Artificial Intelligence-Assisted EUS for Risk Stratification of Pancreatic Cystic Lesions: A Narrative Review of Current Evidence and Future Directions for Predicting High-Grade Dysplasia, Invasive Cancer, and the Need for Surgical Referral

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background: Pancreatic cystic lesions are increasingly detected on cross-sectional imaging, yet accurate risk stratification remains challenging. A clinically important minority, particularly intraductal papillary mucinous neoplasms and mucinous cystic neoplasms, may progress to high-grade dysplasia or invasive carcinoma. EUS plays a central role in cyst evaluation, but interpretation remains operator-dependent, and guideline-based risk categories have imperfect predictive accuracy. Methods: This narrative review was conducted using a structured literature search of PubMed, Scopus, Web of Science, and Embase with terms including artificial intelligence, deep learning, machine learning, endoscopic ultrasound, pancreatic cystic lesions, intraductal papillary mucinous neoplasm, high-grade dysplasia, and pancreatic cancer. Studies reporting AI applications in EUS image interpretation, multimodal risk prediction, radiomics, and cyst-fluid or molecular marker integration were included. A qualitative synthesis was performed given the heterogeneity of study designs and AI methodologies. Results: AI may enhance EUS-based assessment by extracting quantitative imaging features and integrating EUS data with CT, MRI, cyst-fluid biomarkers, cytology, molecular markers, and longitudinal cyst behavior. The most clinically meaningful goal — though not yet established — is individualized prediction of high-grade dysplasia and invasive cancer to support surgical referral decisions, particularly where guideline-based criteria are discordant or borderline. Current models remain preliminary and require prospective validation. Retrospective designs, small datasets, lack of external validation, and uncertainty regarding explainability and clinical integration limit evidence. Conclusions: Future prospective multicentre studies should determine whether AI-assisted EUS improves patient-centered outcomes by reducing unnecessary surgery while preventing delayed diagnosis of advanced neoplasia.