A multimodal fusion model integrating Vision Transformer, radiomics, and clinical features for predicting bone...

To evaluate the performance of a multimodal fusion framework integrating a Vision Transformer (ViT), radiomics, and clinical features for predicting bone metastasis (BM) status in patients with prostate cancer (PCa).

Patients with pathologically confirmed PCa were retrospectively included. Based on clinical features and apparent diffusion coefficient (ADC) images, three single-modal models were constructed: the clinical model (Model_Clin), the radiomics model (Model_Rad), and the ViT model (Model_ViT). Subsequently, a multimodal fusion model (Model_Fusion) was constructed by integrating ViT, radiomics, and clinical features. Model performance was evaluated using the receiver operating characteristic (ROC) curve and the DeLong test. The clinical utility and interpretability of the Model_Fusion were assessed using decision curve analysis (DCA) and Shapley additive explanations (SHAP).

Model_ViT demonstrated the best performance among the single-modal models, achieving AUCs of 0.909 and 0.872 in the training and validation sets, respectively, outperforming both Model_Rad (AUC = 0.885 and 0.842) and Model_Clin (AUC = 0.861 and 0.781). By integrating multimodal information, Model_Fusion achieved superior performance (AUC = 0.944 and 0.894). DeLong test results showed that, in the training set, Model_Fusion had a significantly higher AUC than Model_Clin, Model_Rad, and Model_ViT (all P < 0.05), whereas in the validation set, a significant difference was observed only when compared with Model_Clin. DCA further demonstrated that Model_Fusion provided a higher net benefit. SHAP analysis indicated that the predicted probability of ViT contributed the most to Model_Fusion.

A fusion model integrating ViT, radiomics, and clinical features provides a non-invasive framework for predicting BM in PCa, which may help guide personalized clinical decision-making and prognostic evaluation.