Enhancing gastrointestinal stromal tumor risk stratification: A novel deep learning approach applied to EUS imaging.

Background and Objectives: Effective management of gastrointestinal stromal tumors (GISTs) requires accurate risk assessment due to their variable carcinogenic potential.This study aimed to improve GIST classification using a novel deep learning-based GIST risk prediction model based on EUS imaging modalities. Methods: We retrospectively analyzed EUS images of 341 patients diagnosed with GIST at a tertiary medical center between January 2016 and March 2022. Patients were selected based on specific criteria, including pathologic validation of surgical outcomes. The dataset allowed for the development and validation of a deep learning risk prediction model (DLRPM), a traditional risk prediction model (TRPM), and a combined risk prediction model (CRPM). Model performance was evaluated using sensitivity, specificity, positive and negative predictive values, accuracy, and statistical analysis. Results: The efficacy of the 3 prognostic models (TRPM, DLRPM, and CRPM) for GIST classification was evaluated using a dataset consisting of 1019 EUS images from 341 patients. These models were developed using a training subset of 310 patients and subsequently validated in a defined group of 31 consecutive patients. Using multivariate logistic regression, TRPM showed an accuracy rate of 71.10%. Using a DenseNet-121 framework developed specifically for medical imaging, the DLRPM demonstrated superior predictive capabilities with an accuracy rate of 92.65%. The CRPM achieved a prediction accuracy of 90.32%. In addition, receiver operating characteristic curve analysis revealed area under the curve values of 0.909, 0.932, and 0.843 for CRPM, DLRPM, and TRPM, respectively. However, comparative statistical evaluation between these models showed no significant differences in area under the curve. Conclusions: The novel DLRPM improved the accuracy of GIST risk stratification by EUS. CRPM is a promising method to integrate deep learning with traditional statistical methods, potentially optimizing clinical decision-making and patient outcomes in gastroenterological oncology.