or ImageJ 1.50i software. The radiologists were blinded to the pathologic and EGFR mutation test results.

3. Feature extraction and selection: The number of features extracted from 38 to 3344, mostly extracted with in-house built software, such as LIFEx, Pyradiomics, and MATLAB. The studies extracted similar classes of features: first-order statistics, shape-based features, and texture features. Feature selection only applied to the studies developing multivariate prediction models, where these studies opted for a different approach, with the least absolute shrinkage and selection operator (LASSO) most often. Two studies used more than one feature selection method. ^14,15 The study of Koyasu et al ¹⁸ did not report on the feature selection method. The number of features included in the models varied between 3 and 11 features (Supplementary Table 2). Figure 3 shows the overlapped features among the included features. The first-order statistics (e.g. maximum and median) were the most reproducible features that significantly related to EGFR mutations.

4. Radiomics analysis and modeling: All studies presented individual features significant for EGFR mutation status. Three studies developed clinical, radiomic, and combined models. ^13,16,17 The clinical models usually consisted of sex and smoking history, other variables included age, tumor stage, and location. The combined models were superior to clinical model and radiomic model alone. Two studies analyzed the predictive value of PET/CT metabolic parameters alone or together with radiomic signature and/or clinical variables. ^13,15 One study compared the predictive value of radiomic features derived from PET/CT, PET, and CT, respectively, and found that the combined PET/CT radiomic features outperformed radiomics signature based on PET alone and CT alone. ¹⁷ All studies predicted general EGFR mutation except one study ¹⁴ predicted EGFR mutation subtypes (EGFR exon 19 deletions and exon 21 L858R missense). Three studies ^15,17,18 compared at least two machine learning methods in predicting EGFR mutation status, especially Shiri et al ¹⁵ . To validate the models, the majority of studies performed 10-fold cross-validation. All studies divided their data into a training and a validation cohort. No study externally validated their models.

5. Prediction performance evaluation: The prediction performance of models was distinctly different. The AUCs results of the multivariate analysis in differentiating EGFR mutation status, for the validation cohort, ranged from 0.659 to 0.870. The performance of models based on clinical or PET/CT or radiomic signature alone was relatively low and could be significantly improved when incorporated, highlighting the necessity of combining multisource data to develop prediction models prior to clinical application.

Figure 3.

The overlap features among the included studies.

Discussion

The systematic review provides an overview of the recent literature on PET/CT-based radiomics for EGFR mutation status in NSCLC patients. Overall, the studies indicated large methodologies heterogeneity in each step of radiomic workflow, such as differences in image acquisition, feature extraction, feature selection, and modeling strategy. Despite the radiomic workflow heterogeneities, the results seem promising. The AUCs of the models in differentiating EGFR mutation status, for the validation cohort, ranged from 0.659 to 0.870.

¹⁸F-FDG PET/CT cannot only depict anatomical structure, but also reflect the biological characteristics of tumors. However, the PET/CT parameters in predicting EGFR mutation status remain controversial. ^21–24 Shiri et al performed a univariate analysis of conventional PET metrics for EGFR mutation prediction showed that SUV_peak had the highest value, with an AUC of 0.69, whereas the AUCs of other parameters were below 0.60. ¹⁵ Li et al ¹⁷ found that the AUCs of conventional PET parameters to predict EGFR mutational status in NSCLC ranged from 0.615 to 0.621. Previous studies demonstrated that the predictive performance of radiomic features extracted from PET/CT images was significantly higher than conventional PET/CT parameters. ^13,15 The addition of clinical variables and radiomic features into PET/CT parameters could further improve the discrimination of models, achieving an AUC of around 0.80 in the validation cohorts. ^13,16,17 The radiomic features incorporated in the prediction models greatly varied across studies, the overall robustness of features was relatively poor, however, the susceptibility of those features of variation was significantly different. The findings showed that first-order features (e.g. maximum, median) were potentially more related to EGFR mutations in NSCLC. Variations in scanners, imaging protocols, segmentation strategies, feature selection, and feature-intrinsic characteristics pose a threat to robustness and reproducibility of features. Identification of robust radiomics features may improve the generalizability of radiomic models in future studies and fasten the clinical application.

