Predicting Kirsten Rat Sarcoma Virus Gene Mutation Status in Patients With Colorectal Cancer by Radiomics Models Based on Multiphasic CT

Objective To develop and validate radiomics models based on multiphasic CT in predicting Kirsten rat sarcoma virus (KRAS) gene mutation status in patients with colorectal cancer (CRC). Materials and Methods A total of 231 patients with pathologically confirmed CRC were retrospectively enrolled and randomly divided into training(n=184) and test groups(n=47) in a ratio of 4:1. A total of 1316 quantitative radiomics features were extracted from non-contrast phase (NCP), arterial-phase (AP) and venous-phase (VP) CT for each patient. Four steps were applied for feature selection including Spearman correlation analysis, variance threshold, least absolute contraction and selection operator, and multivariate stepwise regression analysis. Clinical and pathological characteristics were also assessed. Subsequently, three classification methods, logistic regression (LR), support vector machine (SVM) and random tree (RT) algorithm, were applied to develop seven groups of prediction models (NCP, AP, VP, AP+VP, AP+VP+NCP, AP&VP, AP&VP&NCP) for KRAS mutation prediction. The performance of these models was evaluated by receiver operating characteristics curve (ROC) analysis. Results Among the three groups of single-phase models, the AP model, developed by LR algorithm, showed the best prediction performance with an AUC value of 0.811 (95% CI:0.685–0.938) in the test cohort. Compared with the single-phase models, the dual-phase (AP+VP) model with the LR algorithm showed better prediction performance (AUC=0.826, 95% CI:0.700-0.952). The performance of multiphasic (AP+VP+NCP) model with the LR algorithm (AUC=0.811, 95%CI: 0.679-0.944) is comparable to the model with the SVM algorithm (AUC=0.811, 95%CI: 0.695-0.918) in the test cohort, but the sensitivity, specificity, and accuracy of the multiphasic (AP+VP+NCP) model with the LR algorithm were 0.810, 0.808, 0.809 respectively, which were highest among these seven groups of prediction models in the test cohort. Conclusion The CT radiomics models have the potential to predict KRAS mutation in patients with CRC; different phases may affect the predictive efficacy of radiomics model, of which arterial-phase CT is more informative. The combination of multiphasic CT images can further improve the performance of radiomics model.


INTRODUCTION
Colorectal cancer (CRC) is the second most common cancer and the fourth-leading cause of cancer death in China (1). Kirsten Rat Sarcoma virus (KRAS) is the most common mutated oncogene in colorectal cancer, about 30%-45% of patients with CRC have mutations in the KRAS, which is one of the high-risk factors that drive distant metastasis of tumor cells (2). Those patients with CRC who have KRAS mutations have no benefit of the antibodytargeted therapies to the epidermal growth factor receptor (EGFR) (3). Hence, KRAS mutational test has been recommended by the National Comprehensive Cancer Network (NCCN) guidelines for patients with suspected or proven metastatic CRCs for guiding targeted therapy (4).
At present, the gold standard for determinate KRAS mutation status is the pathological examination of tumor tissue in clinical practice (4). However, some patients cannot tolerate biopsy due to its invasiveness, and the insufficient quality of biopsy specimens may hinder efficient and robust mutation testing. In addition, tumor tissues have the characteristics of spatial and temporal heterogeneity (5), which makes biopsy samples may not accurately reflect the tumor genotype expression, especially after multiple treatments (6,7). Therefore, it would be meaningful for developing a relatively simple and non-invasive method for identifying KRAS mutational status in patients with CRC.
Some non-invasive methods had been used to predict KRAS mutation status in previous studies, and the most used imaging technique was fluorine-18 fludeoxyglucose (18F-FDG) positron emission tomography (PET)-CT (8)(9)(10)(11). However, the sample size of these studies was generally small and the research results had been conflicting between different studies (10,12).
In recent years, radiomics is an emerging technique that has been widely studied in the early diagnosis, efficacy evaluation, and prognosis prediction of tumors (13)(14)(15)(16). Previous studies indicated that radiomics has shown great prediction performance and clinical potential for predicting the genetic mutations status of glioma (17,18), lung cancer (19,20), and breast cancer (21). In addition, radiomics has been studied in CRC for predicting KRAS mutation (22)(23)(24)(25)(26)(27), but most of these studies have only used portal venous phase CT images for radiomics analysis. It is not yet clear whether the non-contrast phase (NCP), arterial phase (AP), venous phase (VP) CT images can be used to predict KRAS mutation in patients with CRC, and the value of the combination of multiphasic radiomics features has yet to be investigated.
Therefore, the aim of our study was to investigate the performance of CT radiomics analysis based on multiphasic CT imaging for predicting KRAS mutation in patients with CRC.

