Multicenter Observational Study of Multimodal AI for Upper GI Mesenchymal Tumor Diagnosis

Overview

This study develops a multimodal AI model using endoscopic ultrasound, white-light endoscopy, and clinical information to support the diagnosis of upper GI mesenchymal tumors and the risk stratification of gastric GISTs.

This is a multicenter, observational study designed to evaluate the diagnostic performance of a multimodal artificial intelligence (AI) model for the classification of upper gastrointestinal subepithelial lesions (SELs) and risk stratification of gastric gastrointestinal stromal tumors (gGISTs). The study combines retrospective image data for training and validation with prospectively recruited cases for testing. Endoscopic ultrasound (EUS) images, white-light endoscopy (WLE) images, and relevant clinical data will be collected according to strict image quality control criteria. The multimodal AI model integrates these inputs using a multi-branch fusion strategy. A cross-validation trial will be conducted using prospectively recruited patients' data from multiple centers to compare the diagnostic and predictive performance of endoscopists with and without AI assistance for both lesion classification and risk stratification. According to existing literature, no multimodal AI model has yet reported diagnostic performance for classifying SELs or for risk stratification of gastric gGISTs. It is assumed that the multimodal AI model will achieve a diagnostic accuracy of 95% for classifying upper gastrointestinal SELs and 95% for gGIST risk stratification. In comparison, the diagnostic accuracy of endoscopists is approximately 73.3%-75% for differentiating GIST from non-GIST and 72.4%-78.9% for risk stratification of gGISTs . GISTs account for about 67-68% of all lesions . Using a two-sided confidence interval with α = 0.05 and β = 0.2, and considering a 20% potential dropout rate, the minimum sample size required for prospective SEL classification is 65 cases, and 88 gGIST cases for risk stratification. Since the risk stratification task requires a larger sample size and GISTs are the common target of both tasks, the final planned sample size is 130 patients with upper GI SELs, which meets the statistical requirements for all primary endpoints. The study team will screen patients based on the inclusion and exclusion criteria, ensure that all required examinations are completed to confirm eligibility, and record pre-treatment test results. All prospective participants will provide written informed consent before any study-related examinations. This study is purely observational. No additional interventions will be performed on participants, nor will any additional costs be incurred. Patients' access to optimal diagnostic or treatment options will not be affected. The primary potential risk is the breach of patient privacy; the research team will establish a strict data security and monitoring plan and inform participants that their data will be used for clinical research purposes. This study is purely observational. No additional interventions will be performed on participants, nor will any additional costs be incurred. Patients' access to optimal diagnostic or treatment options will not be affected. The primary potential risk is the breach of patient privacy; the research team will establish a strict data security and monitoring plan and inform participants that their data will be used for clinical research purposes. Each enrolled participant will undergo diagnostic assessment by both the multimodal AI model and expert endoscopists. The AI model and expert interpretation will be blinded to each other. Final diagnosis will be confirmed by histopathology. Diagnostic performance will be compared using paired analysis. All statistical tests will be two-sided, and differences will be considered statistically significant at P \< 0.05. Continuous variables will be described as mean ± standard deviation. Categorical variables will be presented as counts and percentages. (1) Diagnostic Performance: The diagnostic performance of endoscopists and the AI model will be compared using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and area under the curve (AUC). F1-score (harmonic mean) and balanced accuracy will be calculated to address class imbalance (e.g., GIST vs. other lesions). (2) Continuous Data: Comparisons with baseline values will be conducted using paired t-tests, ANOVA, or rank-sum tests as appropriate. (3) Categorical Data: Group comparisons will use Chi-square tests (including CMH Chi-square test) or Fisher's exact test. (4) Baseline Comparability: Demographic and baseline characteristics will be compared using independent t-tests or Chi-square tests to assess group balance. (5) Effectiveness Analysis: The primary effectiveness endpoint is the diagnostic accuracy for GI subepithelial lesions. The difference in proportions and Youden index will be compared using the approximate normal Z test or Chi-square test with center effect control. (6) Software: All statistical analyses will be performed using SPSS version 26.0.

Conditions

Interventions

Eligibility

Locations (1)

Outcomes

Central Contacts