A novel radiomics–platelet nomogram for the prediction of gastroesophageal varices needing treatment in cirrhotic patients

Highly accurate noninvasive methods for predicting gastroesophageal varices needing treatment (VNT) are desired. Radiomics is a newly emerging technology of image analysis. This study aims to develop and validate a novel noninvasive method based on radiomics for predicting VNT in cirrhosis. In this retrospective–prospective study, a total of 245 cirrhotic patients were divided as the training set, internal validation set and external validation set. Radiomics features were extracted from portal-phase computed tomography (CT) images of each patient. A radiomics signature (Rad score) was constructed with the least absolute shrinkage and selection operator algorithm and tenfold cross-validation in the training set. Combined with independent risk factors, a radiomics nomogram was built with a multivariate logistic regression model. The Rad score, consisting of 14 features from the gastroesophageal region and 5 from the splenic hilum region, was effective for VNT classification. The diagnostic performance was further improved by combining the Rad score with platelet counts, achieving an AUC of 0.987 (95% CI 0.969–1.00), 0.973 (95% CI 0.939–1.00) and 0.947 (95% CI 0.876–1.00) in the training set, internal validation set and external validation set, respectively. In efficacy and safety assessment, the radiomics nomogram could spare more than 40% of endoscopic examinations with a low risk of missing VNT (< 5%), and no more than 8.3% of unnecessary endoscopic examinations still be performed. In this study, we developed and validated a novel, diagnostic radiomics-based nomogram which is a reliable and noninvasive method to predict VNT in cirrhotic patients. NCT04210297.


Introduction
Gastroesophageal varices (GEV) are the principal complications of cirrhotic portal hypertension. Studies have demonstrated that GEV develops in approximately 50% of cirrhotic patients and ruptured GEV occurs in approximately 10-15% per year [1,2]. The mortality rate in cirrhotic patients with a first hemorrhage from GEV is 20%, and patients who survive the first hemorrhage without intervening are at high risk of rebleeding (greater than 60% at 1 year), with a mortality rate of approximately 33% [3].
Given to the mortality and morbidity associated with GEV, the guidelines recommend that all cirrhotic patients should be screened for GEV [4,5]. Upper endoscopy is recommended as the golden standard for GEV. However, endoscopic examination is invasive. In addition, a large proportion of cirrhotic patients do not present with varices needing treatment (VNT), so the majority of cirrhotic patients are exposed to the risk of invasive procedure and sedation complications without detecting VNT. It would cause much useless endoscopic examination leading to the medical burden.
The 2015 Baveno VI consensus workshop proposed a noninvasive method that the patient with liver stiffness < 20 kPa and a platelet count > 150,000/mm 3 could avoid endoscopic screening safely [6]. However, the efficacy of Baveno VI criteria was criticized owing to a substantially low number of spared endoscopies (15-30%) [7][8][9]. Thus, researches on developing a more accurate noninvasive method are encouraging.
Computed tomography (CT) is widely applied in liver cirrhosis, vastly contributing to the diagnosis and evaluation of the complications of cirrhosis. However, the assessment is highly dependent on the experience and subjectivity of radiologists. Previous studies have attempted to quantitatively analyze CT findings by measuring varices size, liver, and spleen volume, but these radiological parameters did not show a satisfactory performance [10][11][12][13][14].
Radiomics is a newly emerging technology of image analysis which refers to extracting high-throughput and quantitative features from medical images, revealing the correlation between these features and the disease using data mining algorithms and statistics analysis, then builds an appropriate model with refining features [15,16]. Previous studies suggest the potential application of radiomics in predicting VNT [17][18][19]. However, the previous radiomics models do not contain the esophageal and gastric radiomics features which are important evidence for the radiologist to determine the existence and severity of GEV. What's more, the efficacy and safety of the radiomics model for predicting VNT are unclear.
Therefore, in this study, we aimed to develop a novel radiomics model containing the esophageal and gastric radiomics features for predicting VNT in multiple etiological cirrhotic patients and to assess its performance in clinical application, particularly to assess its efficacy and safety.

Study design and patients
This retrospective-prospective study was performed in Qilu Hospital of Shandong University (Institution 1) and Jinan Central Hospital (Institution 2). The study design and procedures were presented in full in the study protocol (Clini-calTrials.gov: NCT04210297). Ethical committee approval was granted by the Medical Ethics Committee of involved institutions. The informed consent was waived in the training set for the retrospective analysis. All procedures involving human participants were performed following the Helsinki declaration and its later amendments.
The training set consists of patients who were retrospectively collected in institution 1 from January 2018 to December 2019. The internal and external validation set consists of patients who were prospectively enrolled from January 2020 in institution 1 and institution 2, respectively.

