Ultrasound-based intratumoral and peritumoral radiomics for preoperative prediction of lymph node metastasis in pancreatic ductal adenocarcinoma.

BACKGROUND: Accurate preoperative assessment of lymph node metastasis (LNM) in pancreatic ductal adenocarcinoma (PDAC) remains challenging. We developed and compared ultrasound-based intratumoral, peritumoral, clinical, and combined models for LNM prediction and explored the complementary value of multi-regional imaging. METHODS: Ninety-nine patients with pathologically confirmed PDAC who underwent preoperative ultrasound were retrospectively enrolled. Intratumoral and 3-mm peritumoral ROIs were manually delineated. Radiomics features were extracted using PyRadiomics and selected via reproducibility filtering, correlation analysis, and LASSO. Nine machine-learning algorithms were evaluated to identify the optimal classifier for each region (intratumoral: logistic regression; peritumoral: random forest). A decision-level combined model was constructed by integrating regional model outputs with clinical information. Discrimination was assessed by AUC with sensitivity, specificity, accuracy, PPV, and NPV; calibration by calibration curves and the Hosmer-Lemeshow test; clinical utility by decision curve analysis (DCA). Model interpretability and inter-model relationships were explored using SHAP, correlation, and Bland-Altman analyses. RESULTS: The intratumoral and peritumoral models achieved AUCs of 0.815 and 0.792, respectively. The combined model yielded the highest performance (AUC = 0.898, 95% CI: 0.770-1.000) with good calibration (Hosmer-Lemeshow p = 0.091) and the greatest net benefit on DCA. DeLong tests showed no statistically significant AUC differences among models. Intratumoral and peritumoral outputs were moderately correlated (Pearson's r = 0.711, p < 0.001), and Bland-Altman analysis demonstrated overall agreement with several outliers, suggesting complementary region-specific information. CONCLUSION: Intratumoral and peritumoral ultrasound radiomics provide complementary information for preoperative LNM prediction in PDAC. The decision-level combined model achieved numerically higher discrimination with favorable calibration, supporting further validation before clinical translation.