Prediction model based on contrast-enhanced computed tomography images and clinical indicators for the prognosis of pancreatic necrosis in acute pancreatitis

The clinical outcomes of acute necrotizing pancreatitis (ANP) including necrotic tissue absorption, persistent walled-off necrosis (WON) formation, and infectious pancreatic necrosis (IPN) require different treatment modalities. In this prospective observational study, patients with ANP admitted to our hospital between January 2021 and December 2022 underwent contrast-enhanced computed tomography (CECT) and clinical tests within 24 hours of admission and were followed up for 6 months. CECT images of the pancreas were automatically segmented using deep learning (DL) model, and 3D ResNet DL and logistic regression (LR) models were developed using CECT images and selected clinical indicators, respectively. Prediction models were obtained via the integration of the DL and LR models, and comparison of their respective performances. Of the 133 patients with ANP, absorption of necrotic tissues, persistent WON, and IPN were found in 45.86, 30.83, and 23.31% of the patients, respectively. For pancreatic segmentation, the Attention U-Net model performed better than the U-Net model. Blood glucose, urea nitrogen, lactate dehydrogenase, and C-reactive protein levels were then used to construct an LR model with .714 accuracy. The accuracies of the 3D ResNet models using manually and automatically segmented pancreatic images were .821 and .750 initially, respectively, and .857 and .786 when combined with LR, respectively. The model developed in this study may be clinically applied to improve the accuracy of ANP prognosis prediction.