Upper endoscopic examination
Every patient received an upper endoscopic examination for the screening of EV and identifying the risk of bleeding. Upper endoscopic examination was performed by experienced endoscopists. The endoscopic findings were checked by two experienced endoscopists and recorded in a standard format. VNT was defined as small

Radiomics analysis
The workflow of the radiomics analysis is summarized in Fig. 1 and can be divided into four steps: CT image acquisition, region of interest (ROI) segmentation, feature extraction, and radiomics signature construction.

CT image acquisition
Every patient underwent an abdominal enhanced CT scan after an overnight fast using one of the following systems: Discovery CT750 HD (GE Healthcare), Brilliance iCT (Philips Healthcare), or Sensation 16 CT (Siemens). The following parameters were used: tube voltage, 120 or 140 kVp; tube current, 150 − 600 mAs; slice thickness, 1.25 mm; pitch, 1.375. Ultravist (2.5 mL/kg, 300 mg/mL) was injected intravenously at a rate of 3 mL/s. Arterial phase scan began at the 30 s after injection, while the venous phase and delayed phase scan were started at 70 and120 s, respectively. Portal venous phase CT images were retrieved from the picture archiving and communication system (PACS).

Region of interest (ROI) segmentation and feature extraction
The liver at the porta hepatis level, the spleen at splenic hilum level, and the level from the lower esophagus to gastric fundus were selected as the ROI. ROI was delineated manually by two radiologists (reader 1: Dexin Yu and reader 2: Zhuyun Liu with 20 and 3 years of clinical experience in abdominal CT interpretation, respectively) using the ITK-SNAP 3.8 (www. itksn ap. org). The two radiologists were blinded to the endoscopic findings.
Radiomic features were extracted from each ROI using the MATLAB 2018b (MathWorks, Natick, USA) by utilizing the open-source radiomics feature extraction package. Textural and nontextural feature extractions were conducted (Table S1). We extracted 4 nontextural features (volume, shape, size, and solidity) and 43 textural features. The textural features consist of two parts: (1) first-order statistics texture features (histogram feature) included three features (Variance, Skewness, Kurtosis) which are extracted from the grayscale histogram of the ROIs. (2) high-order texture features were derived from four statistical feature matrices: the gray level cooccurrence matrix (GLCM), the gray level run length matrix (GLRLM), the gray level size zone matrix (GLSZM), and the neighborhood gray tone difference matrix (NGTDM). First-order statistics features describe the gray level distribution of all voxels within the ROI. Highorder texture features reflect information about the spatial arrangement of voxel intensities and therefore could describe the homogeneity of ROI. Image normalization including Wavelet bandpass filtration, isotropic resampling, and quantization of gray level was performed before radiomic features extraction. For each ROI, 10,324 radiomic features were extracted and a total of 30,972 radiomic features were extracted from each patient. More detailed information about radiomics features extraction methodology is shown in supplementary materials.
The interobserver and intraobserver reliability were analyzed with 30 randomly chosen cases from the training set, two radiologists repeated ROI segmentation and feature extraction twice on those cases with one-month interval. The reliability was calculated by using the intraclass correlation coefficient (ICC), both intraobserver and interobserver ICC values greater than 0.75 were regarded as robust reliability and reproducibility.

Feature selection and radiomics signature construction
After feature extraction, a large scale of redundant features was obtained which resulted in overfitting and reducing the discrimination ability of the model. The least absolute shrinkage and selection operator (LASSO) logistic regression method was used to select the most effective predictive features from the training set. LASSO is a regression method for high-dimensional data analysis, which typically performs variable selection and regularization using L1 penalty to shrink regression coefficients of the redundant features to zero. The penalty parameter lambda (λ) was tuned using tenfold cross-validation based on the minimum partial likelihood deviance. The features with nonzero coefficients in optimal penalty parameter lambda (λ) were chosen for construction of radiomics signature. A formula was generated using a linear combination of selected features that were weighted by their respective LASSO coefficients; the formula was then used to calculate a radiomics signature (Rad score) for each patient to predict the risk of VNT.

Radiomics nomogram construction
The Rad scores and the clinical variables were tested in univariate logistic regression analysis in the training set. All variables with p < 0.05 were entered into the multivariate logistic regression analysis. A radiomics nomogram was then constructed according to the multivariate logistic regression model.