Patients
The study conformed to the provisions of the Declaration of Helsinki (as revised in 2013). Ethical approval is not required for this study as it is based on information collected as part of routine clinical practice. Informed consent was waived because of the retrospective design. We retrospectively analyzed data from patients who were surgically confirmed to have CRC from January 2014 to December 2018. A total of 231 patients met the inclusion criteria for this study. The inclusion criteria and exclusion criteria were shown in Supplementary Material S1. All patients were randomly divided into training and test groups in a ratio of 4:1. There were 184 cases in the training group (80 cases of KRAS mutant type, 104 cases of wild type) and 47 cases in the test group (21 cases of KRAS mutant type, 26 cases of wild type).
Baseline clinical characteristics, including age, gender, maximum tumor diameter, levels of carcinoembryonic antigen (CEA), carbohydrate antigen-199 (CA199), and carbohydrate antigen-724(CA724)were collected from the medical records. The pathological characteristics of tumor surgical specimens, including tumor TNM stage, tumor location, and tumor differentiation grade (well, moderately, and poorly differentiated) were assessed as well.

Identification of KRAS Mutation Status
All surgically resected specimens were processed conventionally by a trained pathologist. DNA was extracted from the formalinfixed paraffin-embedded (FFPE) tumor tissues by the DNA FFPE Tissue Kit (Xiamen Aide Biological Co., Ltd.). Mutations of KRAS (exons 2, 3, and 4) were analyzed by polymerase chain reaction (PCR) and the amplification refractory mutation system (ARMS) method.

Image Acquisition and Segmentation
All patients underwent contrast-enhanced abdominal and pelvic CT by using 64-detector or 128-detector row spiral CT systems in our hospital. The CT image acquisition settings are described in Supplementary Material S2. All NCP, AP, and VP CT images were retrieved from a picture archiving and communication system (PACS) for image segmentation and analysis except for the portal venous phase CT images.
For the lesion segmentation, the region of interest (ROI) was segmented by using software ITK-SNAP (v.3.8.0; http://www. itksnap.org). Firstly, we manually delineated along the contour of tumors on the largest slices on the VP CT images, excluding the air and feces in the intestinal tract. And the ROIs on the NCP and AP CT images were delineated with reference to that on the VP CT images (Figure 1). The ROI of lesions was manually determined by two radiologists with 3 and 8 years of experience, with unanimous agreement. Radiologists were blind to grouping and genetic test results.

Image Preprocessing and Radiomics Feature Extraction
The original images of all cases and the ROI of lesions were preprocessed by using the AK software (Artificial Intelligence Kit, version 3.3.0, GE healthcare) before radiomics feature extraction. The CT image slice and the ROI was resampled to a uniform pixel dimension size of 1×1×1 mm 3 by using Linear Interpolation and Nearest Neighbour Interpolation ( Figure 1).
All radiomics features were extracted using AK software, the detailed information of these features was available in the documentation for PyRadiomics (https://pyradiomics. readthedocs.io/en/latest/features.html), which followed the IBSI radiomics guidelines. Seven categories of feature parameters, including first-order features, shape features, gray level cooccurrence matrix (GLCM), gray level size zone matrix (GLSZM), gray level run length matrix (GLRLM), neighbourhood gray-tone difference matrix (NGTDM), and gray level dependence matrix (GLDM) were selected for feature extraction (Figure 2). In addition, Wavelet transform, Laplacian of Gaussian (LoG) and Local binary pattern (LBP) were applied to the original image respectively and yielded a corresponding derived image. Ultimately, a total of 1316 quantitative 2D radiomics features was exacted based on the original image and its corresponding derived image.

