Single-center, Randomized, Superiority Pivotal Clinical Study to Evaluate the Efficacy of Artificial Intelligence-based Upper Gastrointestinal Endoscopy Image

We will conduct a single-center retrospective study at a university hospital. A total of 3,385 gastroscopic white-light images from patients with pathologically confirmed findings will be analyzed. The AI software will automatically identify images as non-neoplastic or neoplastic (low-grade dysplasia, high-grade dysplasia, early gastric cancer with mucosal or submucosal invasion, or advanced gastric cancer) and highlighted lesion locations. Two experienced endoscopists will independently review the same image set without AI assistance for comparison. Primary outcomes are sensitivity and specificity of the AI in detecting gastric neoplasms (by category and overall), and the localization accuracy measured by the localization receiver operating characteristic (LROC) curve area. Secondary outcomes is includes comparison of the AI's diagnostic performance with that of endoscopists.

Inclusion criteria

Age 19 or older At least one gastric lesion biopsied with a definitive pathological diagnosis Availability of high-quality white-light endoscopy images of the lesion and surrounding mucosa

Exclusion criteria

Poor-quality images (e.g., out of focus or obscured) Lack of histopathological confirmation of the lesion

Each image will be paired with a reference standard diagnosis based on the pathology result for that lesion or region.

Inclusion criteria

• Age 19 or older
• At least one gastric lesion biopsied with a definitive pathological diagnosis
• Availability of high-quality white-light endoscopy images of the lesion and surrounding mucosa

Exclusion criteria

• Poor-quality images (e.g., out of focus or obscured)
• Lack of histopathological confirmation of the lesion