Background: Esophageal cancer is a highly lethal malignancy, the prognosis of which depends largely on early diagnosis. Artificial intelligence (AI) has emerged as a promising tool to enhance endoscopic detection and characterization of early esophageal cancer. This scoping review aims to map and synthesize the available evidence regarding the diagnostic performance and clinical utility of artificial intelligence systems applied to upper gastrointestinal endoscopy for the detection and characterization of premalignant squamous lesions and early-stage esophageal squamous cell carcinoma (ESCC). Methods: A scoping review was conducted according to Arksey and O’Malley, Levac, Joanna Briggs Institute, and PRISMA-ScR recommendations. The review question focused on patients with premalignant lesions or early ESCC, artificial intelligence-based diagnostic systems, and upper gastrointestinal endoscopy. Searches were performed in PubMed, Scopus, and Embase. Original studies reporting sensitivity, specificity, accuracy, AUC, or F1-score were included. Results: A total of 30 publications were included, consisting mainly of retrospective observational and diagnostic test studies (26/30; 86.7%), followed by randomized clinical trials (3/30; 10.0%) and a multicenter validation study (1/30; 3.3%). The studies were predominantly from China (18/30; 60%), followed by Japan (8/30; 26.7%), Taiwan (3/30; 10%), and the United Kingdom + Taiwan (1/30; 3%). Automatic lesion detection was predominant (21/30; 70.0%), followed by diagnostic classification (11/30; 36.7%), while segmentation (3/30; 10.0%), histological prediction (2/30; 6.7%), estimation of invasion depth (3/30; 10.0%), and lesion delineation (1/30; 3.3%) were evaluated less frequently, and in some cases combined within the same model. The most used endoscopic imaging modalities were narrow-band imaging (23/30; 76.7%) and white light endoscopy (20/30; 66.7%), followed by magnifying endoscopy with narrow-band imaging (5/30; 16.7%), blue light imaging (2/30; 6.7%), and hyperspectral imaging (1/30; 3.3%). Conclusions: Available studies suggest that AI has the potential to achieve high diagnostic performance under controlled conditions. However, the current evidence is derived predominantly from single-center retrospective studies using selected high-quality static images, with limited external, prospective, and real-world validation.
Rodríguez et al. (Sun,) studied this question.