Open Access 
| Korean J Radiol. 2020 Jun;21(6):670-683. English. Published online Apr 06, 2020. https://doi.org/10.3348/kjr.2019.0607 | |
| Copyright © 2020 The Korean Society of Radiology | |
Kai Xu , MD,1
Lin Liu, MD, PhD,1
Wenhui Li, PhD,2
Xiaoqing Sun, MD,1
Tongxu Shen, MD,1
Feng Pan, MD,1
Yuqing Jiang, MD,1
Yan Guo, MD,3
Lei Ding, MD,1
and Mengchao Zhang, MD, PhD 1
| |
|
1Department of Radiology, China-Japan Union Hospital of Jilin University, Changchun, China. | |
|
2College of Computer Science and Technology, Jilin University, Changchun, China. | |
|
3Life Sciences, GE Healthcare, China, Shenyang, China. | |
| Corresponding author: Mengchao Zhang, MD, PhD, Department of Radiology, China-Japan Union Hospital of Jilin University, NO. 126 Xiantai Street, Changchun 130033, China. Tel: (86) 15948705198, Fax: (86) 431-849957,
| |
| Received August 31, 2019; Revised December 09, 2019; Accepted January 27, 2020. | |
|
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by- | |
 This article has been cited by 6 articles in  This article has been cited by Google Scholar. This article has been cited by 4 articles in PubMed Central. This article has been cited by 7 articles in Scopus. This article has been cited by 5 articles in Web of Science. | |
|
Abstract
| |
|
Objective
The presence of coagulative necrosis (CN) in clear cell renal cell carcinoma (ccRCC) indicates a poor prognosis, while the absence of CN indicates a good prognosis. The purpose of this study was to build and validate a radiomics signature based on preoperative CT imaging data to estimate CN status in ccRCC.
Materials and Methods
Altogether, 105 patients with pathologically confirmed ccRCC were retrospectively enrolled in this study and then divided into training (n = 72) and validation (n = 33) sets. Thereafter, 385 radiomics features were extracted from the three-dimensional volumes of interest of each tumor, and 10 traditional features were assessed by two experienced radiologists using triple-phase CT-enhanced images. A multivariate logistic regression algorithm was used to build the radiomics score and traditional predictors in the training set, and their performance was assessed and then tested in the validation set. The radiomics signature to distinguish CN status was then developed by incorporating the radiomics score and the selected traditional predictors. The receiver operating characteristic (ROC) curve was plotted to evaluate the predictive performance.
Results
The area under the ROC curve (AUC) of the radiomics score, which consisted of 7 radiomics features, was 0.855 in the training set and 0.885 in the validation set. The AUC of the traditional predictor, which consisted of 2 traditional features, was 0.843 in the training set and 0.858 in the validation set. The radiomics signature showed the best performance with an AUC of 0.942 in the training set, which was then confirmed with an AUC of 0.969 in the validation set.
Conclusion
The CT-based radiomics signature that incorporated radiomics and traditional features has the potential to be used as a non-invasive tool for preoperative prediction of CN in ccRCC. |
|
Keywords:
Coagulative necrosis; Clear cell renal cell carcinoma; CT; Radiomics
|
|
|
INTRODUCTION
|
Renal cell carcinoma (RCC) is the most common primary malignancy of the kidney, accounting for approximately 85–90% of renal malignancies. Clear cell renal cell carcinoma (ccRCC) is the most common subtype of RCC, accounting for approximately 70% of the cases (1). With the increased health awareness among people and the development of advanced examination methods, the proportion of RCC cases that are being incidentally detected is gradually increasing (2, 3, 4, 5). Among incidental RCCs, ccRCC is a common pathological type and has a higher risk of a poor prognosis (6, 7); therefore, it has received more attention in clinical practice. Radiofrequency treatment or cryoablation have recently become options for the treatment of renal cancer and are suitable for tumors with good prognosis (8, 9). Therefore, there is an urgent need for preoperative assessment of the prognosis of ccRCC. Histological coagulative necrosis (CN) has been widely recognized as an important independent prognostic factor for ccRCC (7, 10, 11, 12, 13). Studies have shown that the 10-year cancer-specific survival in patients showing CN in ccRCC tumors is 29.2%, while the corresponding value in those without necrosis is as high as 77.6% (11) . Moreover, for ccRCC, the risk ratio of death in patients with CN and non-necrosis in the tumor is 5.27 (11). Therefore, prediction of the presence or absence of CN within the tumor before surgery is a very important factor influencing the choice of treatment strategy for RCCs. Although needle biopsy is an effective method for obtaining pathological findings before surgery, because of tumor heterogeneity, needle biopsy does not yield convincing results related to CN (14). Moreover, as an invasive examination, this technique can cause a variety of complications (15). Therefore, it is worth exploring whether noninvasive methods can be used to predict CN in ccRCC accurately.
