UTUC-PNet: A Multimodal Deep Learning Model with Feature Attribution for Postoperative Recurrence Risk Prediction

Background and Objective: Upper tract urothelial carcinoma (UTUC) is a highly aggressive malignancy characterized by a high postoperative recurrence rate, posing substantial challenges in clinical management and prognosis. Accurate prediction of recurrence risk is critical for optimizing treatment strategies. However, conventional imaging-based assessments predominantly rely on expert interpretation, which is subjective and limited in capturing subtle prognostic features.Methods: This study proposes UTUC-PNet, a deep learning-based recurrence risk prediction framework that integrates contrast-enhanced CT imaging and key clinical indicators. The model combines deep feature learning with a multimodal fusion strategy to extract recurrence-associated patterns. A logistic regression-based feature attribution module is incorporated to quantify the contribution of clinical and imaging-derived variables, ensuring interpretability. A complete prediction pipeline is constructed, including feature representation, weighted risk computation, and score calibration.Results: UTUC-PNet demonstrated competitive predictive performance compared to conventional CNN models, with improvements in accuracy, generalization, and interpretability. On a real-world dataset, the model achieved an accuracy of 95.03% and an F1-score of 97.62%. The interpretability mechanism provides clinically meaningful insights by highlighting influential features in risk assessment.Conclusions: The proposed framework provides a promising and interpretable approach for recurrence risk stratification in postoperative UTUC patients. While the results are encouraging, further validation on larger, multi-center cohorts is required before clinical application.