Artificial Intelligence (AI), Gastric Leiomyoma, Gastric Subepithelial Tumors, Gastrointestinal Stromal Tumors, Multimodal Imaging
Conditions
Brief summary
Background: Gastrointestinal Stromal Tumors (GISTs) are the most common mesenchymal tumors of the gastrointestinal tract. Accurate pre-operative diagnosis, risk stratification, and genotyping are critical for determining the appropriate surgical approach and targeted therapy (such as Imatinib). However, current methods often rely on invasive postoperative pathology and expensive genetic testing. Study Objective: The purpose of this study is to develop and validate a multimodal Artificial Intelligence (AI) model that integrates clinical data, CT radiomics (imaging features), and pathomics (digital pathology features) to improve the precision of GIST management. Study Design: This is a prospective, observational study. The researchers will recruit patients with suspected gastric submucosal tumors who are scheduled for surgery or biopsy at The Fourth Hospital of Hebei Medical University. Core Tasks: The AI model will be trained to perform three specific tasks: Diagnosis: Distinguish GISTs from other non-GIST mesenchymal tumors (e.g., leiomyomas, schwannomas). Risk Assessment: Stratify GISTs into risk categories (e.g., Low vs. High risk) to predict malignant potential. Genotyping: Predict specific gene mutations (e.g., KIT or PDGFRA mutations) to guide immunotherapy or targeted therapy. Methodology: Patient data (CT scans, pathology slides, and clinical history) will be collected and analyzed by the AI system. The AI's predictions will be compared against the "Gold Standard" results derived from postoperative pathological examination and Next-Generation Sequencing (NGS). This study is non-interventional; the AI results will not affect the standard of care received by the patients.
Interventions
DIAGNOSTIC_TESTMultimodal AI Analysis System
CT-based multitask deep learning system (GIST-Net). Input is the routine preoperative contrast-enhanced CT only; no pathology, molecular or laboratory data are used at inference, and no extra imaging, radiation, blood sampling or biopsy is required. The tumour is segmented on the portal venous phase, and four task-specific heads output probabilities for: (1) GIST vs non-GIST submucosal lesions; (2) modified NIH risk category; (3) driver genotype (KIT exon 11/9, PDGFRA non-D842V, D842V, wild-type); (4) recurrence within 24 months after R0 resection. Steps 2-4 are conditioned on step 1. A prespecified reader component evaluates human-AI interaction: 10 radiologists of three experience levels read the same cases unaided, then re-read with model scores and attention maps after a 4-week washout, giving paired within-reader comparisons of AUC, accuracy, agreement, confidence and reading time. Observational only; outputs are blinded to treating physicians and do not affect management.
Sponsors
Qun Zhao
Study design
Observational model
COHORT
Time perspective
PROSPECTIVE
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Inclusion criteria
Age ≥ 18 years, gender not limited. Clinical diagnosis of gastric submucosal tumor (SMT) or suspected gastrointestinal stromal tumor (GIST) based on gastroscopy or ultrasound. Scheduled for surgical resection or endoscopic biopsy at the study center. Standard preoperative contrast-enhanced CT scans are available (performed within 2 weeks prior to surgery). Patients or their legal guardians have signed the informed consent form.
Exclusion criteria
Received neoadjuvant therapy (e.g., Imatinib, chemotherapy, or radiotherapy) prior to surgery/biopsy. Poor quality of CT images (e.g., severe motion artifacts) affecting radiomics analysis. Insufficient tissue samples for pathological diagnosis or genetic testing. Confirmed diagnosis of other primary malignancies. Incomplete clinical data or lost to follow-up immediately after surgery.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Diagnostic Accuracy of the AI Model for Distinguishing GIST from Non-GIST Tumors | Up to 30 days post-surgery | The diagnostic accuracy is calculated as the proportion of correctly classified patients (GIST vs. Non-GIST) by the multimodal AI model, compared to the gold standard postoperative pathological diagnosis. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Concordance Rate between AI-predicted Risk Grade and Pathological Modified NIH Criteria | Up to 30 days post-surgery | The proportion of patients whose risk category (Very Low/Low vs. Intermediate/High) predicted by the AI model matches the actual risk grade determined by postoperative pathology according to the modified National Institutes of Health (NIH) criteria. This will be reported as a percentage (0-100%) |
| Sensitivity and Specificity of the AI Model in Predicting KIT/PDGFRA Gene Mutations | Up to 30 days post-surgery | The AI model's performance in identifying specific mutations (e.g., KIT exon 11, PDGFRA) compared to the results of Next-Generation Sequencing (NGS). Data will be reported as percentages with 95% confidence intervals. |
| Area Under the Receiver Operating Characteristic Curve (AUC) for All Tasks | Up to 30 days post-surgery | The AUC values will be calculated to evaluate the overall performance of the AI model in diagnosis, risk stratification, and genotype prediction. Sensitivity and Specificity will also be reported. |
Countries
China
Outcome results
None listed