Previous studies have shown that traditional image features, such as enhancement characteristics, could provide valuable predictive information for identifying benign and malignant tumors, tumor subtypes, and tumor grades in RCC (16, 17). Nowadays, with the development of radiomics technology, radiomics methods that translate medical imaging data into high-dimension data can also be used as non-invasive biomarkers for prognosis or prediction (18, 19, 20, 21). However, it remains unclear whether it is possible to predict the presence or absence of CN in ccRCC tumors by using radiomics features and traditional features based on CT images. Moreover, the types of features that could yield a higher prediction accuracy are unknown.
Therefore, the purpose of this study was to build and validate a radiomics score and a traditional predictor based on CT imaging for prediction of the CN status in ccRCC. Moreover, we developed an inclusive radiomics signature incorporating the radiomics score and traditional predictors for preoperative estimation of the CN status in ccRCC patients.
|
MATERIALS AND METHODS
|
Patients
The retrospective study was approved by the Ethics Review Committee of our hospital. The requirement for informed consent was waived because CT image acquisition is part of a routine non-invasive examination protocol for suspected RCC patients.
Between March 2013 and March 2019, 105 patients with ccRCC underwent surgical resection in our hospital; the obtained pathological results were collected in this study. Patient inclusion/exclusion criteria are presented in Figure 1. Furthermore, a renal mass was obtained from each patient. After strict screening of the enrolled patients, 105 patients were included in the study.
|
Sample Size Consideration
Based on sample size calculation methods in clinical research (22) introduced by Shein-Chung Chow, Ph.D., we estimated the validation sample size that was sufficiently independent to test whether our model was robust and efficient.
First, considering the two groups to be A and B, µ represents the mean value of the average radiomics score for each group, with the hypotheses of interest being as follows:
The sample size and power are calculated respectively as follows:
Where n is the sample size in the training group and N is the sample size for the validation group, Φ is the standard normal distribution function, α is the type I error, β is the type II error, 1 − β is the power, and σ2 is the variance of the covariate.
In our study, the sample sizes in the training groups were nA = 28 and nB = 44 with means of µA = −2.086 and µB = 0.643, respectively, and with a variance of σ2 = 3.3508. Therefore, the minimum number of validation samples was 15 (without CN) and 10 (with CN) in the two groups with the desired two-sided significance level of α = 0.05 and power of 1 − β = 95%. In our study, the validation set included 20 cases without CN and 13 cases with CN in the two groups, respectively, which were greater than the minimum required sample sizes.
CT Examination
Triple-phase CT-enhanced images were obtained using a 64-slice CT scanner (Discovery CT750 HD, GE Healthcare, Boston, MA, USA). The scanning parameters were as follows: tube voltage, 120 kV; tube current automatic adjustment technology; scanning range, 500.00 mm; scanning thickness, 1.25 mm; rotation speed, 0.6 s/circle; and matrix size, 512 × 512. The patients were injected with 100 mL of a contrast medium (iohexol; Omnipaque, 300 mg iodine/mL, GE Healthcare) via an elbow vein using a high-pressure syringe (Missouri XD2001, Ulrich Medical, Ulm, Germany) at the rate of 4.5 mL/s. A corticomedullary phase scan was performed 25–30 seconds after the injection of the contrast medium; the nephrographic scan was performed at 60–70 seconds; and the excretory phase was scanned at 2–3 minutes. All patients were scanned while they were holding their breath after deep inhalation.
