Beyond Detection: <scp>AI</scp> in Gastric Endoscopy

I read with great interest the study by Abe et al., evaluating an artificial intelligence (AI) system using image-enhanced endoscopy over white light endoscopy for increased detection of neoplastic lesions [1]. I would like to extend the discussion by highlighting additional considerations for upper gastrointestinal endoscopy. Linked color imaging (LCI) offers distinct advantages over narrow-band imaging (NBI) and blue laser imaging (BLI), particularly in enhancing subtle mucosal color contrasts. This is especially valuable in Helicobacter pylori–negative cases, where the absence of diffuse redness may obscure early gastric cancer (EGC) [2]. AI-driven quantification of these subtle shifts could standardize detection across varying levels of expertise. Similarly, automated recognition of the regular arrangement of collecting venules (RAC) may serve as a surrogate biomarker of H. pylori–negative mucosa, reducing unnecessary biopsies and improving specificity [3]. Another important frontier is the differentiation of histologic subtypes. Differentiated EGC often demonstrates distinct microvascular and microsurface patterns, whereas undifferentiated EGC may mimic benign erosions [4]. AI-based pattern recognition could provide real-time histologic prediction, guiding therapeutic decisions between endoscopic resection and surgical referral. Equally critical is distinguishing malignant ulcers from benign conditions such as peptic ulcer disease. Training AI on large, annotated datasets encompassing these entities could reduce misclassification and enhance diagnostic confidence. As AI systems expand their discriminative capacity, the phenomenon of “discriminative hallucinations” must be acknowledged. These occur when algorithms confidently assign a neoplastic label to benign or artifact-driven patterns, often due to overfitting or underrepresentation of benign lesions in training datasets. In practice, such errors may lead to unnecessary biopsies, overtreatment, or patient anxiety. Balanced training in this context requires proportional inclusion of benign, inflammatory, and neoplastic lesions, stratified by histologic subtype and clinical context. Recent reviews on AI hallucinations in medical imaging highlight the importance of probabilistic outputs that communicate uncertainty to clinicians [5]. A.T.: concept and design, and critical revision of the manuscript for important intellectual content. The author has nothing to report. No use of AI for text or concept generation. Zotero Software for citations and Grammarly software for spelling checks and grammar used in manuscript preparation. The author declares no conflicts of interest. This article is linked to Abe et al. papers. To view this article, visit https://doi.org/10.1111/den.70015.