Objective To investigate the role of endoscopic ultrasonography (EUS) in differentiating between autoimmune pancreatitis (AIP) and pancreatic cancer (PC). Methods Data of 133 patients with AIP and 113 patients with PC who underwent EUS because of obstructive jaundice at Peking Union Medical College Hospital from January 2013 to December 2018 were retrospectively analyzed in the study, and were randomly divided into either a derivation sample or a validation sample using 1∶1 allocation according to the random number. In the derivation sample, 10 EUS characteristics were used to construct a prediction model to distinguish between AIP and PC, in which predictors were identified by multivariate stepwise logistic regression analysis and predictive efficacy was evaluated by receiver operating characteristics (ROC) curve analysis. The predictive efficacy was assessed in the validation sample. In view of the subjectivity in the judgment of diffuse/focal hypoechogenicity, 2 prediction models were designed in order to avoid bias. Results By multivariate stepwise logistic regression analysis, diffuse hypoechogenicity (OR=591.0, 95%CI: 98.8->999.9, P<0.001) and vessel involvement (OR=11.9, 95%CI: 1.4-260.2, P=0.023) were identified as statistically significant predictors for distinguishing AIP from PC. EUS characteristics excluding diffuse/focal hypoechogenicity were stepped by logistic regression, which showed that hyperechoic foci/strands (OR=177.3, 95%CI: 18.7->999.9, P<0.001), pancreatic duct dilation (OR=60.5, 95%CI: 6.2->999.9, P=0.004), bile duct wall thickening (OR=35.4, 95%CI: 3.7->999.9, P=0.009), lymphadenopathy (OR=16.8, 95%CI: 1.7-475.2, P=0.038) and vessel involvement (OR=22.7, 95%CI: 2.0-725.7, P=0.028) were statistically significant predictors to distinguish the two diseases. Both prediction models were built in the derivation sample, with area under the ROC curve of 0.995 and 0.979 respectively. In the validation sample, sensitivity, specificity, accuracy, positive predictive value and negative predictive value of both prediction models were all >90% by using the optimal cutoff value. Even for discrimination between focal AIP and PC, sensitivity and accuracy of both models were >90%, and specificity, positive predictive value and negative predictive value were all >85%. Conclusion The 2 prediction models have good differential predictive value, and EUS is a useful tool to differentiate between AIP and PC.