Figure 2 shows the radiomics workflow, which included the feature extraction after the CT imaging, which was followed by analysis and signature building.
|
Feature Extraction
Traditional Feature Extraction
The traditional features included both qualitative and quantitative features. The qualitative features were as follows: side, defined by the location of the tumor in the left or right kidney; location, defined by the location of the tumor in the upper pole, lower pole, or the interpolar region of the kidney; arteryIntratumoral, classified as “presence” when tortuous arteries were observed within the renal mass in the corticomedullary phase, and conversely classified as “absence,” as shown in Figurea 3A; peritumoral neovascularity, defined by the presence of a blood-supplying artery around the tumor in the corticomedullary phase, as shown in Figure 3B; calcification, defined by the presence of calcification in the tumor; completeness of the pseudocapsule, defined by the presence of a complete, high- or low-attenuation rim surrounding the renal neoplasm in the coronal or sagittal planes for the nephrographic or excretory phase (23), as shown in Figure 3C and D. The quantitative features included diametermax, the tumor attenuation value (TAV), the renal cortex attenuation value (CAV), and the difference ratio. Diametermax was defined as the longest diameter of the largest layer of the tumor in the transverse planes; TAV and CAV were defined as the attenuation of areas with the most obvious enhancement in the tumor and the renal cortex on the same plane, respectively. The difference ratio was defined as the ratio of the difference between TAV and CAV to the renal CAV. These quantitative feature data were obtained in the corticomedullary phase. TAV and CAV were measured by drawing the region of interest (ROI) with a size of approximately 25 mm2 in the most obvious enhancement area of the tumor and the same level of the renal cortex, respectively (Fig. 3E). All the measurements and evaluations were conducted by two independent radiologists with 8 (reader 1) or 6 years (reader 2) of experience in abdominal CT interpretation, who were blinded to the pathological results. Among the two sets of findings, the image analysis results obtained by the radiologist with 8 years of diagnostic experience was used for data analysis in this study. The above image analysis processes were performed on the Picture Archiving and Communication System viewer.
|
Radiomics Feature Extraction
First, data preprocessing was performed to address the differences in image quality and image noise between images and to ensure that image features were calculated using the same specifications. All the images were resampled into voxel sizes of 1 × 1 × 1 mm3 using linear interpolation. In addition, a Gaussian filter was used for denoising. Then, tumor segmentation and feature extraction were performed.
The CT images were stored in the Digital Imaging and Communications in Medicine format and uploaded to the ITK-SNAP software (http://www.itk-
|
Pathological Assessment
Whole-tumor specimens were placed in formalin solution and sent to a pathology laboratory. After staining with hematoxylin and eosin, a histopathological evaluation of the specimens was performed by a pathologist with more than 10 years of experience, who observed the microscopic CN inside the tumor under a microscope. Pathological images of tumors with and without CN are shown in Figure 5.
|
Consistency Test
Inter-observer agreement was determined to assess traditional features. We used the intraclass correlation coefficient (ICC). An ICC value > 0.75 was considered indicative of good agreement.
For radiomics features, we randomly selected image data from 20 patients, of which 10 had CN in the tumor. The VOI was delineated by another abdominal radiologist with 6 years of CT interpretation experience (reader 2), and the data were acquired. Finally, the same methods and standards were used to assess the consistency of traditional features.
Statistical Analysis
The categorical variables were compared using the chi-squared test, and the continuous variables were compared using the Mann–Whitney U test. Binary logistic regression analysis was used to analyze the correlation between traditional features and CN.
