Objective To explore predictive value of radiological indexes for hemorrhage in patients with portal hypertension. Methods The clinical data and radiological data of patients with portal hypertension accompanied with hepatitis B from June 2008 to June 2014 in the Nanjing Drum Tower Hospital were analyzed retrospectively. Patients with hepatocellular carcinoma, portal vein thrombosis, or portal hypertension due to other causes, such as autoimmune hepatitis, pancreatitis, or hematological diseases were excluded. Results Ninety-eight patients were studied and subsequently divided into a hemorrhage group (n=57) and a non-hemorrhage group (n=41). There were no statistical differences in the clinical indexes such as the age, prothrombin time, serum albumin, serum creatinine, serum sodium, white blood cell count, and blood platelet count (P>0.05). However, the differences were statistically significant in the serum total bilirubin, hemoglobin, and liver function with theP values of 0.023, 0.000, and 0.039, respectively. For the radiological indexes, the hemorrhage was correlated with the diameter of posterior gastric vein (P=0.028 3) or grading of esophageal varices (P=0.022 1). Logistic procedure was used to construct the model with stepwise selection and finally the diameter of inferior mesenteric vein, diameter of posterior gastric vein, grading of esophageal varices, and diameter of short gastric vein were enrolled into this model. These indexes were scored, the risk of bleeding increased with increasing the points. Then the model was validated with 26 patients with portal hypertension from July 2014 to December 2014, the area under the receiver operating characteristic curve was 0.884 9 by this radiological model. Conclusions A radiological scoring model is constructed including diameter of inferior mesenteric vein, grading of esophageal varices, diameter of posterior gastric vein, and diameter of short gastric vein, which might predict risk of hemorrhage in patients with portal hypertension. However, further protective study of large sample is needed to validate this model.