سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / اولین همایش بین المللی تحول خدمات سلامت: تلفیق مراقبت، درمان، بازتوانی و امیدبخشی

AI-Assisted Risk Stratification of Pancreatic Cystic Lesions: Integrating Imaging, EUS, and Clinical Decision-Making

نویسندگان :

Mobina Katal¹

1- دانشگاه آزاد واحد ساری

کلمات کلیدی :

Pancreatic Cyst،Endosonography،Artificial Intelligence،Risk Assessment،Pancreatic Neoplasms

چکیده :

Background Pancreatic cystic lesions (PCLs) are increasingly detected on cross-sectional imaging. A clinically important subset, particularly intraductal papillary mucinous neoplasms and mucinous cystic neoplasms, has malignant potential. Conventional risk stratification combines CT or MRI morphology, clinical indicators, guideline criteria, and, when indicated, endoscopic ultrasound (EUS). Because these approaches have imperfect sensitivity and specificity, both unnecessary surgery and delayed treatment may occur. This review assessed whether artificial intelligence (AI), particularly models using EUS and related high-resolution imaging, provides incremental diagnostic value for malignant risk stratification. Method A narrative review was conducted of English-language literature published from January 2020 to December 2025. PubMed, ScienceDirect, and SpringerLink were searched using terms related to pancreatic cystic lesions, IPMN, artificial intelligence, deep learning, convolutional neural networks, EUS, and risk stratification. Studies addressing AI or machine-learning applications to CT, MRI, EUS, or EUS-guided needle-based confocal laser endomicroscopy were considered. Data on model architecture, imaging modality, reference standard, diagnostic performance, and comparison with guideline-based criteria were extracted and synthesised narratively. No formal meta-analysis was undertaken because of heterogeneity in endpoints and validation methods. Finding The reviewed evidence showed three main applications: cyst-subtype differentiation, discrimination of high-grade dysplasia or carcinoma from low-grade lesions, and refinement of surgical candidacy among guideline-defined high-risk lesions. Representative CNN models using EUS or nCLE demonstrated sensitivities of approximately 83-96% and accuracies of approximately 83-95%. A multicentre EUS CNN study reported sensitivity of 95.7%, specificity of 88.7%, and AUROC of 0.951 for distinguishing high-grade dysplasia or carcinoma from low-grade IPMNs. Machine-learning models combining clinical and CT features also achieved AUCs around 0.93-0.95 in selected high-risk populations. Several studies suggested improved specificity or accuracy compared with conventional guideline criteria, although external validation was limited. Conclusion Current evidence indicates that EUS-based AI can provide incremental diagnostic information beyond conventional imaging and clinical indicators in pancreatic cystic lesion risk stratification. Nevertheless, most evidence remains retrospective, with heterogeneity in image acquisition, reference standards, and validation. Prospective multicentre studies are needed before routine implementation. In the near term, AI outputs may be most appropriately integrated as decision-support information within multidisciplinary surveillance-versus-resection pathways rather than used as autonomous clinical decisions. Keywords Pancreatic Cyst, Endosonography, Artificial Intelligence, Risk Assessment, Pancreatic Neoplasms