The least absolute shrinkage and selection operator (LASSO) algorithm was used to identify the best radiomics features that were significantly associated with CN in the ccRCC. A multivariate logical regression model combining the candidate variables selected by the LASSO algorithm was built to ensure efficiency.
An receiver operating characteristic (ROC) curve analysis was used to illustrate the prediction performance of the selected features. The optimal cutoff value was selected as the point when the sensitivity plus specificity was maximal, and the area under the ROC curve (AUC) value was calculated. The DeLong test was used as a difference test on the AUC of different results.
The Mann–Whitney U test, chi-squared test, ICC calculation, and the kappa test were performed using SPSS Statistics (version 22.0, IBM Corp., Armonk, NY, USA). The confidence level was maintained at 95%, and a p value of less than 0.05 was considered significant. The DeLong test, LASSO algorithm, multivariate logical regression model construction, and ROC analyses were performed using R Studio (Version 1.0.143© 2009–2016, R Studio, Inc.: https://www.r-
|
RESULTS
|
The patients were divided into training and validation sets based on the principle of random allocation. The training and validation sets were also analyzed for data pertaining to patient characteristics and traditional and radiomics features.
Patient Characteristics
Using random allocation, 72 patients (48 men, 24 women; mean age, 57.0 ± 8.3 years) were assigned to the training group, while 33 patients (27 men, 6 women; mean age, 54.3 ± 9.7 years) were allocated to the validation group. The training and validation sets contained 28 and 13 patients with CN in their tumors, respectively. There was no significant intergroup difference in the age and sex of the patients with and without CN in the training and validation sets. However, significant differences were observed in the International Society of Urological Pathology (ISUP) grade, pathology of tumor (pT) stage, and the existence of intratumoral CN in the training and validation sets. CN was more likely to occur in tumors with higher pT stage and ISUP grade in this study. An analysis of the patient characteristics, pathological features, and CN in the training and validation sets is shown in Table 1.
|
Performance of the Traditional Predictors
Among traditional features, a significant difference between ccRCCs with and without CN was found for two features—diametermax and completeness of the pseudocapsule—with all p values < 0.05 in the training and validation sets. The AUC values were 0.713 (95% confidence interval [CI], 0.594–0.833) and 0.760 (95% CI, 0.639–0.880) for diametermax and 0.738 (95% CI, 0.563–0.914) and 0.771 (95% CI, 0.595–0.947) for completeness of the pseudocapsule in the training and validation sets, respectively. No significant difference was observed in the two features among tumors regardless of the presence or absence of CN (all, p > 0.05). In the validation set, besides these two features, significant differences were found for TAV and CAV. Among these four features, the ROC curve for the complete pseudocapsule had the highest AUC value in the two sets. A comparison of the ROC analysis of the imaging features is shown in Tables 2 and 3. We used logistic regression to analyze the correlation between traditional features and CN in the training and validation sets. We found that diametermax and completeness of the pseudocapsule were the only two selected features in the training and validation sets. There were significant differences between the two features and the existence of intratumoral CN. When there was no clear complete pseudocapsule in the ccRCC, the incidence of intratumoral CN was about 11 and 13 times that observed in cases with a clear pseudoenveloped tumor in the training and validation sets, respectively. Moreover, the incidence of CN in the tumor increased by 40% and 80%, respectively, in the two sets when the diametermax of the tumor increased by 1 cm.
|
|
A higher AUC value was obtained when the two traditional features (diametermax and completeness of pseudocapsule) were combined by a computational model (Model 1) constructed using multivariate logical regression analysis. The calculation formula for Model 1 is as follows:
The AUC value was 0.843 (95% CI, 0.750–0.935; sensitivity, 0.821 and specificity, 0.773) in the training set, and 0.858 (95% CI, 0.718–0.998; sensitivity, 0.692 and specificity, 0.750) in the validation set. It was higher than the AUC of diametermax and the completeness of the pseudocapsule (all, p < 0.05). Moreover, there was no significant difference in other features observed among tumors regardless of the presence or absence of CN (all, p > 0.05).
