Objective To establish and validate methods for differentiating gastrointestinal stromal tumor (GIST) from non-GIST based on endoscopic ultrasound radiomics and machine learning. Methods A total of 435 eligible patients were enrolled, and 3 279 endoscopic ultrasound images of GIST (257 cases) and non-GIST (including 145 cases of gastric leiomyoma and 33 cases of schwannoma) were collected and assigned (case proportion, 7:3) to training set or test set. Pyradiomics software was used to extract tumor radiomics features, and principal component analysis (PCA), least absolute shrinkage and selection operator (LASSO), extreme gradient boosting (XGBoost), random forest, and recursive feature elimination (RFE) algorithms were used to design feature screening schemes. Based on the selected features, the models were established by support vector machine classifier. Receiver operating characteristic (ROC) curve was used to evaluate the predictive performance of the models for GIST and non-GIST. Results The radiomics prediction models were established based on the selected features. The area under curve values of 5 models based on different feature screening methods (PCA, PCA+LASSO, PCA+XGBoost, PCA+random forest, and PCA+RFE) were 0.581, 0.870, 0.874, 0.860, and 0.661, respectively. Conclusion PCA+XGBoost algorithm has the best feature screening effect. A model based on the radiomics and machine learning methods in this study for distinguishing GIST from non-GIST can be used for preoperative prediction of patients.