Radiomics and artificial intelligence in pancreatic cyst characterization: future or fiction?
Authors
Affiliations (7)
Affiliations (7)
- Departement of Diagnostic Radiology, Fondazione IRCCS San Gerardo dei Tintori, Monza, Italy. [email protected].
- Departement of Diagnostic Radiology, Fondazione IRCCS San Gerardo dei Tintori, Monza, Italy. [email protected].
- Department of Diagnostics and Public Health - Institute of Radiology, University of Verona, Verona, Italy.
- Interventional Radiology Unit, Ospedale Generale Regionale Francesco Miulli, Acquaviva delle Fonti, Italy.
- Università Italiana in Puglia Bari Casamassima, Bari, Italy.
- Departement of Diagnostic Radiology, Fondazione IRCCS San Gerardo dei Tintori, Monza, Italy.
- School of Medicine, University of Milano-Bicocca, Milan, Italy.
Abstract
Pancreatic cystic lesions (PCLs) are increasingly detected due to the widespread use of cross-sectional imaging and represent a significant diagnostic challenge because of their heterogeneous biological behavior, ranging from benign lesions to neoplasms with malignant potential. Accurate characterization and risk stratification are essential to guide appropriate management and avoid unnecessary surgical interventions. Conventional imaging modalities, including computed tomography (CT), magnetic resonance (MR) imaging, and endoscopic ultrasound (EUS), remain central to the diagnostic work-up; however, their ability to reliably differentiate cyst subtypes and predict malignant transformation remains limited. In recent years, artificial intelligence (AI) and radiomics have emerged as promising approaches for improving the non-invasive characterization of PCLs by extracting quantitative imaging features beyond those appreciable through visual assessment. This narrative review summarizes the current evidence regarding CT- and MR-based radiomics and AI in pancreatic cyst characterization, focusing on their role in differentiating mucinous from non-mucinous cysts, identifying high-risk intraductal papillary mucinous neoplasms (IPMNs), and supporting clinical decision-making. The potential advantages of these techniques are discussed alongside main methodological limitations, including variability in imaging acquisition protocols, segmentation reproducibility, small and often retrospective datasets, limited external validation, and interpretability of AI-based models. Further multicenter studies, standardized radiomic pipelines, and prospective validation are required before these tools can be reliably integrated into routine clinical practice.