Value of contrast-enhanced CT based radiomic machine learning algorithm in differentiating gastrointestinal stromal tumors with KIT exon 11 mutation: a two-center study

Background Knowing the genetic phenotype of gastrointestinal stromal tumors (GISTs) is essential for patients who receive therapy with tyrosine kinase inhibitors. Methods We enrolled 106 patients (80 in the training set, 26 in the validation set) with clinicopathologically conrmed GISTs from two centers. Preoperative and postoperative clinical characteristics were selected and analyzed to construct the clinical model. Arterial phase (A-phase), venous phase (V-phase), delayed phase (D-phase), and combined radiomics algorithms were generated from the training set based on contrast-enhanced computed tomography (CE-CT) images. Various radiomics feature selection methods were used, namely least absolute shrinkage and selection operator (LASSO); minimum redundancy maximum relevance (mRMR); and generalized linear model (GLM) as a machine-learning classier. Independent predictive factors were determined to construct preoperative and postoperative radiomics nomograms by multivariate logistic regression analysis. The performances of the clinical model, radiomics algorithm, and radiomics nomogram in distinguishing GISTs with the KIT exon 11 mutation were evaluated by area under the curve (AUC) of the receiver operating characteristic (ROC). The combined radiomics algorithm was found to be the best prediction model for differentiating the expression status of the KIT exon 11 mutation (AUC = 0.836; 95% condence interval (CI), 0.640–0.951) in the validation set. The clinical model, and preoperative and postoperative radiomics nomograms had AUCs of 0.606 (95% CI, 0.397–0.790), 0.715 (95% CI, 0.506–0.873), and 0.679 (95% CI, 0.468–0.847), respectively, with the validation set. GISTs exon 11 CE-CT


Background
Gastrointestinal stromal tumors (GISTs) are the most commonly diagnosed subepithelial tumors of the gastrointestinal tract (1)(2)(3). However, GISTs respond poorly to chemotherapy and prognosis prediction is challenging because of their biological diversity and histological heterogeneity (2). Mutations in the KIT gene, which encodes a tyrosine kinase, are known to lead to the development of GISTs. Besides KIT, mutations in other genes (e.g., platelet-derived growth factor receptor alpha (PDGFRA)) also have been associated with GISTs (4). The KIT mutations have been widely studied and targeted therapy with tyrosine kinase inhibitors (TKIs) has been developed and implemented successfully in the current treatment of GISTs. Different mutations in the different genes associated with GISTs have been shown to affect the response of GISTs to TKIs (2,5).
Most GISTs (75-85%) have KIT mutations, 5-7% have PDGFRA mutations, and 12-15% have no detectable mutation in KIT or PDGFRA (1,4). The National Comprehensive Cancer Network (NCCN) guidelines highly recommend the genetic analysis of GISTs, because the expression status of speci c regions of the KIT and PDGFRA gene sequences are correlated with a speci c response to TKIs (2). The KIT exon 11 mutation, which accounts for about 75% of KIT mutations in GISTs, is the most common imatinib-target mutation and GISTs with this mutation respond well to standard imatinib therapy (1,(5)(6)(7). Because there are relatively few patients who do not have the KIT exon 11 mutation, genetic analysis is required to de ne a wide variety of treatment effects (2,6,7). Therefore, before clinical treatment with TKIs, it is important to determine the expression status of the KIT exon 11 mutation. However, considering tolerance of the procedure and the heterogeneity of GISTs, conventional invasive biopsy or surgery to obtain specimens has many defects (2,8,9). Further, because of the cost, routine genetic analysis of GISTs is still not performed widely (10,11).
Consequently, it is urgent to establish a non-invasive, repeatable, economical, and accurate pretreatment method to identify the expression status of the KIT exon 11 mutation.
Contrast-enhanced computed tomography (CE-CT) is a basic diagnostic method that is used to detect lesions and evaluate the treatment response of GISTs (10)(11)(12). Medical imaging has been playing an increasing key role in personalized precision medicine. Radiomics has been applied to oncology studies (13) and has been widely used in diagnosis (14), prognostic evaluation (15), prediction of biological behaviors (16), and even genetic prediction (11,17).
Previous studies applied CT texture analysis with a small amount of single-center data to distinguish GISTs without the KIT exon 11 mutation (11). To our knowledge, the radiomics approach has not been evaluated for predicting GISTs with the KIT exon 11 mutation. Thus, the aim of this study was to develop and validate a triphase CE-CT based radiomics model with two-center data that can distinguish the patients with the KIT exon 11 mutation.

