A perioperative multi-modal fusion and deep learning-based prognostic system for upper tract urothelial carcinoma: a multi-institutional study.
Authors
Affiliations (9)
Affiliations (9)
- Department of Urology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China.
- Department of Urology, Urologic Surgery Center, Xinqiao Hospital, Third Military Medical University (Army Medical University), Chongqing, China.
- Center for Reproductive Medicine, Women and Children's Hospital of Chongqing Medical University, Chongqing Health Center for Women and Children, Chongqing, China.
- Department of Urology, Chongqing University Fuling Hospital, Chongqing, China.
- Department of Urology, The Third Affiliated Hospital of Chongqing Medical University, Chongqing, China.
- Department of Urology, Yongchuan Hospital of Chongqing Medical University, Chongqing, China.
- Department of Urology, The Second Affiliated Hospital of Chongqing Medical University, Chongqing, China.
- Department of Urology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China. [email protected].
- Department of Urology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, China. [email protected].
Abstract
Precise perioperative risk stratification for upper tract urothelial carcinoma (UTUC) is essential. We developed a multimodal prognostic model integrating perioperative clinical data, radiomics, and deep learning (DL) features from baseline CT urography to improve survival prediction and guide adjuvant management. We retrospectively enrolled 623 patients from six institutions, divided into training, internal validation, and independent external validation sets. Four single-modal models (clinical, radiomics, 2D DL, and 2.5D DL) were developed, and an integrated combined model was constructed by fusing their prognostic scores. Performance was evaluated using the C-index, area under the curve (AUC), calibration curves, and decision curve analysis (DCA). The combined model consistently outperformed all single-modal models across all cohorts. C-indices reached 0.758 (95% CI: 0.712-0.804), 0.725 (95% CI: 0.651-0.798), and 0.704 (95% CI: 0.631-0.777) in the training, internal validation, and external validation sets, respectively, numerically surpassing the best single-modal models. Notably, our 2.5D DL model (C-index: 0.705) demonstrated a consistent incremental improvement over the 2D DL model (C-index: 0.681) in capturing prognostic information. In external validation, the combined model achieved a 3-year AUC of 0.766. DCA indicated the comprehensive model exhibited excellent calibration and provided the highest net benefits. This multimodal system, featuring a robust 2.5D DL strategy, improves overall survival prediction in UTUC. It offers a valuable tool for accurate perioperative risk stratification immediately after radical nephroureterectomy, demonstrating particularly reliable value for 3-year intermediate-term clinical decision-making. This multimodal system advances clinical radiology by fusing perioperative clinical data, radiomics, and DL features from CTU images, enhancing risk stratification accuracy to guide postoperative adjuvant management for UTUC. Current prognostic models for UTUC lack accuracy, creating an unmet clinical need for precise perioperative risk stratification to guide adjuvant management. A multimodal prognostic model fusing clinical, radiomic, and DL features from baseline CT urography consistently outperformed single-modality models in predicting overall survival.