A radiomics model of contrast-enhanced computed tomography for predicting post-acute pancreatitis diabetes mellitus.

BACKGROUND: Diabetes mellitus can occur after acute pancreatitis (AP), but the accurate quantitative methods to predict post-acute pancreatitis diabetes mellitus (PPDM-A) are lacking. This retrospective study aimed to establish a radiomics model based on contrast-enhanced computed tomography (CECT) for predicting PPDM-A.

METHODS: A total of 374 patients with first-episode AP were retrospectively enrolled from two tertiary referral centers. There were 224 patients in the training cohort, 56 in the internal validation cohort, and 94 in the external validation cohort, and there were 86, 22, and 27 patients with PPDM-A in these cohorts, respectively. The clinical characteristics were collected from the hospital information system. A total of 2,398 radiomics features, including shape-based features, first-order histogram features, high order textural features, and transformed features, were extracted from the arterial- and venous-phase CECT images. Intraclass correlation coefficients were used to assess the intraobserver reliability and interobserver agreement. Random forest-based recursive feature elimination, collinearity analysis, and least absolute shrinkage and selection operator (LASSO) were used for selecting the final features. Three classification methods [eXtreme Gradient Boosting (XGBoost), Adaptive Boosting, and Decision Tree] were used to build three models and performances of the three models were compared. Each of the three classification methods were used to establish the clinical model, radiomics model, and combined model for predicting PPDM-A, resulting in a total of nine classifiers. The predictive performances of the models were evaluated by the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1-score.

RESULTS: Eleven radiomics features were selected after a reproducibility test and dimensionality reduction. Among the three classification methods, the XGBoost classifier showed better and more consistent performances. The AUC of the XGBoost's radiomics model to predict PPDM-A in the training, internal, and external cohorts was good (0.964, 0.901, and 0.857, respectively). The AUC of the XGBoost's combined model to predict PPDM-A in the training, internal, and external cohorts was good (0.980, 0.901, and 0.882, respectively). The AUC of the XGBoost's clinical model to predict PPDM-A in the training, internal, and external cohorts did not perform well (0.685, 0.733, and 0.619, respectively). In the external validation cohort, the AUC of the XGBoost's radiomics model was significantly higher than that of the clinical model (0.857 vs. 0.619, P<0.001), but there was no significant difference between the combined and radiomics models (0.882 vs. 0.857, P=0.317).

CONCLUSIONS: The radiomics model based on CECT performs well and can be used as an early quantitative method to predict the occurrence of PPDM-A.