To investigate the value of endoscopic ultrasound (EUS)-based deep learning, radiomics, and their combined model in differentiating gastric gastrointestinal stromal tumors (GIST) from leiomyomas.

Clinical and imaging data from 895 patients (totaling 4, 802 EUS images) at the First Medical Center of the Chinese PLA General Hospital between January 2010 and April 2025 were retrospectively analyzed.Cases were randomly divided into training and test sets at an 8∶2 ratio based on lesions.Radiomic features and fine-grained deep learning features based on the Transformer architecture were extracted from EUS images to construct single-modality and combined-modality machine learning classification models (support vector machine, extreme gradient boosting, random forest).Model performance and clinical utility were comprehensively evaluated using receiver operating characteristic (ROC) curves, the DeLong test, calibration curves, and decision curve analysis.

In the test set, both the deep learning model and the combined model demonstrated significantly better diagnostic performance than the traditional radiomics model. Among these, the combined model using the support vector machine achieved the best performance, with an area under the curve (AUC), accuracy, sensitivity, and specificity of 0.852, 0.792, 0.831, and 0.707, respectively.

EUS-based deep learning features serve as the primary driver for differentiating gastric GIST from leiomyomas, with diagnostic performance significantly superior to that of traditional radiomic features.