In this review, we applied the RQS checklist to assess the quality of included studies and reported the radiomic workflow. The inter rater agreement was good for most items, as indicated by some previous reviews. ^25,26 The studies employing multivariate analysis achieved a mean RQS of 12.7 (35.2%) points, ranging from 5 (13.9%) to 17 (47.2%) points. The RQS is largely affected by the items prospective study registered in a trial database and validation because they account for 7 and 5 points, respectively (i.e. 33.3% of the maximum score). In this review, all studies were retrospective and internally validated. Besides, none of the studies conducted multiple segmentations, phantom study, imaging at multiple time points, cut-off analyses, clinical utility, and cost-effectiveness analysis. Some previous systematic reviews have evaluated the quality of radiomic studies across various tumors (e.g. bone, soft-tissues and visceral sarcomas, prostate cancer, NSCLC, lymphoma, renal cell carcinoma, neuro-oncology, breast cancer, rectal cancer, liver cancer, head and neck cancer). ^25–40 The authors found that the mean RQS scores of all studies were less than 50%, ranging from 9.4 to 41.2%. Substantial quality limitations were found in study design, radiomic workflow methodology, validation, prospective study, and open science and data. Most of radiomic studies were retrospective and performed in single center or two centers. Few studies shared their code, images, and features. As radiomics is still a very young field, radiomic research has not yet been translated into clinical use. Current promising radiomic models can be externally validated to demonstrate the generalizability of findings and their impact assessed within the clinical pathway prior to they can be implemented as a clinical decision-making tool to facilitate targeted treatment for patients with NSCLC.

Some limitations in this systematic review should be acknowledged. Firstly, PET/CT-based radiomics studies for the prediction of EGFR mutation status are limited. Hence, only six studies with approximately 300 patients were analyzed. Secondly, the RQS has limitations, several components are not always applicable to every study, for example, the clinical utility for pre-clinical studies or “comparison to gold standard” for outcomes in which no gold standard exists for the assessment of outcome (e.g. intratumoral heterogeneity). Other components are not described in detail in the scoring table (e.g. what should be specifically be mentioned on clinical utility) or are relatively ambiguous (e.g. in case that only a subgroup of features has been extracted that does not feature selection in strict sense). Moreover, the interpretability of the RQS score is left to a large extent to the raters, since it is not meant to report on the overall quality of a manuscript or the research presents, but rather to guide the reader on a radiomic workflow appraisal. Therefore, widespread application of RQS, pre-trained RQS scoring procedure, and modification of RQS in response to actual clinical practice needs are necessary considering radiomics is a developing field. Thirdly, a meta-analysis could not be performed because most studies did not report the necessary data. In addition, the heterogeneity across studies were substantial, so it would be better to conduct systemic review. However, a network meta-analysis may relieve the limitation caused by missing data and increase the possibility of conducting a meta-analysis of radiomic studies on the same topic. Thus, future works can use this network meta-analysis approach to accelerate the clinical translation of predicting EGFR mutation using radiomics. Finally, this review did not investigate the role of CT- or PET-based radiomics for EGFR detection or survival prediction in patients with EGFR mutations. ^41–43 However, radiomics using PET/CT may provide more information than that using CT or PET alone.

In conclusion, the quality of PET/CT-based radiomic studies for predicting EGFR mutation status in NSCLC was somewhat poor. Methodologies in the radiomic workflow varied greatly. Although these scores are relatively low, and radiomics has not yet reached clinical utility in NSCLC, it is important to underscore the fact that these early studies pave the way for the radiomic field with a focus on NSCLC. Further efforts are needed to make PET/CT-based radiomic studies in NSCLC useful and reproducible with an acceptable level of evidence. A better knowledge of the RQS and the reporting guidelines of radiomics could improve the quality of NSCLC radiomic studies. Increasing education about the particular radiomics’ methodology and quality metrics are necessary to fasten and accelerate clinical applications. Overall, our review stresses the urgent need for improving the quality and reproducibility of radiomic studies. This review will be beneficial for the construct of a more consistent or reproducible research in the field of radiomics.