Preoperative radiomics models for predicting risks of distinct recurrence patterns in pancreatic ductal adenocarcinoma based on contrast enhanced CT

This study aimed to establish a noninvasive preoperative radiomics–clinical model combining contrast–enhanced computed tomography (CT) features and serum biomarkers to accurately predict the risk of heterogeneous recurrence patterns in resectable pancreatic ductal adenocarcinoma (PDAC), with the goal of optimizing personalized postoperative management. This retrospective study included 290 patients with pathologically confirmed PDAC who underwent curative resection between May 2014 and December 2023. Patients were randomly divided into the training and test cohorts: overall recurrence model (n = 203 vs. 87), local recurrence model (n = 132 vs. 56), and distant metastasis model (n = 111 vs. 47). Radiomic features extracted from preoperative contrast–enhanced CT were selected using Lasso–Cox regression and combined with clinical variables to construct combined models. Model performance was assessed using time–dependent AUC curves, calibration curves, and decision curve analysis. Radiomics–clinical models incorporating carbohydrate antigen 19–9 (CA19–9), American Joint Committee on Cancer (AJCC) stage, tumor enhancement patterns, and radiomics scores achieved superior predictive performance over radiomics–only or clinical–only models. The 12–month AUCs were 0.801 for overall recurrence, 0.896 for distant metastasis, and 0.808 for local recurrence. Calibration and DCA confirmed good agreement and clinical utility. High– and low–risk groups stratified by model scores showed significantly different RFS (P < 0.001). The preoperative radiomics–clinical model accurately predicts the risk of distinct recurrence patterns in resectable PDAC, supporting individualized treatment planning.