Efficacy and safety assessment
We assess the efficacy and safety of radiomics nomogram according to the number of spared endoscopies, unneeded endoscopies and the missed VNT. Spared endoscopies = the number of patients who were classified as a non-VNT group by noninvasive methods. The rate of unneeded endoscopy = the number of non-VNT patients / the number of patients who were classified as VNT group by noninvasive methods. The rate of VNT missed = the number of VNT patients / total VNT patients.

Statistical analysis
Differences of clinical characteristics between the training set and the validation set as well as between the VNT group and non-VNT group in their respective datasets were assessed using independent sample t test or Mann-Whitney U test. The optimal cutoff value for Rad score was determined using Youden's index in the training set, which maximizes the sum of sensitivity and specificity. The predictive accuracy of the radiomics signature was quantified by the area under the receiver-operator characteristic (ROC) curve (AUC) in both training and validation sets. The likelihood ratio test with a backward stepwise selection was applied to the multivariate logistic regression model. The pointbiserial correlation test was used to correlate the radiomics nomogram and the CT image findings. In addition, a decision curve analysis was performed to evaluate the clinical usefulness and net benefits of the developed radiomics nomogram. Statistical analysis was performed using the R software (version 3.6.2, R Project for Statistical Computing, http:// www.r-proje ct. org). Two-sided p values less than 0.05 were considered statistically significant.

Study population
A total of 245 cirrhotic patients from two institutions were included in this study. 111 patients retrospectively collected from January 2018 to December 2019 at institution 1 were taken as the training set, 71 patients prospectively enrolled from January 2020 to October 2020 at institution 1 were used as the internal validation sets and 63 patients prospectively enrolled from January 2020 to October 2020 at institution 2 were used as the external validation set. The results of blood tests within 1 week of endoscopic examination were recorded. Baseline characteristics of the study population were outlined in Table 1.

Radiomic features selection and radiomics signature (Rad score) development
After extracting features from ROIs, 30,972 radiomics features were retrieved from patients in the training set and reduced to 19 potential predictors with 14 features from the gastroesophageal region and 5 from the splenic hilum region using the LASSO regression analysis. The intraobserver ICCs ranged from 0.883 to 0.990 and the interobserver ICCs ranged from 0.839 to 0.935, indicating favorable intra-and interobserver feature extraction reproducibility. Thus, the radiomics signature (Rad score) formula was generated using a linear combination of selected features that were weighted by their respective LASSO coefficients. The detailed calculation formula was provided in the supplementary materials.

Diagnostic performance of the rad score for VNT
The Rad score showed a good diagnostic performance for identifying VNT in cirrhotic patients with the AUC of 0.983 (95% CI 0.964-1.00) in the training set. And the Rad score yielded the AUC of 0.970 (95% CI 0.930-1.00) and 0.932(95% CI 0.853-1.00) in internal and external validation sets, respectively ( Supplementary Fig. S1).

Radiomics nomogram development and validation
For univariate analysis, the Rad score, platelet counts, albumin, and spleen diameters were found significantly associated with VNT. When multiple logistic regression analyses were performed, the Rad score and platelet counts remained significant for VNT (Table 2). Then, a radiomics nomogram for predicting VNT was constructed using the above regression coefficients (Fig. 2). The point-biserial correlation test indicated a strong positive correlation between the radiomics-platelet nomogram and the CT imaging findings (r = 0.762, p < 0.001). The calibration curve and a nonsignificant Hosmer-Lemeshow test statistic (p = 0.49) showed good calibration in training sets. The calibration curve was confirmed in the internal and external validation sets with a nonsignificant Hosmer-Lemeshow test statistic p = 0.29 and p = 0.33, respectively (Fig. 3A). The radiomics nomogram showed favorable predictive efficacy with AUC of 0.987 (95% CI 0.969-1.00), which confirmed with AUC of 0.973 (95% CI 0.939-1.00) and 0.947 (95% CI 0.876-1.00) in internal validation set and external validation set, respectively (Fig. 3B). Compared with other noninvasive methods that were proposed in previous researches such as PSR, APRI, and FIB-4, the radiomics nomogram still showed the highest diagnostic performance. The comparison between the radiomics nomogram and liver stiffness measured by To evaluate the clinical usefulness of our radiomics nomogram, decision curve analysis (DCA) was applied in this study (Fig. 3C). The DCA curve indicated that if the threshold probability for a patient or a doctor is within a range from 0 to 0.95, the radiomics nomogram adds more benefit than either the treat-all-patients or the treat-none strategies. Within this range, the radiomics nomogram also was better than the PSR which showed a good performance in predicting VNT in previous studies.

