Development and validation of an ultrasound-based AI-radiomics model for diagnosing and risk-stratifying gastrointestional stromal tumors: a retrospective diagnostic study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Ultrasound, The First Affiliated Hospital of Chongqing Medical University, No. 1 Youyi Road, Yuzhong District, Chongqing, 400016, China.
- School of Resources and Safety Engineering, Chongqing University, No. 174 Shazhengjie, Shapingba District, Chongqing, 400044, China.
- Department of Gastroenterology, The First Affiliated Hospital of Chongqing Medical University, No. 1 Youyi Road, Yuzhong District, Chongqing, 400016, China. [email protected].
- Department of Ultrasound, The First Affiliated Hospital of Chongqing Medical University, No. 1 Youyi Road, Yuzhong District, Chongqing, 400016, China. [email protected].
Abstract
Accurate discrimination between gastrointestinal stromal tumors (GISTs) and leiomyomas is essential for guiding clinical management. This study developed and validated three predictive models: a baseline characteristics model (BCM), an endoscopic ultrasound-based morphological model (EUS-Morph), and a multimodal fusion model (MMF) incorporating clinical, EUS, and radiomics features. Their diagnostic performance in distinguishing GISTs from leiomyomas was systematically evaluated, along with their potential utility for GIST risk stratification. A retrospective analysis was conducted on 3,393 EUS images from 265 pathologically confirmed GIST/leiomyoma patients to differentiate tumors and stratify GIST risk. The diagnostic performance of various models was compared with each other and against endoscopists in the development cohort. We screened out several key independent variables that distinguish leiomyoma from GIST, including the Length/Short-axis (L/S) Ratio. After comparison, The MMF Model demonstrated optimal diagnostic and stratification performance, achieving the areas under the curve (AUC) values of 0.975 and 0.992, respectively. In contrast, the AUC range for diagnosing GIST by three experienced endoscopists was 0.68–0.70, and the AUC range for risk stratification was 0.72–0.74. Its performance in diagnosing GIST is over 40% higher than that of experienced endoscopists. All three developed AI models demonstrated favorable performance. Notably, the MMF model, which integrates patient baseline data, clinical characteristics, two-dimensional EUS images and radiomic features, effectively differentiated GISTs from leiomyomas GILs and achieved accurate risk stratification of GISTs. This EUS image-based integrated AI-radiomics model demonstrated improved diagnostic performance compared to endoscopists and may serve as a complementary tool in clinical practice. The online version contains supplementary material available at 10.1186/s12880-025-02050-z.