Multimodal PET/CT–MRI Radiomics and Deep Learning Fusion for Individualized Prognostic Prediction in Nasopharyngeal Carcinoma

Objectives Nasopharyngeal carcinoma shows considerable biological heterogeneity that leads to wide variations in clinical outcomes. Anatomy‐based staging and single‐modality imaging cannot fully represent the complexity of tumor phenotype. This study aimed to develop and validate a prognostic model for progression‐free survival by integrating multimodal positron emission tomography, computed tomography, magnetic resonance imaging, quantitative radiomics, deep learning representations, and clinical variables. Methods A total of 261 patients with locoregionally advanced disease were retrospectively included. Radiomics and deep learning features were extracted from pretreatment positron emission tomography/computed tomography and T1‐weighted magnetic resonance imaging. Fourteen survival models were constructed using Cox proportional hazards regression. Model performance was assessed through concordance index, time‐dependent receiver operating characteristic analysis, Brier score, and calibration. The Friedman–Nemenyi procedure compared overall performance across models, and SHapley Additive Explanation identified key prognostic contributors. Results Multimodal fusion models outperformed clinical‐only and single‐modality models. The fully integrated model achieved the highest discrimination and stable performance across multiple follow‐up intervals, with consistently strong calibration. Ranking analysis showed that deep‐feature‐dominant models performed worst, radiomics‐based models achieved intermediate performance, and multimodal models formed the top‐performing group. Feature attribution analysis highlighted several texture descriptors and deep representations as major predictors associated with adverse outcomes. Conclusion The integrated multimodal model demonstrated superior accuracy, robustness, and interpretability for predicting progression‐free survival. This approach provides a reliable imaging‐based tool for individualized risk stratification and may contribute to more precise treatment planning in clinical practice.