Artificial Intelligence in Endoscopic Ultrasonography for Gastrointestinal Subepithelial Lesions: Evidence, Validation, and Clinical Translation

Artificial intelligence is increasingly being applied to endoscopic ultrasonography for gastrointestinal subepithelial lesions, but the maturity of the evidence differs substantially across clinical tasks. This narrative review critically appraises contemporary studies of lesion detection and localization, wall-layer identification, diagnostic classification, pathological risk classification of gastrointestinal stromal tumors, human–artificial intelligence interaction, and potential management support. Thirty-one original studies published from 2021 through 2026 were included in the contemporary evidence map. Binary differentiation of gastrointestinal stromal tumors from leiomyomas or other selected comparators has the largest evidence base and includes multicenter, external, prospective, and real-time evaluations. Evidence for multiclass diagnosis, layer-of-origin assessment, small lesions, and pathological risk classification is less mature and is strongly influenced by case spectrum, reference standard, equipment, image selection, and validation design. No study has demonstrated that artificial intelligence improves tissue-acquisition strategy, surveillance, treatment selection, or patient outcomes in subepithelial lesions. Progress toward clinically useful decision support will require patient-level data separation, device- and center-held-out validation, analysis of complete examinations rather than curated images, explicit uncertainty handling, prospective human–artificial intelligence studies, and trials with endoscopy-specific management and safety outcomes.