The consistency test results for the traditional features were good (all, > 0.750), and were in the range of 0.775–1.000 The specific results for each traditional feature are shown in Table 2.
Performance of the Radiomics Score
Seven optimal features, called ClusterProminence_angle90_ offset7, ClusterShade_angle0_offset7, Compactness2, HaralickCorrelation_angle135_offset7, Inertia_AllDirection_offset1_SD, LongRunLowGreyLevelEmphasis_angle0_offset7, and ShortRunEmphasis_ angle45_offset7, were screened using the LASSO algorithm. The introduction and equations for the seven optimal features are attached in Supplementary Materials.
The multivariate logical regression model (Model 2) built with the seven radiomics features was called the “radiomics score.” The calculation formula for this score is as follows:
The ROC curves constructed using the model had a high AUC value. The AUCs in the training and validation sets were 0.855 (95% CI, 0.770–0.940; sensitivity, 0.893; specificity, 0.750) and 0.885 (95% CI, 0.766–1.000; sensitivity, 0.923; specificity, 0.800), respectively.
The consistency test results for these seven radiomics features were good (all, > 0.750). Among these, the ICC result for Inertia_AllDirection_offset1_SD was the smallest, which was 0.91; the result for Compactness 2 was the largest, which was 0.99. The ICC results of the other five features were in the range of 0.91–0.99.
Development of the Radiomics Signature Incorporating Traditional Predictors and Radiomics Score and Performance Assessment
The weighted coefficients of the selected radiomics features in the multivariate logical regression model were presented in a linear formula for radiomics score calculation. A high AUC was obtained with a computational model (Model 3) constructed using the traditional features and the radiomics score. The calculation formula for Model 3 is as follows:
The AUC was 0.942 and 0.969 in the training and validation sets, respectively. We constructed a nomogram with the diametermax, completeness of the pseudocapsule, and radiomics score as predictors to display the prediction performance of model 3 intuitively. The nomogram is shown in Figure 6.
|
A BarChart diagram was used to visualize the classification accuracy of model 3 in the training and validation set. The pink and green bars represent tumors with and without CN, respectively. Therefore, the green part below the threshold and the pink part above the threshold are misclassified data. The BarChart diagram of the training group and the validation group is shown in Figure 7.
|
A comparison of the ROC curves constructed by the two traditional features, Model 1, Model 2, and Model 3 in the training and validation sets is shown in Figure 8.
|
|
DISCUSSION
|
The results of this study showed that CT-based imaging features, irrespective of whether they were traditional or radiomics features, could accurately predict the presence or absence of CN in ccRCC. The results were confirmed further in the analysis of the validation group.
The results of this study indicated that the two artificial recognition features of diametermax and completeness of the pseudocapsule could accurately distinguish between the presence and the absence of CN within the tumor in both training and validation sets. When the two features were combined, a higher AUC could be obtained. The diametermax of the tumor positively correlated with the occurrence of CN, which might be attributed to the greater likelihood of CN within large-diameter tumors (24). In this study, the incidence of coagulation necrosis in ccRCCs increased by 40% and 80% in the training and validation sets, respectively, when the diametermax of the tumor increased by 1 cm. This result is similar to that obtained in previous studies, in which CN was shown to be often present in ccRCCs with size > 10 cm (11). The ROC curves constructed using the feature of completeness of the pseudocapsule had the highest AUC values in the training and validation sets, and its accuracy was higher than that of the other traditional features. This suggested that the completeness of the pseudocapsule was superior in predicting CN in ccRCC. Pseudocapsule formation is a result of tumor growth, which causes compression, ischemia, and necrosis of the adjacent renal parenchyma, and results in the deposition of fibrous tissue (23, 25). Previous pathological studies have shown a higher proportion of CN in ccRCC cases with an incomplete pseudocapsule (25). This conclusion supports the results of our study.
