Rebleeding is a real emergency, with significant mortality and high medical costs. In this study, the rate of rebleeding was 35.77% after endoscopic treatment.The purpose of this study is to establish non-invasive models according to the CT radiomics features of different organs, to identify groups of patients at high risk of rebleeding after endoscopic treatment, and to intervene early and improve the prognosis. Among the organ-based CT radiomics models, the liver-based model performed better than the spleen- or esophagus-based model. And the predicting performance could be further enhanced by combining the clinical model and the liver-based CT radiomics models.
Some previous studies have shown that some laboratory data and imaging findings are associated with bleeding[23, 24] and rebleeding[25] of EGVs.To our knowledge, the higher the portal vein pressure or if patients were infected before treatment, the greater the risk of rebleeding[26–28]. In our study, the factors indirectly reflecting the level of portal vein pressure (portal vein diameter and upper and lower spleen diameter) and RBC and LYM counts were also associated with rebleeding.Meanwhile, rebleeding was also related to ALP, which may indicate that patients with cholestasis are more likely to experience rebleeding. The model based on the clinical characteristics had an AUC and accuracy of 0.818 and 77.2%, respectively, in our study.
Some studies have proved that a CT radiomics model can noninvasively detect portal hypertension36and predict the first bleeding[29]. Portal hypertension is the direct cause of varices, which indicates that the CT radiomics model is a feasible method of predicting rebleeding within 1 year after endoscopic treatment of EGVs. Three independent organ-based radiomics models were constructed based on the radiomics features extracted from the CT images of liver, spleen, and lower esophagus. Compared with the clinical model, the diagnostic efficiencies of the radiomics models constructed from the liver and spleen were better. In chronic liver disease, portal hypertension and EGVs are associated with increased resistance to portal blood flow because of the obvious architectural changes in the liver vasculature (fibrosis, nodules, angiogenesis, and sinusoidal remodeling)[30].The pathophysiological changes of the liver directly lead to EGVs, and the texture features of the liver can reflect the changes of the liver microenvironment[31]. The irreversible microstructural changes of the liver lead to EGVs, which can be better reflected by the radiomics features of CT images. So, the liver-based radiomics model is the most effective in predicting rebleeding. Among the three organ-based radiomics models, the lower esophageal-based radiomics model performed poorly compared with the others. This may be related to the fact that the esophagus is in motion, and the lower esophagus may be in different states (systolic or diastolic) during the CT scans. At the same time, our study only extracted 2D images, which might not reflect the situation of the varicose veins stereoscopically and comprehensively. This can be further improved in the future.
At the same time, the performance of the radiomics-clinic combined models was better compared with the independent clinical model or radiomics models. Among the combined radiomics models, the clinical-liver model had the highest AUC of 0.931 (95% CI: 0.887–0.974). In our study, creating further combinations of the liver, spleen, esophagus, and clinical models did not significantly enhance the AUC. Although the process of image feature extraction and combination model establishment of some of these models was rather complicated, they did not exhibit better prospects for clinical application. This further illustrates that the liver CT texture features constitute a relatively realistic, objective reflection of the internal liver microenvironment changes. Therefore, the liver radiomics model and clinical-liver combined model are based on good clinical judgment as to the likelihood of rebleeding. At the same time, the non-invasive CT radiomics model is also a simple and cost-effective method of predicting rebleeding, which may be more readily accepted by patients. In conclusion, the clinical-liver combined model could have the potential to evaluate the prognosis of endoscopically treated patients, which will aid clinical decision making, and ensure that effective secondary prevention measures are implemented in a more timely manner to reduce overall mortality. To our knowledge, this study may have derived the first CT-based radiomics models to consider multiple organs for predicting rebleeding after endoscopic treatment of cirrhotic varices.
There are some limitations in this study. First, the sample size of included cases was small and we are trying to collect more data to improve the model generalization. Second, there is a lack of multi-center controlled trials and this study lacks external validation. Therefore, the image standardization, ICC-based feature reliability and internal validation by1000-times bootstrapping were carefully considered in the current study to guarantee the rationality of the model as much as possible. Third, the largest cross-section of the organ used in this study might lead to the loss of important information. In future work, a more detailed exploration of the region of interest selection for radiomics analysis and whether image-based deep learning without ROI segmentation could be applicable should be considered and attempted.
In conclusion, the independent and combined liver-based CT radiomics models performed well in predicting rebleeding within 1 year after endoscopic treatment of EGVs. The organ-based radiomics models might provide helpful information in addition to clinical factors.