Efficacy and safety assessment
Furthermore, we evaluated the efficacy and the safety of the radiomics nomogram for the prediction of VNT in cirrhotic patients. As shown in Table 3, the radiomics nomogram could spare 42.3%, 49.3%, and 44.4% endoscopies with a low risk of VNT missed (< 5%) and no more than 8.3% of unnecessary endoscopic examinations still be performed in the training set, internal validation set, and external validation set, respectively. When compared with the radiologist, the radiomics nomogram was more accurate than radiologist interpretation, especially in the discrimination of mild varices. In our study, the radiologist could classify 13 patients among 27 patients with mild varices correctly while the radiomics nomogram could classify 23 patients correctly. When compared with Baveno VI criteria in a subgroup, the radiomics nomogram still showed better performance. The detailed result comparing to Baveno VI criteria was provided in supplementary materials.

Discussion
Poor compliance to endoscopic examination and excessive unneeded endoscopic examination are the two major problems that we encountered in clinical practice. For precise prevention for varices bleeding, we aim to develop and validate a noninvaisve, convenient, easily accessible and highly accurate method for the detection of VNT in cirrhosis.
Radiomics has been recognized as an emerging image analysis technology in clinical disease assessment which could extract a large scale of image features in a noninvasive way and describes modules for quantitatively assessing the distribution of gray level and pixels in medical images, and these features cannot be detected by human vision. In recent years many studies explored the application of radiomics in the diagnosis, treatment and prognosis of the disease and the results are satisfying [21]. A prospective multicenter study developed a radiomics model that showed good performance in the prediction of patients with HVPG > 10 mmHg [19]. And following studies used similar radiomics methods to predict gastroesophageal varices needing treatment and varices rebleeding [18,22], but the results were not satisfying. When predicting HVPG and GEV by analyzing an abdominal CT image, the main difference between them is that GEV is visualized. However, the former radiomics model lacked radiomics features from gastroesophageal ROI. In our study, we developed an improved radiomics model containing the esophageal and gastric radiomics features and this radiomics nomogram showed excellent performance in the prediction of VNT in cirrhotic patients.
For the construction of Rad score, 19 potential predictors with 14 features from the gastroesophageal region and 5 from the splenic hilum region were obtained. Major of Fig. 3 a The calibration curve of the radiomics nomogram for prediction of VNT in cirrhotic patients in training and validation sets. Calibration curves depict the calibration of the radiomics nomogram in terms of agreement between the predicted risk of VNT and observed outcomes. The blue line represents a perfect prediction, and the dotted pink lines represent the predictive performance of the nomogram. The closer the dotted line fit is to the ideal line, the better the predictive accuracy of the nomogram is. b The ROC of the radiomics nomogram for prediction of VNT. c Decision curve analysis for radiomics nomogram and PSR. The y axis measures the net benefit. The red line represents the radiomics nomogram. The blue line represents the PSR. The black line represents the hypothesis that no patients had VNT. The gray line represents the hypothesis that all patients had VNT. The x axis represents the threshold probability. The threshold probability is where the expected benefit of treatment is equal to the expected benefit of avoiding treatment. For example, if the possibility of VNT involvement of a patient is over the threshold probability, then a treatment strategy for VNT should be adopted. VNT varices needing treatment, PSR platelet-spleen ratio, ROC receiver-operator characteristic, AUC area under the curve, PSR platelet-spleen ratio, APRI AST-to-platelet ratio index, FIB-4 fibrosis-4 score potential predictors are from the gastroesophageal region which confirmed that gastroesophageal radiomics features are more important. However, none of the hepatic features while five splenic features were selected by LASSO analysis. This finding may result from the prevalence of splenomegaly is over 50% in our study and the extrahepatic factors contribute more to the rise of portal pressure. The recent studies suggested that splenomegaly may be superior to liver fibrosis in the prediction of VNT [23][24][25][26][27][28]. Our data also showed that PSR which reflects the splenomegaly performed a better diagnostic ability than liver fibrosis-associated parameters such as APRI and FIB-4.