Among the radiomics features, seven quantitative features were selected by the LASSO algorithm to distinguish ccRCC with CN from tumors without CN. The results showed that the multivariate logical regression model constructed using radiomics features was effective in both the training set (AUC, 0.855) and the independent validation set (AUC, 0.885). Radiomics was thus proven to show a high prediction value. Moreover, most of the selected features were texture features, which reflected the heterogeneity of the tumor ROI (21). For example, cluster prominence is a measure of asymmetry of a given distribution, and high values of this feature indicate that the symmetry of the image is low. In this study, the values of cluster prominence extracted from tumors containing CN were higher than the values extracted for tumors without CN. Furthermore, the radiomics features considered in this experiment were extracted based on the whole-tumor delineation on contrast-enhanced CT images. A whole-tumor ROI delineation can reflect more accurately and comprehensively the characteristics and the heterogeneity of the tumor (26, 27). In addition, on contrast-enhanced images, texture features will also reflect the distribution of the contrast agent between the intra- and extravascular extracellular spaces. One hypothesis is that CN results from the tumor's growth beyond the supply of the existing vasculature (28). Therefore, contrast-enhanced images can more comprehensively reflect the existence of tumor CN.
Moreover, as for traditional features, a higher AUC could be obtained when completeness of the pseudocapsule and diametermax were combined in this study. This indicated that in clinical practice, we could use these traditional features to obtain a preliminary prediction of the presence or absence of CN in a tumor and that doing so would in fact be more convenient and practical than the use of radiomics features. However, evaluation of the traditional features requires experience of image diagnosis, and the results will be influenced by subjective factors. The results for the radiomics features and combined features were slightly higher than those for the traditional features were. Moreover, radiomics features are less affected by subjective factors. However, the entire tumor had to be delineated and specific software was used in the process of analysis; therefore, the process is relatively time- and effort-intensive. This indicated that in clinical practice, if conditions permit, radiomics features could be used to predict CN in ccRCC cases to obtain more accurate results or to evaluate tumors more comprehensively.
In this study, we also analyzed demographic and clinical features. The results indicated that there was no significant difference in demographic characteristics (age and sex) between the ccRCC groups with and without CN, which was consistent with the previous findings (11). Statistically significant differences were found in ISUP grading and pT staging between the ccRCC cases with and without CN. CN in ccRCC was significantly associated with adverse pathologic features, including ISUP grade and pT stage. Moreover, CN was more likely to occur in relatively high-grade and high-stage tumors. This is similar to the results of certain previous studies (11, 29).
The current study had several inherent limitations. First, we only constructed the models using VOIs sketched on the corticomedullary phase CT images and did not extract other phases for the multi-parameter analysis. However, we constructed the ROC curve using the features extracted on the nephrographic phase images in the pre-experimental assessments. The results showed that the AUC value in the nephrographic phase was lower than that in the corticomedullary phase. Second, although our sample size met the standard for a diagnostic experiment after estimation of the sample size, a prospective and multi-center experimental study is still needed for experimental verification of these models in the future.
Overall, according to our current research, the accuracy of the multivariate logical regression model constructed by combining traditional features and radiomics features in predicting CN in the training and validation sets could reach 0.942 and 0.969, respectively. Thus, imaging methods could be used to assess the prognostic risk of ccRCC to determine which strategy could be used. In addition, the noninvasive nature of the method allowed for repeated evaluations during follow-up and compensated for the limitations of needle biopsies in obtaining accurate findings for CN.
In conclusion, CN in ccRCC could be detected by using traditional features or radiomics features selected based on CT imaging.
|
Supplementary Materials
|
The Data Supplement is available with this article at https://doi.org/10.3348/kjr.2019.0607.
|
Notes
|
This study was supported by the science and Technology Development Plant of Jilin Province (No.20180101015JC), the research grant from the Jilin Province Science and Technology Development Plan Project (NO.20190303182SF).
Conflicts of Interest:The authors have no potential conflicts of interest to disclose.
|
References
|
Related citations