A deep learning and radiomics fusion model enhances endoscopic ultrasonography diagnosis of gastric tumors.

Objectives: Gastric tumors are one of the most common digestive system diseases, and their accurate diagnosis largely depends on the expertise of medical professionals in interpreting imaging data. As medical imaging technology has advanced, the volume of imaging data for gastric tumors has rapidly increased. This wealth of data provides valuable diagnostic information but also poses challenges for experts in diagnosis and analysis. To address these challenges, a more accurate and efficient diagnostic model is urgently needed. Methods: In this study, we propose a gastric tumor classification model that combines deep learning and radiomic features to differentiate gastrointestinal stromal tumors from other gastric tumor types. This retrospective study collected 4,806 endoscopic ultrasound images from 219 patients. Deep learning and radiomic features were extracted, and feature selection was performed using t-tests and the LASSO method. Additionally, clinical features were recorded. Single-parameter models were constructed based on these three categories of features. Following this, a gastrointestinal tumor classification model was developed by integrating the predictions from the single-parameter models. Results: The model achieved an area under the receiver operating characteristic curve of 0.95, with sensitivities, specificities, positive predictive values, negative predictive values, and accuracies of 100.00%, 82.86%, 60.00%, 100.00%, and 86.36%, respectively. Conclusions: The experimental results indicate that the gastric tumor classification model has demonstrated strong performance in classifying gastric tumors using endoscopic ultrasound images, offering valuable diagnostic support for clinicians.