CLIP-CGPNet: A Multimodal Clinical Decision Support System for Postoperative UTUC Recurrence Prediction

Accurate prediction of postoperative recurrence is essential for individualized management and follow-up planning in patients with upper tract urothelial carcinoma (UTUC). However, reliable risk stratification remains challenging due to the heterogeneous nature of imaging characteristics and clinical information. In this study, we propose a multimodal clinical decision support framework that integrates contrast-enhanced computed tomography (CT) images with routinely available clinical variables for postoperative recurrence prediction. The proposed framework adopts an image-clinical integrated learning strategy to enhance semantic alignment between visual representations and patient-specific clinical context. Imaging features are guided by clinically informed semantic prompts, while a dynamic fusion mechanism adaptively balances the contributions of different modalities. This design enables robust multimodal interaction and improves the individualized risk estimation. The framework is evaluated using a retrospective UTUC cohort of 160 patients treated between 2016 and 2023, including 59 recurrence cases and 101 nonrecurrence cases. Experimental results demonstrate that the proposed method achieves an area under the receiver operating characteristic curve (AUC) of 0.87, demonstrating superior performance compared with conventional convolutional neural network-based models and representative multimodal baselines. Qualitative analyses further indicate improved alignment between imaging patterns and clinical factors, providing interpretable insights through attention-based visualization. These findings suggest that the proposed multimodal framework provides a potential approach for postoperative UTUC recurrence risk stratification and has the potential to support clinical decision-making and risk-adapted postoperative management.