Related Experiment Video
Updated: Sep 20, 2026
Intraoperative Gastroscopy for Tumor Localization in Laparoscopic Surgery for Gastric Adenocarcinoma
Published on: August 9, 2016
Endoscopic Artificial Intelligence for Invasion-Depth Prediction in Superficial Upper Gastrointestinal Cancer: A
Hyuk Lee1, Yang Won Min1, Sehun Kim2
1Department of Medicine, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea; Samsung Precision Genome Medicine Institute, Samsung Medical Center, Seoul, Republic of Korea.
Background And Aims:
Pre-resection prediction of submucosal invasion is pivotal for endoscopic submucosal dissection (ESD) candidate selection in early gastric and esophageal cancers. We performed a diagnostic test accuracy meta-analysis of artificial intelligence (AI) for invasion-depth prediction.
Methods:
PubMed/MEDLINE, Embase, Cochrane CENTRAL, and Web of Science were searched through 11 June 2026. Eligible studies applied AI to endoscopic images or videos of superficial upper gastrointestinal cancers with histopathological reference standards. Pooled sensitivity and specificity were estimated with a bivariate random-effects model; subgroup and meta-regression analyses examined validation framework, histology, AI architecture, and imaging modality. Risk of bias was assessed with QUADAS-2.
Results:
Twenty-six studies were eligible; 18 (14 gastric adenocarcinoma, 4 esophageal squamous cell carcinoma) contributed to the primary pool of 6,568 validation/test patients or lesions. Pooled sensitivity was 0.77 (95% CI 0.71-0.82) and specificity 0.88 (0.85-0.90); positive and negative likelihood ratios were 6.30 and 0.26. Externally validated specificity was lower than internally validated specificity (0.85 vs. 0.93); validation framework was the only covariate significantly associated with performance in prespecified meta-regression (specificity pseudo-R2 77.9%), whereas AI architecture and imaging modality were not. In within-study paired comparisons (n = 8 studies), mean per-study sensitivity was higher for AI alone than for unassisted endoscopists (82.5% vs. 63.8%; +18.7 percentage points; paired t = 3.29, p = 0.013); specificity did not differ significantly. Limited AI-assisted reader-comparison data were directionally consistent with a second-reader role but remained exploratory.
Conclusions:
Endoscopic AI shows clinically meaningful, but validation-dependent, accuracy for invasion-depth prediction, supporting stronger rule-in than rule-out utility for submucosal invasion. Its most plausible near-term role is second-reader support rather than stand-alone determination of ESD candidacy; independent patient-level external validation should be considered essential before clinical deployment. (PROSPERO registration: CRD420261400366).
