Radiomics model of contrast‐enhanced MRI for early prediction of acute pancreatitis severity

Abstract

Background Computed tomography (CT) or MR images may cause the severity of early acute pancreatitis (AP) to be underestimated. As an innovative image analysis method, radiomics may have potential clinical value in early prediction of AP severity. Purpose To develop a contrast‐enhanced (CE) MRI‐based radiomics model for the early prediction of AP severity. Study Type Retrospective. Subjects A total of 259 early AP patients were divided into two cohorts, a training cohort (99 nonsevere, 81 severe), and a validation cohort (43 nonsevere, 36 severe). Field Strength/Sequence 3.0T, T 1 ‐weighted CE‐MRI. Assessment Radiomics features were extracted from the portal venous‐phase images. The "Boruta" algorithm was used for feature selection and a support vector machine model was established with optimal features. The MR severity index (MRSI), the Acute Physiology and Chronic Health Evaluation (APACHE) II, and the bedside index for severity in acute pancreatitis (BISAP) were calculated to predict the severity of AP. Statistical Tests Independent t ‐test, Mann–Whitney U ‐test, chi‐square test, Fisher's exact tests, Boruta algorithm, receiver operating characteristic analysis, DeLong test. Results Eleven potential features were chosen to develop the radiomics model. In the training cohort, the area under the curve (AUC) of the radiomics model, APACHE II, BISAP, and MRSI were 0.917, 0.750, 0.744, and 0.749, and the P value of AUC comparisons between the radiomics model and scoring systems were all less than 0.001. In the validation cohort, the AUC of the radiomics model, APACHE II, BISAP, and MRSI were 0.848, 0.725, 0.708, and 0.719, respectively, and the P value of AUC comparisons were 0.96 (radiomics vs. APACHE II), 0.40 (radiomics vs. BISAP), and 0.46 (radiomics vs. MRSI). Data Conclusion The radiomics model had good performance in the early prediction of AP severity. Level of Evidence: 3 Technical Efficacy Stage: 2 J. Magn. Reson. Imaging 2020;51:397–406.