A Multi-Data Fusion Deep Learning Model for Prognostic Prediction in Upper Tract

Hongdi Sun¹Siping Chen²Yongxing Bao³Fengyan You³HongHui Zhu³Xin Yao³Lianguo Chen³Jiangwei Miao²Fanggui Shao^3*Xiaomin Gao^3*Binwei Lin^2*

• ¹Department of Hematology, The Third Clinical Institute Affiliated to Wenzhou Medical University (Wenzhou People’s Hospital), Wenzhou, Zhejiang, 325006, P.R. China, Wenzhou, China
• ²Department of Urology, Rui’an People’s Hospital, The Third Affiliated Hospital of the Wenzhou Medical University, Wenzhou, China
• ³The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China

Upper tract urothelial carcinoma (UTUC) is a rare but highly invasive malignancy of the urinary system, characterized by a high recurrence rate. This study aimed to develop a multi-modal deep learning (DL) model to predict postoperative prognosis in UTUC patients using multi-phase contrast-enhanced CT imaging and clinical data. A total of 133 patients diagnosed with UTUC who underwent radical nephroureterectomy between 2005 and 2017 were retrospectively included. Of these, 103 patients were allocated to the training set and 30 to the testing set. A multi-modal DL model-termed the Multi-modal Image-Clinical Combination Classifier (MICC)-was developed by integrating CT imaging and clinical data. The prognostic performance of MICC was compared with two unimodal models: ImageNet (CT-based) and ClinicalNet (clinical data-based), as well as with traditional clinical parameters, including pathological T stage. Feature importance was assessed using SHapley Additive exPlanations (SHAP). The MICC model demonstrated superior prognostic performance in both training and testing sets, achieving AUCs of 0.918 and 0.895, respectively, and outperforming both the ImageNet and ClinicalNet models. Robust classification metrics were observed (accuracy: 0.854; sensitivity: 0.889; specificity: 0.836; NPV: 0.933; PPV: 0.744). Precision-recall analysis further supported the model's strength in identifying high-risk patients within an imbalanced dataset. SHAP analysis revealed that CT imaging features contributed most significantly to model predictions. Therefore, the MICC model, which integrates multi-phase CT imaging and clinical variables, provides accurate prognostic prediction for UTUC.This approach may assist clinicians in tailoring personalized treatment strategies and improving patient outcomes.