Determining malignant potential of pancreatic intraductal papillary mucinous neoplasm with CT: interobserver variability and diagnostic performance.

Abstract

Objectives: To evaluate interobserver agreement and diagnostic performance of computed tomography (CT) in assessing the malignant potential of pancreatic intraductal papillary mucinous neoplasms (IPMNs) based on the 2024 International evidence-based Kyoto guidelines, focusing on the impact of reviewer experience.

Methods: In this multicenter retrospective study, CT images of 120 patients with pathologically confirmed IPMNs (53 malignant, 44.2%) were independently reviewed by eight observers (four experienced abdominal radiologists and four radiology residents). Malignant potential was graded using a five-point scale based on high-risk stigmata (HRS) and worrisome features (WF) defined by the 2024 Kyoto guidelines. Diagnostic performance (AUC) and interobserver agreement were compared between experience-based and guideline-based interpretations according to reviewer experience.

Results: Among 120 patients (mean age, 65.9 ± 10.5 years; range, 41-87), 53 (44.2%) had malignant IPMNs. Interobserver agreement was substantial (α = 0.66) using the Kyoto guidelines and was higher among more experienced reviewers (α = 0.73 vs. 0.66; p < 0.001). Diagnostic performance was good (AUC, 0.75-0.89), with no statistically significant difference between more and less experienced reviewer groups (p = 0.26). For less experienced reviewers, guideline-based interpretation improved diagnostic performance (AUC 0.71-0.80 to 0.75-0.85; p < 0.001) and interobserver agreement (α = 0.43 to 0.66; p < 0.001) compared with experience-based assessment. In contrast, for more experienced reviewers, guideline-based interpretation did not result in a statistically significant change in diagnostic performance (AUC 0.88-0.92 to 0.82-0.89; p = 0.08), while interobserver agreement showed a modest decrease (α = 0.81 to 0.73; p < 0.001).

Conclusion: The 2024 Kyoto guidelines demonstrated an experience-dependent effect, improving diagnostic performance and interobserver agreement among less experienced readers, with limited additional benefit for experienced radiologists.