Bridging Performance, Annotation, and Human-Machine Interaction Gaps in Artificial Intelligence-Assisted Diagnosis of Barrett’s Neoplasia

Abstract

Artificial intelligence (AI)-assisted diagnosis of early neoplasia in Barrett’s esophagus now operates in real time during endoscopy, a step from the offline, patch-based image analysis that dominated earlier work. Three problems still limit its use in routine practice. The most physically detailed imaging methods, such as volumetric laser endomicroscopy, remain too slow and expensive for everyday use, whereas lightweight white-light systems that are easy to deploy generate frequent false-positive alerts. Most existing models depend on pixel-level annotation, an approach that is difficult to scale; weakly supervised attention methods have reduced this burden in histopathology but have not been tested on endoscopic video or flat, poorly defined lesions. Little is known about how interface design, operator experience, and diagnostic uncertainty jointly shape real-time decisions between clinicians and AI systems. This review reconstructs how AI-assisted Barrett’s esophagus diagnosis evolved and examines the 3 gaps in turn. It closes with 3 directions: a performance-accessibility index applicable across imaging modalities; self-supervised and weakly supervised learning adapted to endoscopic video; and validation frameworks that bring human-factors research alongside diagnostic accuracy.