Selection of Patients and Their Characteristics
This retrospective study was approved by the Institutional Review Boards of the A liated Hospital of Qingdao University (institution 1) and the Shandong Provincial Hospital a liated to Shandong First Medical University (institution 2), which waived patient informed consent. We screened the patients with pathologically con rmed GISTs in the databases of these institutions from July 2008 to August 2019.
Patients who ful lled the following inclusion criteria were included in the study: (a) de nitive genetic analysis results could be obtained; (b) complete clinical characteristics could be obtained; (c) tri-phase CE-CT images before treatment (neoadjuvant therapy or surgery) could be obtained; and (d) the GIST size was ≥ 2 cm and could be outlined in the CE-CT images for volume of interest segmentation.
According to these criteria, 106 patients who also underwent genetic analysis between August 2010 and August 2019 were included: 93 were from institution 1 and 13 were from institution 2. Because the number of patients from institution 2 was small, their data could not be used as an external validation set alone. Therefore, the data of patients from both institutions were merged, then randomly allocated to training and validation sets in a 3:1 ratio. The number of patients with the KIT exon 11 mutation was 45, the number of patients without the KIT exon 11 mutation was 61. Among the 106 patients, 57 were male and 49 were female (mean age 59.5 ± 9.9 years). Detailed information on the clinical characteristics of the selected patients is given in Table 1. A detailed description of the work ow process is shown in Fig. 1. ITK-SNAP (v3.6.0; www.itksnap.org) was used for three-dimensional volume of interest (VOI) segmentation of the CE-CT images. Radiologist 1, who had nine years' clinical experience of using ITK-SNAP, outlined the edge of the tumor on the tri-phase CE-CT images layer by layer and fused them into the VOI. The same radiologist repeated the VOI segmentation and feature extraction two weeks later. In addition, radiologist 2, who had 14 years' clinical experience of using ITK-SNAP, performed the VOI segmentation and feature extraction once. To evaluate inter-and intra-observer reproducibility, the two radiologists randomly selected 30 cases for VOI segmentation and feature extraction. Inter-/ intra-class correlation coe cients (ICCs) were used to evaluate the reproducibility and stability of radiomics feature extraction to identify robust radiomics features. The inter-observer ICC was derived from the rst feature extraction by radiologist 1 and the feature extraction by radiologist 2. The intra-observer ICC was derived from the feature extraction performed twice by radiologist 1.
Generally, ICCs > 0.75 are considered to indicate good reproducibility or reliability. In this study, to ensure highly accurate features were selected, an ICC > 0.9 was used as a criterion for good robustness.
All the CE-CT images were processed using AnalysisKit software (v1.0.3; GE Healthcare, China) for texture feature extraction and quanti cation analysis of tissue heterogeneity. A total of 396 textural parameters were extracted, including the characteristics of 42 histograms, 10 Haralick features, 9 form factors, 48 grey-level co-occurrence matrix (GLCM) features with an offset of 1/4/7, and 60 gray-level run length matrix (GLRLM) features with an offset of 1/4/7.

Combat Compensation Method
The distinguishing features of the texture patterns were retained using the combat compensation method (21), while eliminating the in uence of the scanner and the protocol. This method was helpful for multi-center radiomics analysis (22) and can be used for CE-CT images.

Clinical Characteristics
The preoperative demographic and clinicopathologic data, including age, sex, and blood test results, were To select the clinical factors, Fisher's exact test or the chi-square test was used to compare the categorical variables between the training and validation sets. The Mann-Whitney U test or independent t-test was used to evaluate the discrimination in clinical characteristics among different groups of continuous variables. The relationship between the KIT exon 11 mutation and clinical factors was assessed with univariate analysis.

