Preoperative prediction of postoperative pancreatic fistula after pancreatic head resection using radiomics and machine learning based on computed tomographic diagnostics

Abstract

Abstract Purpose This study employed machine learning and radiomics to determine whether postoperative pancreatic fistulas (POPF) and perioperative drain amylase dynamics can be predicted prior to pancreaticoduodenectomy by evaluating the radiologic appearance of the pancreatic tissue. Methods 68 patients were included. Radiomic features of the pancreas were extracted from the arterial phase of computed tomography (CT) at a 1 mm slice thickness for each patient. the Radiomic features with highest correlation with POPF for our models, controlling for autocorrelation and applying Bonferroni correction for P-values were selected. For amylase prediction model (APM), radiomic features were correlated with postoperative maximum drain amylase levels at a cut-off of 1000U/l. ROC analysis was performed for evaluation of the resulting prediction models. Results POPF prediction model (PPM) showed an area under the curve (AUC) of 0.897 (confidence interval (CI) = 82.3–97.1%) in the cohort. The AUC of PPM was higher than that for Roberts’ score, but the difference was not statistically significant. An attempt to predict postoperative amylase dynamics in the drainage fluid achieved an AUC of 0.936 (CI = 88%-99.1%). Conclusions Preoperative prediction of POPF and drain amylase dynamics using radiomics showed promising results. Both models offer new approaches to the clinical management of POPF.