Dynamic contract-enhanced CT-based radiomics for differentiation of pancreatobiliary-type and intestinal-type periampullary carcinomas.

Aim: To investigate whether computed tomography (CT) radiomics can differentiate pancreatobiliary-type from intestinal-type periampullary carcinomas.
Materials and Methods: CT radiomics of 96 patients (54 pancreatobiliary type and 42 intestinal type) with surgically confirmed periampullary carcinoma were assessed retrospectively. Volumes of interest (VOIs) were delineated manually. Radiomic features were extracted from preoperative CT images. A single-phase model and combined-phase model were constructed. Five-fold cross-validation and five machine-learning algorithms were utilised for model construction. The diagnostic performance of the models was evaluated by receiver operating characteristic (ROC) curves, and indicators included area under the curve (AUC), accuracy, sensitivity, specificity, and precision. ROC curves were compared using DeLong's test.
Results: A total of 788 features were extracted on each phase. After feature selection using least absolute shrinkage and selection operator (LASSO) algorithm, the number of selected optimal feature was 18 (plain scan), nine (arterial phase), two (venous phase), 23 (delayed phase), 15 (three enhanced phases), and 29 (all phases), respectively. For the single-phase model, the delayed-phase model using the logistic regression (LR) algorithm showed the best prediction performance with AUC, accuracy, sensitivity, specificity, and precision of 0.89, 0.83, 0.80, 0.88, and 0.93, respectively. Two combined-phase models showed better results than the single-phase models. The model of all phases using the LR algorithm showed the best prediction performance with AUC, accuracy, sensitivity, specificity, and precision of 0.96, 0.88, 0.90, 0.93, and 0.92, respectively.
Conclusion: Radiomic models based on preoperative CT images can differentiate pancreatobiliary-type from intestinal-type periampullary carcinomas, in particular, the model of all phases using the LR algorithm.
(Copyright © 2021 The Royal College of Radiologists. Published by Elsevier Ltd. All rights reserved.)