Benefit of Linked-Color Imaging in Artificial Intelligence-Assisted Diagnosis of Early Gastric Cancer: A Pilot Study With Propensity Score Adjustment.
Linked-color imaging reduced false-positive AI detections in endoscopy for early gastric cancer compared to white-light imaging, but results should be interpreted with caution due to study design limitations.
Where it sits
this study against the rest of the glutathione (gsh) corpusSummary and findings
This study evaluated the effectiveness of linked-color imaging (LCI) in reducing false-positive detections in AI-assisted endoscopy for early gastric cancer. A total of 63 patients were analyzed, with LCI showing a reduction in false positives compared to white-light imaging (WLI). The median false positives were 2 for LCI versus 5 for WLI, with a p-value of <0.001.
Abstract
<h4>Background and aims</h4>Artificial intelligence (AI)-assisted endoscopy represents a promising approach for lesion detection, yet frequent false-positive detections impair clinical utility by disrupting examinations and diminishing physician confidence. Linked-color imaging (LCI), an image-enhanced endoscopy technique that amplifies mucosal and vascular contrast, may address this limitation. This investigation evaluated whether LCI reduces false-positive AI detections compared with white-light imaging (WLI).<h4>Methods</h4>This retrospective study analyzed consecutive AI-assisted upper endoscopies performed between March 2024 and June 2025. WLI and LCI were performed sequentially within the same endoscopic session in each patient. False-positive AI detections were compared between modalities using two computer-aided detection (CAD) versions. Propensity score adjustment was used as a sensitivity analysis for baseline differences between CAD Versions I and II.<h4>Results</h4>Of 66 initially screened cases, 63 remained after excluding patients with prior gastric surgery. LCI reduced false-positive AI detections compared with WLI (median 2 vs. 5; <i>p</i> < 0.001). In CAD version-stratified sensitivity analyses, LCI reduced false-positive AI detections in both Version I (5 to 2; <i>p</i> = 0.01) and Version II (2 to 0; <i>p</i> = 0.03). This reduction remained consistent across atrophic grades. Both imaging modalities identified all gastric lesions, achieving 100% detection sensitivity.<h4>Conclusions</h4>LCI assessment performed after WLI observation yielded fewer false-positive CAD-EYE detections while maintaining lesion detection sensitivity. However, because the observation sequence was fixed, these findings should be interpreted cautiously and require confirmation in prospective or counterbalanced studies.<b>Trial Registration</b>: N/A (retrospective study).
Background
This paper addresses the challenge of false-positive detections in AI-assisted endoscopy for early gastric cancer, which can disrupt examinations and affect physician confidence. Prior studies have indicated that image-enhanced techniques may improve detection accuracy. This study is significant as it explores the potential of linked-color imaging (LCI) to enhance the performance of AI in this context.
Methods
This retrospective study analyzed 63 consecutive AI-assisted upper endoscopies performed between March 2024 and June 2025, excluding patients with prior gastric surgery. Both white-light imaging (WLI) and linked-color imaging (LCI) were performed sequentially in each patient. False-positive detections were compared between the two imaging modalities using two computer-aided detection (CAD) versions, with propensity score adjustment for baseline differences.
Results
LCI significantly reduced false-positive AI detections compared to WLI, with a median of 2 false positives for LCI versus 5 for WLI (p<0.001). In sensitivity analyses stratified by CAD version, LCI reduced false positives from 5 to 2 in Version I (p=0.01) and from 2 to 0 in Version II (p=0.03). Both modalities maintained 100% sensitivity in detecting gastric lesions.
Interpretation
The findings suggest that LCI may improve the accuracy of AI-assisted endoscopy by reducing false-positive detections, which is a notable advancement. However, the effect sizes, while statistically significant, may not be clinically meaningful given the small sample size and the nature of the study design. The fixed sequence of imaging modalities may introduce confounding factors that limit the conclusions drawn from this study.
Key findings
- Median false positives of 2 for LCI vs 5 for WLI, p<0.001.
- In CAD Version I, false positives reduced from 5 to 2, p=0.01.
- In CAD Version II, false positives reduced from 2 to 0, p=0.03.
- Both imaging modalities achieved 100% detection sensitivity for gastric lesions.
Limitations
- Retrospective study design.
- Small sample size, n=63.
- Fixed observation sequence may introduce bias.
- Further confirmation needed in prospective studies.