Intraductal papillary mucinous neoplasms (IPMNs) are increasingly diagnosed due to rising life expectancy and widespread use of abdominal imaging. Despite the publication of several guidelines to drive the management of IPMNs, a significant proportion of patients still undergo surgical resection for lesions ultimately classified as low-grade dysplasia (LGD), while others managed conservatively may progress to invasive carcinoma (IC). Pancreatic surgery could have high perioperative morbidity and mortality, highlighting the need for improved preoperative stratification. Among the available tools, endoscopic ultrasound with fine-needle aspiration (EUS-FNA) represents a key diagnostic method for assessing the risk of malignancy in IPMN. This study aimed to evaluate the diagnostic accuracy of EUS-FNA in predicting histological dysplasia grade in surgically resected IPMNs. 

We retrospectively included patients who underwent both EUS-FNA and surgery at San Raffaele Hospital. Surgical diagnoses were classified into LGD, high-grade dysplasia (HGD), or IC, while cytological diagnoses also included “Negative for malignant cells” (NMC). We defined STRICT criteria (SC) as exact cytological-histological concordance (considering NMC as non-concordant), while LARGE (LC) considered NMC equal to LGD. Concordance was assessed using Fleiss’ kappa (κ) and linear weighted κ. Uni- and multivariate logistic regressions were conducted to identify predictors of concordance. Diagnostic performance of EUS cytology in distinguishing LGD and HGD/IC was evaluated through sensitivity, specificity, PPV, NPV, and accuracy. A Random Forest classifier was developed to predict concordance before FNA utilizing three variables: patient age, main pancreatic duct (MPD) dilation, and IPMN subtype. The model was trained on the entire dataset with 100 decision trees and a fixed random state to ensure reproducibility.

Ninety-four patients were included (36% branch duct-, 41% main duct-, 23% mixed type-IPMN). The concordance using SC and LC criteria was moderate (Fleiss’ κ respectively 0.41 and 0.36). The weighted κ was 0.55. When stratified, concordance was κ=0.52 for LGD 0.437 for IC, and 0.109 for HGD. Univariate analysis identified diffuse MPD dilation (p=0.0016 SC, p=0.0002 LC), cyst wall enhancement (p=0.0955 SC), grade of dysplasia (p=0.0809 SC, <0.00001 LC), and IPMN subtype (p=0.0433 LC) as significant. Multivariate analysis confirmed cyst wall enhancement (OR=0.128, p=0.0334) and grade of dysplasia (OR=0.239, p=0.0094) as independent predictors using SC, while the latter one alone (OR=0.354, p=0.0027) when LC were considered. Diagnostic performance of EUS cytology in distinguishing LGD vs HGD/IC showed sensitivity 75%, specificity 89.2%, PPV 93.1%, NPV 64.7%, accuracy 79.8%. The Random Forest model demonstrated high predictive performance, achieving an accuracy of 94.7% and an area under the receiver operating characteristic curve (AUC) of 0.99. It correctly classified 89 out of 94 cases, with 4 false negatives and 1 false positive. Precision and recall were 97.9% and 92.3% for the concordant class.

EUS-FNA shows moderate concordance with final pathology, especially for LGD and IC, while HGD remains challenging. Pre-FNA imaging features are predictive of concordance. A machine learning model demonstrated high accuracy using preprocedural variables only, supporting its clinical value in biopsy guidance. Strengths include the high-volume pancreatic center setting. Limitations include selection bias from a surgical cohort, which is inherent to histology-based studies.