Multimodal fusion of radiomics, dosiomics and deep learning for predicting radiation-induced hypothyroidism after radiotherapy in nasopharyngeal carcinoma: a multicenter study

Abstract

To develop and externally validate a prediction framework integrating clinical factors, dose‒volume histogram (DVH) parameters, radiomics, dosiomics and deep - learning (DL) features for individualised radiation-induced hypothyroidism (RIHT) risk assessment. This retrospective multicentre study included 396 patients with nasopharyngeal carcinoma (NPC) treated with radiotherapy between January 2020 and December 2024. Patients from Fujian Cancer Hospital were randomly divided into training and internal validation cohorts, whereas patients from the First Affiliated Hospital of University of South China formed an independent external-test cohort. RIHT was treated as a binary endpoint defined by persistent thyroid-stimulating hormone (TSH) elevation above the institutional upper reference limit during follow-up. Clinical factors, thyroid DVH parameters, computed tomography (CT) radiomic features, dose-distribution dosiomic features, and ResNet18-derived DL features were extracted and integrated to develop single-modality and multimodal fusion models. Model performance was evaluated via the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, calibration analysis, decision curve analysis (DCA), SHapley Additive exPlanations (SHAP) attribution and gradient-weighted class activation mapping (Grad-CAM) visualisation. Among 396 patients, 150 developed RIHT. In the internal validation cohort, the early-fusion and combined models achieved AUCs of 0.822 and 0.775, respectively; in the external test cohort, the corresponding AUCs were 0.827 and 0.813, with models achieving an accuracy of 0.750 and 0.775, sensitivity of 0.933 and 0.867, and negative predictive value of 0.941 and 0.900. The combined model showed the lowest Brier scores in both validation cohorts, whereas the early fusion model showed stronger discrimination but substantial calibration-slope deviation. Calibration and DCA showed improved agreement and greater clinical net benefit for the fusion models compared with single-modality models. SHAP quantified multimodal feature contributions, and Grad-CAM supported partial anatomical localisation. Multimodal fusion of clinical, DVH, radiomic, dosiomic, and DL features improved the robustness and external generalizability of RIHT prediction after radiotherapy for NPC. The early-fusion and combined model showed the most favorable overall performance, although calibration varied and the external cohort was small. These findings support further prospective multicentre evaluation rather than immediate clinical deployment.