Results
The 106 patients with GISTs were divided into two groups: those without the KIT exon 11 mutation and those with the KIT exon 11 mutation. The clinicopathologic and demographic data of these patients are summarized in Table 1. The expression status of CA724 and CD117 were signi cantly different between the two groups. No statistically signi cant differences were found between any of the other clinical characteristics.
The univariate logistic regression analysis of the clinical results showed that CA724 was signi cantly associated with GISTs with the KIT 11 mutation in the training set ( Table 2). The multivariate logistic regression analysis of the clinical characteristics showed that CA724 (P = 0.004) and mitotic count (P = 0.020) were independent predictors of the KIT exon 11 mutation in GISTs ( Table 2)  Note: OR, odds ratio;95% CI ,95% con dence interval. Because VOIs of tri-phase CE-CT were analyzed, we developed arterial phase (A-phase), venous phase (Vphase), delayed phase (D-phase), and tri-phase combined radiomics algorithms. This study calculated radiomics scores that could be used to predict GISTs with the KIT exon 11 mutation. Patients without the KIT exon 11 mutation generally had lower radiomics scores than patients with the KIT exon 11 mutation. The selected radiomics features of the A-phase, V-phase, D-phase, and combined radiomics algorithms are shown in Fig. 2. The prediction e ciencies of the radiomics algorithms to distinguish GISTs with the KIT exon 11 mutation in the training and validation sets are shown in Table 3. The AUCs of the radiomics algorithms with the training and validation sets were 0.773 and 0.709, for the A-phase, 0.745 and 0.558 for the V-phase, 0.738 and 0.600 for the D-phase, and 0.826 and 0.836 for the combined radiomics algorithms respectively. These results show that the predictive performance of the combined radiomics algorithm was better than that of any of the single-phase algorithms. The preoperative and postoperative clinical models and combined radiomics algorithm were incorporated into the multivariate logistic regression analysis to generate preoperative and postoperative radiomics nomograms, respectively (Fig. 4A,D). The performances of preoperative and postoperative radiomics nomograms in distinguishing the patients with the KIT exon 11 mutation in the training and validation sets are shown in Table 4. For the preoperative radiomics nomogram with the training and validation sets, the AUCs were 0.913 and 0.715 respectively; for postoperative radiomics nomogram with the training and validation sets, the AUCs were 0.919 and 0.679 respectively. The ROC curves for the training and validation sets were used to compare the prediction e ciencies of the combined radiomics algorithm and preoperative and postoperative radiomics nomograms (Fig. 3A,B). Interestingly, with the training set, the preoperative and postoperative radiomics nomograms both had higher AUCs than the combined radiomics algorithm, whereas, with the validation set, the combined radiomics algorithm had a higher AUC than the three models.
The calibration curves of the combined radiomics algorithm(  Fig. 5A,B).

