Objective:

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).

Approach:

• Model Construction: Three single-modal models were constructed: the clinical model (Model_Clin), the radiomics model (Model_Rad), and the ViT model (Model_ViT). A multimodal fusion model (Model_Fusion) was created by integrating these models.
• Performance Evaluation: Model performance was evaluated using the receiver operating characteristic (ROC) curve and the DeLong test. Clinical utility and interpretability were assessed using decision curve analysis (DCA) and Shapley additive explanations (SHAP).

Key Findings:

• Model_ViT achieved AUCs of 0.909 and 0.872 in training and validation sets, respectively.
• Model_Fusion outperformed all single-modal models with AUCs of 0.944 and 0.894.
• Model_Rad achieved AUCs of 0.885 and 0.842, while Model_Clin achieved AUCs of 0.861 and 0.781.
• DeLong test indicated significant performance differences in the training set for Model_Fusion compared to other models (all P < 0.05).
• DCA showed that Model_Fusion provided a higher net benefit.
• SHAP analysis revealed that the predicted probability of ViT was the most influential in Model_Fusion.

Interpretation:

Limitations:

• The study is retrospective and may be subject to selection bias.
• The sample size and diversity of the patient population may limit generalizability.

Conclusion:

The multimodal fusion approach shows potential in improving the prediction of bone metastasis in prostate cancer patients.