In our study, the 19-features-based Rad score was found to be effective for VNT classification, this Rad score could stratify patients into non-VNT and VNT group with an AUC of 0.983 (95% CI 0.964-1.00), 0.970 (95% CI 0.930-1.00) and 0.932(95% CI 0.853-1.00) in the training set, internal validation set and external validation set respectively. Next, we considered clinical risk factors, a multivariate logistic regression analysis for accessible clinical parameters indicated that platelet count was a significant predictive factor distinct from Rad score. The diagnostic performance was further improved by combining the Rad score with platelet counts, achieving an AUC of 0.987 (95% CI 0.969-1.00), 0.973 (95% CI 0.939-1.00) and 0.947 (95% CI 0.876-1.00) in the training set, internal validation set and external validation set, respectively. Although the AUC was mild improvement when Rad score combined with the platelet count, the combined nomogram could reduce the risk of missing VNT and unneeded endoscopies in the following evaluation. When compared to other noninvasive methods that were comprehensively validated in previous studies such as PSR, APRI, and FIB-4, both Rad score and radiomics nomogram showed better performance in the prediction of VNT.
Although our radiomics nomogram showed excellent discrimination and calibration, it could not represent favorable clinical usefulness. Previous studies seldom assess the clinical usefulness and safety of the VNT-prediction radiomics model. In this study, we applied decision curve analysis and the result suggested that our radiomics nomogram could derive good net benefit in clinical application. Furthermore, our radiomics nomogram could spare more than 40% of endoscopic examinations with a low risk of missing VNT, and no more than 8.3% of unneeded endoscopies still be performed.
Previous studies demonstrated that radiologists could distinguish VNT with an accuracy of nearly 90% by analyzing portal-phase abdominal CT images [29]. The radiologists could predict the risk of VNT by observing enhanced esophageal or periesophageal varices, splenomegaly and the total portosystemic shunt area. In our study, the accuracy of medical image reports was 90.1%, 87.3%, and 84.1% in the training set, internal validation set, and external validation set, respectively. This result suggested radiologist interpretation may be unstable owing to the subjectivity and experience of radiologists. In addition, it may be difficult for the radiologist to distinguish the mild varices and VNT sometimes. When compared with radiologist interpretation, the radiomics nomogram yielded a robust accuracy and better performance in the classification of mild varices and VNT overall. Therefore, the application of the noninvasive and reproducible radiomics-platelet nomogram could improve screening compliance in cirrhotic patients. Another merit of the developed radiomics-platelet nomogram is that it could significantly reduce the unnecessary endoscopic examination and medical burden due to its excellent discrimination with high sensitivity and specificity.
When compared with the previous studies reporting the radiomics model for detecting VNT, our study has the following advantages. First, our study has a larger study population and validated our radiomics model in an independent external population, which improved test power and predictive ability. Secondly, we analyzed not only hepatic and splenic features but also gastroesophageal features, which contribute to a better diagnostic performance than the previous radiomics model. In addition, our study has more detailed validation in efficacy and safety assessment which previous studies lacked.
Some limitations of our study should be discussed. One of the potential criticisms is our training set was collected retrospectively and not large, so we prospectively enrolled more patients as the validation set to reduce the related deviation. Another shortcoming of our study was that we only performed liver stiffness measurement in those patients who were hardly diagnosed with cirrhosis according to clinical presentation, blood tests, or medical images. Only a total of 42 patients with liver stiffness measurement were enrolled in the internal and external validation set, so we only evaluated the comparison of radiomics nomogram and Baveno VI criteria in this subgroup. A more rigorous comparison will be studied in our future researches. Thirdly, an application limitation of the developed radiomics-platelet nomogram is its price. The radiomics analysis in our study was based on the enhanced contrast CT examination which is expensive than liver transient elastography (TE), but the enhanced contrast CT examination could provide more information such as portosystemic shunt, portal thrombosis and the sign of liver cancer rather than a simple value of liver stiffness provided by TE. Lastly, medical image is easily accessible in routine clinical practice and many studies explored the potential application of advanced radiological technology in the diagnosis of liver fibrosis and esophageal varices. For example, some studies used diffusion-weighted magnetic resonance imaging for quantification of liver fibrosis and prediction of esophageal varices which showed encouraging results [30][31][32]. We lacked the comparison between our developed radiomics nomogram and other advanced radiological technology. More studies about the comparison and cost-effective assessment of the radiomics-platelet nomogram in the future are needed, which will help us to find its place in clinical practice.

Conclusion
In this study, we developed and validated a novel, diagnostic radiomics-based nomogram which is a reliable and noninvasive method to predict VNT in cirrhotic patients.