Identifying a better-prognosis pancreatic cancer from its benign inflammatory mimic: a machine learning approach with contrast-enhanced ultrasound for early intervention.
Authors
Affiliations (4)
Affiliations (4)
- Department of Ultrasound, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
- Department of Biomedical Engineering, Beijing University of Technology, Beijing, China.
- Department of General Surgery, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
- Department of Ultrasound, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China. [email protected].
Abstract
Hypervascular pancreatic ductal adenocarcinoma (PDAC) and mass-forming pancreatitis (MFP) represent a classic diagnostic mimicry on contrast-enhanced ultrasound, as both exhibit similar arterial-phase hyperenhancement, precluding reliable visual distinction. Crucially, the hypervascular PDAC subtype is associated with a more favorable prognosis, rendering its accurate identification from its benign inflammatory mimic (MFP) a clinically significant priority for early and appropriate intervention. This study aimed to develop and validate a small-sample-oriented machine learning framework leveraging quantitative time-intensity curve (TIC) features to achieve this precise differentiation. We retrospectively included 152 patients with pathologically confirmed lesions who underwent CEUS between September 2017 and April 2024 (85 hypervascular PDAC, 67 MFP). A temporally separated split was used: 122 patients (2017-2022) formed the training cohort and 30 patients (2023-2024) served as the internal test cohort. Paired TICs were generated from the lesion and adjacent normal pancreatic parenchyma, and 22 quantitative difference/ratio features describing enhancement amplitude, temporal kinetics and curve morphology were extracted. Based on independent-samples t-tests and clinical interpretability, five representative TIC features were selected to train six classical classifiers (logistic regression, support vector machine, k-nearest neighbors, random forest, naïve Bayes, and decision tree). Model performance was assessed by stratified 10-fold cross-validation on the training cohort and by testing on the temporally separated cohort. In nested cross-validation, all models achieved AUCs of approximately 0.89-0.92. On the test cohort, AUCs ranged from about 0.83 to 0.92, with logistic regression performing best (AUC 0.915). Overall, the six classifiers showed broadly comparable discrimination. A machine-learning model built on a small set of physiologically interpretable CEUS-TIC features can provide stable and explainable quantitative support for differentiating hypervascular PDAC from MFP, even under limited-sample conditions.