Radiomic features from intratumoral and peritumoral regions on portal venous phase CT for multicenter prediction of TP53 mutation in pancreatic cancer.

Background: TP53 mutation, occurring in 50-70% of pancreatic ductal adenocarcinomas (PDAC), is a major determinant of tumor aggressiveness and treatment response. Current assessments rely on invasive biopsy, underscoring the need for reliable non-invasive prediction. Methods: In this multicenter study, 216 PDAC patients (training = 162; external test = 54) who underwent preoperative portal-venous phase CT (PV-phase CT) were analyzed. Intratumoral and 3-mm peritumoral regions were manually segmented, and 1, 561 radiomic features were extracted. Six machine-learning classifiers were trained following feature selection and SMOTE, both of which were strictly nested within the cross-validation training folds to prevent data leakage. Model performance was evaluated by AUC, DeLong test, decision curve, and calibration analyses; interpretability was assessed using SHAP. Results: The Intra-Peri Model (IPM) combining intratumoral and peritumoral features achieved the best performance. The XGBoost classifier yielded an AUC of 0.893 (95% CI, 0.781-1.000) in the external test set, significantly outperforming single-region models (P < 0.05). SHAP analysis identified intratumoral gray-level skewness and peritumoral texture correlation as the most influential predictors, where greater intratumoral asymmetry and lower peritumoral correlation indicated higher likelihood of TP53 mutation. Conclusion: Integrating intratumoral and peritumoral radiomics enables accurate, non-invasive prediction of TP53 status in PDAC. This model serves as a promising auxiliary tool for individualized treatment planning, warranting further prospective validation.