Masterclass: Advanced GI Endoscopic Imaging (On-Demand) | December 2025-Bergman - Computer Aided Detection and Diagnosis in Barrett’s Esophagus

Bergman - Computer Aided Detection and Diagnosis in Barrett’s Esophagus

Pdf Summary

This presentation by Jacques Bergman from Amsterdam University Medical Centers reviews the role of machine learning, particularly AI-based computer-aided detection (CADe) and diagnosis (CADx), in improving endoscopic detection and characterization of early neoplasia in Barrett’s esophagus. While AI has successfully enhanced detection in colonoscopy due to clear lesion distinctions and abundant data, Barrett’s early cancers pose greater challenges because of subtle visual differences, fewer images, and lower prevalence.<br /><br />Recent benchmarking studies involving over a thousand images and experienced versus general endoscopists demonstrate that CADe systems outperform general endoscopists and perform on par with experts in detecting Barrett’s neoplasia in ex vivo video analyses. However, clinical practice involves heterogeneous imaging conditions—operator-dependent factors like image quality, luminal expansion, cleaning, and illumination—which significantly impact AI performance. Small image perturbations, changes in enhancement settings, and compression affect outcome robustness.<br /><br />To build more reliable AI for Barrett’s, strategies include domain-specific pretraining, self-critical AI algorithms, robust architectures, and heterogeneous training data representing varied imaging conditions. Despite improvements, challenges remain: “unseen” variations evade detection, and AI is highly sensitive to image quality variations. Incorporating AI as a complementary tool rather than a replacement, with optimized human-AI interaction and interface design, is emphasized to mitigate false positives and build user trust.<br /><br />For regulatory approval, Bergman argues the FDA should prioritize high-quality ex vivo testing demonstrating that CADe improves general endoscopist performance in neoplasia recognition. Further in vivo trials should confirm that AI aids workflow without adding excessive biopsies or procedure time. Ultimately, AI holds promise to enhance Barrett’s surveillance, but success depends on improving algorithm robustness, integrating quality control, optimizing endoscopist cooperation, and streamlining regulatory pathways. Enhancing overall endoscopic image quality may also have a greater impact than AI detection alone.

Keywords

Machine Learning

Artificial Intelligence

Computer-Aided Detection

Computer-Aided Diagnosis

Barrett's Esophagus

Early Neoplasia Detection

Endoscopy

Image Quality

Algorithm Robustness

Regulatory Approval