Machine learning for differentiating between pancreatobiliary-type and intestinal-type periampullary carcinomas based on CT imaging and clinical findings.

Purpose: To develop a diagnostic model for distinguishing pancreatobiliary-type and intestinal-type periampullary adenocarcinomas using preoperative contrast-enhanced computed tomography (CT) findings combined with clinical characteristics.

Methods: This retrospective study included 140 patients with periampullary adenocarcinoma who underwent preoperative enhanced CT, including pancreaticobiliary (N = 100) and intestinal (N = 40) types. They were randomly assigned to the training or internal validation set in an 8:2 ratio. Additionally, an independent external cohort of 28 patients was enrolled. Various CT features of the periampullary region were evaluated and data from clinical and laboratory tests were collected. Five machine learning classifiers were developed to identify the histologic type of periampullary adenocarcinoma, including logistic regression, random forest, multi-layer perceptron, light gradient boosting, and eXtreme gradient boosting (XGBoost).

Results: All machine learning classifiers except multi-layer perceptron used achieved good performance in distinguishing pancreatobiliary-type and intestinal-type adenocarcinomas, with the area under the curve (AUC) ranging from 0.75 to 0.98. The AUC values of the XGBoost classifier in the training set, internal validation set and external validation set are 0.98, 0.89 and 0.84 respectively. The enhancement degree of tumor, the growth pattern of tumor, and carbohydrate antigen 19-9 were the most important factors in the model.

Conclusion: Machine learning models combining CT with clinical features can serve as a noninvasive tool to differentiate the histological subtypes of periampullary adenocarcinoma, in particular using the XGBoost classifier.