GutCore: An Endoscopy Foundation Model for Whole-Case Gastric Cancer Analysis
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Routine endoscopy image sets contain complementary information about lesions, surrounding mucosa, anatomy, and examination context. However, most endoscopic artificial intelligence systems are developed and evaluated on selected frames for narrowly defined image-level tasks, limiting their ability to support case-level analysis.
Objective
To develop GutCore, an endoscopy foundation model for whole-case patient-level gastric cancer analysis, and assess whether routinely stored endoscopic image sets can support cancer detection, invasion-depth prediction, biomarker inference, and prognosis.
Design
GutCore was pretrained on 5.6 million de-identified endoscopic images from more than ten hospitals. We compared GutCore with general, medical, and endoscopy-specific foundation models using public image-level datasets and an internal retrospective cohort of 11,035 de-identified endoscopic examinations from Samsung Medical Center (2019–2023). Whole-case image sets were aggregated for patient-level prediction of cancer status, invasion depth, biomarker status, and overall survival.
Results
GutCore achieved AUCs of 0.996 for cancer detection, 0.960 for muscularis propria invasion, 0.801 for SM2-or-deeper invasion, and 0.781 for mucosal versus submucosal invasion among ESD-treated early gastric cancer cases. Biomarker prediction was strongest for EBV status and MLH1 loss and weaker for HER2 status, with AUCs of 0.861, 0.822, and 0.648, respectively. In the held-out advanced gastric cancer test set, GutCore-derived risk groups separated overall survival (log-rank P < .0001; high-risk vs low-risk hazard ratio, 13.18; 95% CI, 6.06–28.66), including within pathological stage II and III disease. Public image-level benchmarks further supported the generality of the learned endoscopic representation.
Conclusions
GutCore enabled whole-case patient-level gastric cancer assessment from routinely stored endoscopic images, extending endoscopy foundation model evaluation beyond selected frames. Independent external validation is required before clinical use.