Radiomics Feature Selection and Radiomics Model Building
All cases in the training cohort were used to train the predictive model, while cases in the test cohorts were used to independently evaluate the model's performance. All radiomics features were imported into the IPMs software (Institute of Precision Medicine  Statistics, version 2.4.2, GE healthcare). Before analyses, the missing values were replaced by the median, and the data were standardized by the Z-score method. For the radiomics feature selection, four steps were performed to select the optimal feature subsets for predicting KRAS mutations. First, we used the Spearman correlation analysis. If the correlation coefficient between feature and gene mutation status is small than 0.1, the feature will be eliminated. Second, the Variance threshold method was used to remove features with a variance value lower than 1. Third, the least absolute shrinkage and selection operator (LASSO) algorithm was performed for eliminating the redundancy, this approach can estimate the regression coefficients for every feature and successively shrink them to avoid inflation of the estimated coefficients, resulting in superior predictive performance (28). Fourth, we used multivariate stepwise regression analysis to select the features which were considered to be associated with KRAS mutations. The p-in and p-out of multivariate stepwise analysis were 0.05 and 0.10, respectively. Finally, logistic regression (LR), support vector machine (SVM), and random tree (RT) algorithm were used to build seven groups of prediction models, including the NCP, AP, VP, AP+VP, AP+VP+NCP, AP&VP and AP&VP&NCP models. 5-fold cross-validation was used to select the model with the best performance in the training process. For the SVM and RF, the hyper-parameters of these models were automatic selected by search method, the detailed information of gamma, C, max depth, min samples split, and n estimators were shown in Table S1.

Validation of the Radiomics Model
The receiver operating characteristic (ROC) curve was employed to evaluate the performance of radiomics models for the prediction of KRAS mutation. The area under the curve (including the 95% confidence interval), sensitivity, specificity, and accuracy were also recorded. The calibration curve and the Hosmer-Lemeshow test were used to evaluate the goodness-of-fit of the radiomics model. Decision curve analysis (DCA) was used to evaluate models' net benefits in different threshold probabilities in the training and test cohort. P<0.05 was considered statistically significant.

Statistical Analysis
The clinical and pathological characteristics were analyzed by SPSS Statistics 25.0 software, and a two-sided p value of less than 0.05 was statistically considered significant. We used independent samples ttest or Mann-Whitney U test to compare the differences in continuous variables between the patients in different groups, including age and maximum tumor diameter. The differences in categorical variables, including sex, tumor stage, tumor location, tumor differentiation grade, levels of CEA, CA199 and CA724, were assessed using chi-squared or Fisher's exact tests.

Clinical and Pathological Characteristics
There were no significant differences in the clinical and pathological characteristics between the training and the test cohort (p = 0.210-0.879, Table S2). The clinical and pathological characteristics in the training and test cohorts are listed in Table 1. There were significant differences in TNM stage and M stage between the mutated group and the wild-type group in the training cohort (P < 0.05), but they were not confirmed in the test cohort. There were no significant differences between the mutated group and the wild-type group in both cohorts in terms of age, gender, maximum tumor diameter, tumor location, tumor differentiation grade, T stage, N stage and CEA, CA199, CA724 levels.

Feature Selection and Radiomics Model Building
A total of 1316 radiomics features were extracted from the ROIs of the NCP, AP, and VP CT images for each patient, respectively. After four steps of feature selection, 5, 6 and 7 optimal radiomics features were selected from each phase CT images, respectively (Table 2; Figure 3). Three groups of single-phase radiomics models were built based on corresponding optimal radiomics features, including the NCP, AP, and VP models. The AP+VP model was built based on 13 (6 + 7) features obtained from the combination of the AP and VP. The AP+VP+NCP model based on 18 (5 + 6+7) features was obtained from the combination of three phases. In addition, we combined the 2632 radiomics features of the AP and VP at first and then four steps of feature selection were implemented, 7 radiomics features were selected to build the AP&VP model. In the same way, 12 radiomics features were selected to build the AP&VP&NCP model.
Finally, seven groups of radiomics models for predicting KRAS mutation were constructed by LR, SVM, and RF classifiers using the above-selected features.

