Abstract
BACKGROUND: This study aimed to develop an MRI-based radiomics score (MRscore) derived from CE-T1WI and T2WI sequences for predicting skull bone destruction in patients with nasopharyngeal carcinoma (NPC), and to evaluate its performance across EBV DNA and sex subgroups. METHODS: This retrospective study included 175 patients with histologically confirmed NPC who underwent baseline MRI. Radiomic features were extracted from CE-T1WI and T2WI sequences. Least absolute shrinkage and selection operator (LASSO) logistic regression was used for feature selection and MRscore construction. Receiver operating characteristic (ROC) analysis was used to assess model performance, and subgroup analyses were conducted based on plasma EBV DNA levels and sex. Cutoff stability was evaluated using bootstrap resampling. RESULTS: Nine radiomic features were selected to construct the MRscore. The model showed good predictive performance for skull bone destruction with an AUC of 0.844. Subgroup analysis demonstrated consistent performance in both EBV DNA-high (AUC = 0.880) and low (AUC = 0.810) groups, as well as in male (AUC = 0.853) and female (AUC = 0.812) patients. The MRscore cutoff value remained stable across bootstrapped samples, and precision–recall curves supported its robustness under class imbalance. CONCLUSIONS: The MRscore provides a non-invasive, individualized tool for predicting skull bone destruction in NPC, demonstrating stable performance across EBV DNA and sex subgroups. These results support its potential role in risk stratification and personalized treatment planning. CLINICAL TRIAL NUMBER: Not applicable. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12880-025-02034-z.