Development of a reoperative risk prediction model of muscle-invasive upper tract urothelial carcinoma using clinical and radiomic computed tomography features: Initial results from a multi-institutional Canadian study.
Authors
Affiliations (13)
Affiliations (13)
- Divisions of Urology and Surgical Oncology, Department of Surgery, University Health Network, University of Toronto, Toronto, ON, Canada.
- Institute of Health Policy, Management and Evaluation, Dalla Lana School of Public Health, University of Toronto, Toronto, ON, Canada.
- Joint Department of Medical Imaging, Sinai Health System, University of Toronto, Toronto, ON, Canada.
- Department of Urology, St James's Hospital, Dublin, Ireland.
- Department of Surgery, School of Medicine, Trinity College Dublin, Dublin, Ireland.
- Department of Biostatistics, University Health Network, University of Toronto, Toronto, ON, Canada.
- Department of Urology, King Faisal Specialist Hospital and Research Center, Riyadh, Saudi Arabia.
- Division of Urology, University of Ottawa, Ottawa, ON, Canada.
- Division of Urology, McGill University Health Center, Montreal, QC, Canada.
- Department of Urologic Sciences, University of British Columbia, Vancouver, BC, Canada.
- Division of Urology, Department of Surgery, Université Laval, Quebec City, QC, Canada.
- Department of Urology, Dalhousie University, Halifax, NS, Canada.
- Department of Urology, Queen's University, Kingston, ON, Canada.
Abstract
Accurate pre-intervention staging of upper tract urothelial carcinoma (UTUC) remains a significant clinical challenge, particularly in identifying muscle-invasive disease (≥pT2), where kidney-sparing surgery may not be appropriate. Current imaging and biopsy approaches are often inadequate. Radiomics, which extracts high-dimensional features from medical imaging, may improve non-invasive staging. This study assessed whether computed tomography (CT)-based radiomic features, alone or combined with clinical data, could predict ≥pT2 UTUC in a multicenter Canadian cohort. We retrospectively analyzed clinical, pathologic, and radiographic features of patients with UTUC who underwent extirpative surgery at five academic centers from January 2, 2001, to May 1, 2023. Radiomic features were extracted from machine-learning segmentations of the affected kidney using the excretory phase of CT. Predictive models were developed using clinical only, radiomic only, and combined data to predict stage ≥pT2. Feature selection included univariable logistic regression, correlation filtering, and LASSO. Model performance was assessed via five-fold cross-validation repeated 10 times, with area under the curve (AUC) as the primary metric. Of 441 patients, 208 (47.2%) were included. Of the 208 patients, 97 (46.6%) had ≥pT2 disease. The clinical model (AUC 0.602) included age, hydronephrosis, and high-grade cytology. The radiomics model, based on two texture features, achieved an AUC of 0.653. The combined model achieved an AUC of 0.647. Radiomics and combined models significantly outperformed the clinical model (p<0.01), but did not differ from each other. For 117 patients with renal pelvis cancers, the combined model's discrimination performance was statistically better than the clinical model (AUC 0.708 vs. AUC 0.607, p<0.001). Likewise, the radiomics' AUC discrimination performance was statistically better than the clinical model (AUC 0.694 vs. AUC 0.607, p=0.004). In contrast, we found no significant difference in model performance in the non-renal pelvis subgroup (n=91). Conventional radiomics improved the prediction of muscle-invasive UTUC compared to clinical models alone, but overall accuracy remained suboptimal for clinical use. Heterogeneity in CT protocols and challenges with tumor segmentation were the main limitations. Future work should develop more adaptable AI models trained on larger, more diverse datasets to better reflect real-world imaging conditions.