Predictive Performance of Radiomics Model
The results of the seven groups of radiomics model in the training and test cohort were shown in Figure 4 and Tables 3,  4. Among the three groups of single-phase models, the AP model developed by LR classifiers had the best prediction performance, which had the AUCs of the test cohort was 0.811 (95% CI, 0. The Hosmer-Lemeshow test yielded a non-significant p-value ranging from 0.076 to 0.815 in the LR models. suggesting no departure from the perfect fit. The calibration curve of the multiphasic (AP+VP+NCP) model for KRAS mutation prediction probability shows good accordance between prediction and observation in the training and test cohort ( Figure 5). The DCA for the seven groups of radiomics models in the test cohort was presented in Figure 6. The DCA showed that the AP model, dual-phase model, and multiphasic model showed relatively more area, suggesting the good performance of the radiomics models in terms of clinical application.

DISCUSSION
In this study, we built seven groups of radiomics models based on different phase CT images for predicting KRAS mutation in patients with CRC. We found that single-phase models have the potential to predict KRAS mutation, with the AP model developed by LR showing the better predictive performance. The model developed by LR showed similar results as SVM except for the NCP model. The predictive performance of the AP+VP and AP+VP+NCP model was further improved compared to that of the single-phase model, and the AP+VP model showed the best predictive performance, but the AP+VP+NCP model showed better predictive performance comprehensively, showing that combing the different phase CT radiomics features could elevate the model's prediction ability. There were some previous radiomics studies on the KRAS mutation prediction in CRC. Meng et al. (29) found that radiomic signatures based on multiparametric MRI had the potential to predict KRAS mutation in rectal cancer with an AUC of 0.651 (95% CI, 0.539-0.763). Cui et al. (26)reported that T2WIbased radiomics signature had a moderate performance to predict KRAS mutation in rectal cancer with an AUC of 0.714 (95% CI, 0.602-0.827). The above studies showed an encouraging result for predicting KRAS status by using radiomics, but compare with our study, our best model was the dual-phase (AP+VP) model developed by LR with relatively higher predictive performance, which had an AUCs of 0.826 in the test cohort. Furthermore, as the CT examination is convenient for the patients in clinical practice and also recommended by NCCN guidelines on the management of patients with CRC (4), CT images had been usually used as their research object of radiomics in colorectal cancer. Wu et al. (25) reported that the hand-crafted radiomics signature was associated with the KRAS mutation in CRC with the C-index, sensitivity, and specificity were 0.727, 0.412 and 0.868 in the validation cohort, respectively. Yang et al. (24)found that the proposed CT-based radiomics signature was related to KRAS/NRAS/BRAF mutations with the AUC, sensitivity, and specificity were 0.829, 0.686 and 0.857 in the validation cohort, respectively. Although these studies had higher AUC than that obtained in our study, the sensitivity of these models was relatively low, which may be related to the use of single-phase CT in these studies. While in our study, the singlephase models also showed relatively lower sensitivity, but the multiphasic (AP+VP+NCP) model in our study showed satisfactory predictive performance with the AUC, sensitivity, and specificity were all above 0.8 in the test cohort.
Additionally, many studies in this field have only focused on the portal venous phase CT images (22)(23)(24)(25), without investigating the

