Risk stratification system of gastrointestinal stromal tumors under endoscopic ultrasound elastography based on artificial intelligence

Background and Objectives Gastrointestinal stromal tumors (GISTs) are tumors with malignant potential, and preoperative risk stratification is critical for clinical management. Endoscopic ultrasound elastography (EUS-E) can assess tissue stiffness and may assist in evaluating the malignant potential of GISTs. However, similar studies have not been conducted, and current elasticity evaluation methods are still highly influenced by operator's subjectivity. An effective and objective tool is needed to aid in the risk stratification of GISTs under EUS-E. Methods One hundred eighty-nine patients with submucosal tumors (SMTs) who underwent EUS-E from January 2018 to August 2024 were retrospectively collected, of which 110 cases were GISTs. A total of 2625 EUS B-mode images were collected to construct the classification and segmentation model to distinguish GISTs from non-GISTs and to segment the lesion areas of GIST automatically. The elasticity value (EUS-E-AI) of the lesion area was extracted based on the color features of the elastography images. We evaluated the diagnostic performance of this system in distinguishing GISTs from other SMTs, as well as its ability to stratify GISTs based on their malignant potential. Results The accuracy of the classification model and the Dice coefficient of the segmentation model were 95.8% and 0.967, respectively. The EUS-E-AI value in the low-risk malignancy group (0.268 [IQR, 0.243–0.333]) was significantly higher than that in the high-risk malignancy group (0.186 [IQR, 0.176–0.199], P < 0.001). A cutoff value of 0.224 for the EUS-E-AI was found to effectively differentiate the low-risk from the high-risk group, with an accuracy of 92.6% (95% CI, 89.1–96.1). These findings were also confirmed in small GISTs. Conclusion We developed an AI-based system and elasticity indicator for the accurate and objective identification and risk stratification of GISTs using EUS.