Evaluating the role of the main pancreatic duct in intraductal papillary mucinous neoplasm grading: A multi-structure radiomics-based machine learning approach - 20/02/26

Doi : 10.1016/j.diii.2026.02.002

Emir A. Syailendra ^a,^⁎ , Felipe Lopez-Ramirez ^a, Florent Tixier ^a, Alejandra Blanco ^a, Hajra Arshad ^a, Mohammad Yasrab ^a, Zahra F. Rahmatullah ^a, Ammar A. Javed ^b, Jin He ^c, Anne Marie Lennon ^d, Ralph H. Hruban ^e, Elham Afghani ^f, Satomi Kawamoto ^a, Linda C. Chu ^a, Elliot K. Fishman ^a

^a Russel H. Morgan Department of Radiology and Radiological Science, School of Medicine, Johns Hopkins University, Baltimore, MD 21287, USA

^b Department of Surgery, New York University Grossman School of Medicine, New York, NY 10016, USA

^c Department of Surgery, School of Medicine, Johns Hopkins University, Baltimore, MD 21287, USA

^d Department of Medicine, University of Pittsburgh Medical Center, Pittsburgh, PA 15213, USA

^e Sol Goldman Pancreatic Cancer Research Center, Department of Pathology, School of Medicine, Johns Hopkins University Baltimore, MD 21287, USA

^f Division of Gastroenterology, Department of Medicine, Johns Hopkins University, Baltimore, MD 21287, USA

^⁎ Corresponding author.

Sous presse. Épreuves corrigées par l'auteur. Disponible en ligne depuis le Friday 20 February 2026

Highlight

•	Integrating main pancreatic duct radiomics enhances the performance of machine learning models for intraductal papillary mucinous neoplasm grade prediction over lesion-only radiomics models.
•	Feature importance analysis identifies main pancreatic duct radiomics features as the largest contributors to model’s output.
•	Radiomics-based models outperform models based on conventional morphological size measurements for predicting intraductal papillary mucinous neoplasm grade.

Le texte complet de cet article est disponible en PDF.

Abstract

Purpose

The purpose of this study was to evaluate the contribution of radiomics features extracted from various pancreatic structures on computed tomography (CT) images, including the main pancreatic duct and cystic lesion, for predicting the pathological grade of intraductal papillary mucinous neoplasms (IPMNs) using machine learning models.

Materials and methods

A retrospective study using preoperative CT images obtained during the venous phase of enhancement in patients with pathologically confirmed IPMNs (2003–2024) was conducted. Main pancreatic ducts and cysts were manually segmented. Machine learning models were trained to classify IPMNs into high-grade/associated invasive carcinoma (HG/I) IPMNs or low-grade (LG) IPMNs using radiomics features from three structures ( i.e ., cysts, main pancreatic duct, and a combination of both structures). Model performance was evaluated using area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) values were used to interpret feature importance.

Results

A total of 274 patients with IPMNs were included. There were 149 patients with HG/I IPMNs (70 women [47 %]; median age, 71.0 years; age range: 29–92) and 125 patients with LG IPMNs (73 women [58.4 %]; median age, 68.0 years; range: 35–86). HG/I IPMNs were predominantly mixed-type IPMNs (51.7 %; 77/149). LG IPMNs were unspecified (51/125; 40.8 %), main/mixed (40/125; 32 %), or branch-duct type (34/125; 27.2 %). A support vector machine trained on combined features achieved the largest AUC (0.85; 95 % confidence interval [CI]: 0.85–0.87; P < 0.001), with 90 % sensitivity (95 % CI: 90–93), and 60 % specificity (95 % CI: 58–62). SHAP analysis identified main pancreatic duct radiomics features as having the largest contribution to model output.

Conclusion

Integrating CT-based radiomics features from pancreatic ducts and cysts improves classification performance, with main pancreatic duct features being the most contributive predictor of IPMN grade.

Le texte complet de cet article est disponible en PDF.

Keywords : Artificial intelligence, Intraductal papillary mucinous neoplasm, Machine learning, Main pancreatic duct, Radiomics, Risk stratification

Abbreviations : AUC, BD, CI, CT, FN, FP, HG/I, IPMN, LG, LASSO, LoG, MPD, NPV, PDAC, PPV, ROC, SD, SHAP, SVM, TN, TP, t-SNE