CT phase Category Names
Non-contrast GLSZM [1]original_glszm_SmallAreaEmphasis GLDM [2]wavelet-HHL_gldm_LargeDependenceEmphasis GLCM [3]wavelet-HLH_glcm_Imc1 GLRLM [4]wavelet-LLL_glrlm_GrayLevelNonUniformityNormalized GLSZM [5]log-sigma-3-0-mm-3D_glszm_GrayLevelNonUniformityNormalized arterial GLRLM [1]wavelet-HHL_glrlm_ShortRunEmphasis GLSZM [2]lbp-3D-k_glszm_SmallAreaEmphasis GLSZM [3]wavelet-HLH_glszm_SmallAreaEmphasis GLDM [4]original_gldm_DependenceVariance GLSZM [5]wavelet-LLL_glszm_SmallAreaEmphasis GLDM [6]wavelet-HLH_gldm_SmallDependenceHighGrayLevelEmphasis venous GLSZM [1]wavelet-LHL_glszm_SizeZoneNonUniformityNormalize First Order [2]lbp-3D-m1_firstorder_Maximum First Order [3]lbp-3D-m2_firstorder_10Percentile GLRLM [4]log-sigma-2-0-mm-3D_glrlm_LongRunEmphasis GLDM [5]wavelet-LLL_gldm_LowGrayLevelEmphasis GLRLM [6]wavelet-LHH_glrlm_GrayLevelVariance GLDM [7]log-sigma-2-0-mm-3D_gldm_LargeDependenceLowGrayLevelEmphasis GLCM, gray-level co-occurrence matrix; GLSZM, gray-level size zone matrix; GLRLM, gray level run length matrix; GLDM, gray level dependence matrix.  We also found that the predictive efficacy of the AP model was better than that of the other single-phase model. It may be that the AP CT image mainly reflects the blood perfusion of the tumor tissue, which may better refl ect the tumor microenvironment. Moreover, the predictive performance of the AP+VP and AP+VP+NCP model was slightly higher than those of the AP&VP and AP&VP&NCP model. This may be because the method of feature selection at first and then combined would guarantee that the optimal features of each CT phase can play a role in the combined model. Choosing a proper classifier can improve the stability and predictive performance of the model. The LR classifier is a linear regression method that had been usually used in many machines learning studies for its good interpretability and suitability to solve dichotomous problem (13,19,25). For SVM, it is a robust and effective classifier based on structural risk minimization that had been proved to be a powerful classifier in the previous studies (24,26). Our previous study had used these two algorithms to build model for distinguishing the solid solitary pulmonary lesion based on T 2 WI images and showed relatively better performance (33). In this study, we could find that the model developed by LR showed similar results as SVM except for the NCP model, and both of these classifiers had moderate predictive performance, it may be that these two classifiers are suitable for solving the problem with a small sample. In addition, although the RF classifier had been showed good performance in other studies (29), which has more hyperparameters and is a relatively complicated model, the RF classifier showed overfitting in the training and test cohort in our study, it may be that our sample size is relatively small.
For the radiomics features selection, 5, 6, and 7 features were selected from the NCP, AP, and VP, respectively, to form radiomics models, which were mainly derived from GLDM, GLRLM, and GLSZM. The three sets of higher-order radiomics features could quantify the image uniformity and heterogeneity, which were found to be correlated with KRAS mutation in CRC. Among the 18 radiomics features in the AP+VP+NCP model, the log-sigmaglrlm-LongRunEmphasis contributed the most to the detection of KRAS status. The GLRLM refer to quantify gray level runs, which are defined as the length in number of consecutive pixels that have the same gray level value. LongRunEmphasis is a measure of the distribution of long run lengths, with a greater value indicative of longer run lengths and more coarse structural textures within the ROIs, suggesting that the textures of the images with KRAS mutation were more coarse than those without KRAS mutation. Notably, we found that the wavelet features accounted for the largest proportion of the optimal feature set (11/18), indicating that wavelet features have relatively good predictive performance, which is in line with the previous studies (29,34,35). Wavelet transform is a common method for multi-scale texture analysis in image processing, which can quantify the heterogeneity within tumors at different scales and extract more texture information (36). In addition, Small AreaEmphasis selected from GLSZM accounted for the largest proportion (3/6) in the radiomics feature set of the AP model, and also appeared in the NCP model (1/5), suggesting that this feature may have good stability in predicting KRAS mutation, which is a measure of the distribution of small size zones, with a greater value indicative of more fine textures within the ROIs. There were some limitations in this study. First, the present study is a single-center retrospective study, therefore an independent dataset is needed for external validation. Second, the slice thickness of NCP CT images in this study was not completely consistent, however, we have minimized the effect by resampling in the preprocessing process. Finally, 2D segmentation of the tumor was adopted in this study, however, previous studies had shown that the texture analysis results of 2D segmentation and 3D whole-tumor segmentation are similar (37).
In conclusion, our study showed that different phase CT radiomics features could provide different values in predicting KRAS mutations, the combined model, including the dual-phase (AP+VP) model and multiphasic (AP+VP+NCP) model, showed more satisfactory predictive performance compared with the single-phase models, which may suggest that different phase CT images should be considered in radiomics research, rather than single-phase CT image.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding authors.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by the First Affiliated Hospital of Guangzhou Medical University. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.