Evaluation of Computer-aided Detection for Identifying Missed Gastric Cancer After Endoscopic Submucosal Dissection.

Key Takeaways for GI Nurses

• Computer-aided detection (CADe) demonstrated comparably low per-lesion sensitivity to experienced endoscopists (15.0% vs. 13.8%) for identifying missed gastric cancers (MGCs) after ESD, indicating that AI is not yet a reliable stand-alone safety net for post-ESD surveillance.
• CADe processed images dramatically faster than human reviewers (0.03 seconds vs. 2.86 seconds per image), highlighting its potential value as a rapid, supplementary screening tool during high-volume image review rather than a replacement for careful endoscopist assessment.
• Endoscopists maintained significantly higher specificity (94.8% vs. 88.5%) and positive predictive value, meaning CADe generated more false positives—an important consideration for nurses managing patient anxiety, follow-up testing, and workflow when AI-flagged findings require confirmation.
• This study underscores that current CADe systems require further refinement before they can be trusted to independently reduce missed gastric cancer rates in post-ESD surveillance settings.

Clinical Relevance

For GI nurses involved in endoscopy suite operations, this study offers important context about the current limitations of AI-assisted detection tools in a high-stakes clinical scenario: identifying cancers missed after endoscopic submucosal dissection. While CADe systems are often marketed as enhancing diagnostic accuracy, this research demonstrates that detection performance for MGCs remains suboptimal whether performed by AI or experienced endoscopists, with both groups missing the majority of lesions in this dataset. Nurses coordinating surveillance EGD scheduling and patient education should understand that even with technological support, vigilant endoscopic technique, adequate mucosal visualization, and standardized surveillance protocols remain essential safeguards against missed malignancy.

From an operational standpoint, the significantly faster per-image processing time for CADe (0.03s vs. 2.86s) suggests a potential future role for AI as a triage or double-check mechanism during image review, particularly in high-volume centers or when retrospectively auditing surveillance examinations. Nurses may increasingly be tasked with managing AI-integrated workflows, including flagging discrepancies between CADe alerts and endoscopist findings, documenting AI-assisted diagnostic sessions, and supporting quality assurance initiatives that incorporate CADe as one data point among several. However, the higher false-positive rate associated with CADe (reflected in lower specificity and PPV) means nursing staff should anticipate additional patient counseling needs when AI-flagged areas require biopsy or further evaluation despite ultimately benign findings.

This study also has implications for professional development and interdisciplinary collaboration. As CADe technology evolves, GI nurses will play a critical role in bridging AI outputs with clinical decision-making, ensuring that technology serves as an adjunct rather than a substitute for endoscopist expertise. Nurses should stay informed about the evidence base supporting AI tools in their units, participate in training on CADe system limitations, and advocate for balanced integration that preserves thorough visual inspection protocols during both index ESD procedures and surveillance examinations.

Bottom Line

This study shows that current CADe technology, while offering speed advantages, does not yet outperform experienced endoscopists in detecting missed gastric cancers after ESD and carries a higher false-positive rate—reminding GI nurses that AI should be viewed as a developing adjunct tool rather than a substitute for rigorous surveillance protocols and skilled endoscopic assessment in post-ESD care.