Discussion
The speci c KIT 11 mutation is associated with the clinical response to targeted therapy of GISTs (1,2,5).
Selecting the targeted therapies based on the GIST genotypes has made precision medicine for GISTs possible. Radiogenomics approaches decode image phenotypes using radiomics to predict the genotypes of tumors, and they have been applied successfully in precision medicine studies (11,17). In this retrospective two-center study, the aim was to develop a personalized radiomics method that could distinguish GISTs with the KIT exon 11 mutation by tri-phase CE-CT. We developed and validated clinical models, A-phase, V-phase, D-phase, and combined radiomics algorithms, and preoperative and postoperative radiomics nomograms. The results showed that the combined radiomics algorithm had a well-calibrated clinical bene t and the best ability to distinguish GISTs with the KIT exon 11 mutation in the validation set, which indicated that radiomics had accurate and reproducible value for predicting GISTs with the KIT exon 11 mutation genotype.
In a previous study, the clinical model was used to predict the biological behaviors of GISTs (25), and genotypes have been shown to be closely related to certain clinical characteristics (26)(27)(28). However, the relationship between some clinical characteristics and the KIT 11 mutation remains controversial (26,27,29,30). Most of the previous studies showed that GISTs with the KIT exon 11 mutation usually had high percentages of mitotic count and/or poor prognosis (26,27,30). The relationship between CA724 and GIST genotypes has not been reported until now. We found that, in the training set, the mitotic count and CA724 were correlated with the GISTs with the KIT exon 11 mutation. However, the preoperative and postoperative clinical models had relatively lower predictive effectiveness (AUCs = 0.509 and 0.606) with the validation set than all the other models tested. This result indicated that the models based solely on clinical characteristics made only limited contributions to the discrimination performance. The roles of mitotic count and CA724 for differentiating the expression status of KIT exon 11 need further studies.
The extraction of high-dimensional features from medical images by radiomics has emerged as a promising approach for predicting the characteristic of tumors (11,13,15,17). Radiomics is more comprehensive than morphological visual analysis and can quantitatively reveal tumor heterogeneity. Various types of medical images can be analyzed to help with diagnosis and treatment and support traditional diagnostic methods (14,16,31,32). Considering that different mutation accompanied by different enhancement ratio, heterogeneous blood supply (33). Using VOIs extracted from tri-phase CE-CT images to construct a combined radiomics algorithm (11) can provide more representative information about a tumor when all available slices are considered (34), and may have better clinical application value and generalizability in discriminating mutation (34)(35)(36). In the present study, we extracted more advanced radiomics features and used relatively large data from two institutions compared with the CT texture analysis and single-center data in a previous study (11). We also used various evaluation tools, excellent radiomics feature selection methods, and machine-learning classi er to obtain the generalization and accuracy of the predictive model.
The radiomics algorithm was proven to be more effective than the clinical model. Next, we incorporated the clinical factors with the combined radiomics algorithm to build preoperative and postoperative radiomics nomograms. The AUC and DCA of the combined radiomics algorithm were compared with those of the preoperative and postoperative radiomics nomograms. The results showed that the combined radiomics algorithm had better predictive effectiveness and clinical applicability with the validation set than the radiomics nomograms, which indicated that the radiomics algorithm re ected molecular-level pathology better than the clinical factors and con rmed the enormous potential of the radiomics algorithm to distinguish GISTs with the KIT 11 mutation.
To prevent over tting, the mRMR and LASSO regression methods were applied to the extracted features. For feature selection, we used mRMR, which is a new feature screening method (31), to screen radiomics features that had the fewest redundancies and most credible coe cients. LASSO is a feature screening method for selecting the predominant features of both ridge regression and subset screening (37). We used GLM as a machine-learning classi er because it was reported to produce the best screening performance among the 11 machine-learning methods that were tested (32). With application of the excellent feature selection methods and machine-learning classi er, the combined radiomics algorithm satisfactorily distinguished GISTs with and without the KIT exon 11 mutation.
The development of targeted therapy and genetic analysis have made precision medicine for GISTs feasible (5,11,27,28). The most common imatinib-target mutation is the KIT exon 11 mutation, and patients with this mutation can bene t from standard imatinib therapy (2,4,5,10). The relatively rare patients without the KIT exon 11 mutation require further genetic analysis to identify multiple effects of different treatments (1,2,(4)(5)(6). The combined radiomics algorithm developed in this study will provide an alternative to biopsies, which can produce inconsistent results due to tumor heterogeneity, and surgery, which some special patients cannot tolerate (2,8,10), to distinguish GISTs with the KIT exon 11 mutation non-invasively. For patients who were evaluated as the KIT exon 11 mutation with the combined radiomics algorithm, clinicians could recommend the standard imatinib therapy. For patients who were evaluated not to have the KIT exon 11 mutation, using the combined radiomics algorithm would not delay precision treatment or cause any unnecessary nancial burden.
Our study has several limitations. First, all data were collected retrospectively and limited to Chinese patients, so selection bias was inevitable. Second, although the data were from two centers, the proportion of samples that had been performed genetic analysis was small, possibly because of cost considerations, which highlights the need for and signi cance of radiomics analysis. These ndings need to be validated in international, larger multi-institutional studies. Third, only CE-CT images (slice thickness of 5.0 mm) were analyzed in this study. Whether radiomics could achieve a better performance on different slice thickness of CE-CT is worth further investigation. Fourth, the hand-crafted feature requires time-consuming tumor boundary segmentation, deep learning radiomics could accurately and automatically detect and segment and achieve excellent performances (17). Lastly, this study focused only on GISTs with the KIT exon 11 mutation. GISTs without the KIT exon 11 mutation or secondary mutation could be explored in future work (4,38).

Conclusion
We developed and validated several predictive models that could non-invasively distinguish GISTs with KIT exon 11 mutation based on tri-phase CE-CT images. The combined radiomics algorithm performed better than the other predictive models, which con rmed it as a potential predictor to supplement conventional approaches for selective genetic analysis and to support the clinical decision about imatinib therapy in the precision medicine of GISTs.

Consent for publication
All authors have read and approved the content and agree to submit for consideration for publication in the journal.

Availability of data and materials
All data generated or analysed during this study are included in this published article.

Competing interests
The authors declare that they have no competing interests.

Funding
This work was supported by National Natural Science Foundation of China (81770631).

Author Contributions
All authors have had access to the data and all drafts of the manuscript. Specifc contributions are as follows: design and supervision: JZ, XJ; development of methodology: JZ,XJ, BL, HW; collection and extraction of data: BL, SY, YS, HL, ZZ, JZ, FH; analysis and interpretation of data (e.g., statistical analysis, algorithm and software development):SL; manuscript writing: LB, HW; nal approval of manuscript: all authors.

Figure 1
Work ow process for radiomics processing and analysis in this study.

Figure 2
The selected radiomics features and their coe cients of the A-phase (a), V-phase (b), D-phase, (c) and combined (d) radiomics algorithms. x-axis, coe cients of features with the LASSO logistic regression analysis; y-axis, features selected using the radiomics algorithms.