Can AI Imaging Improve Risk Stratification in Upper Tract Urothelial Carcinoma?

Preoperative decision-making in upper tract urothelial carcinoma (UTUC) remains constrained by the limitations of current diagnostic tools. Computed tomography urography, urinary cytology, and ureteroscopic biopsy all contribute important information, but accurately identifying tumor grade, muscle invasion, and prognosis before surgery is still difficult.

A narrative review published in Frontiers in Oncology examines whether radiomics and deep learning can bridge these gaps by extracting quantitative information from routine imaging that may improve preoperative risk assessment, while also highlighting the challenges that continue to limit translation into practice.

From Conventional Imaging to Quantitative Risk Assessment
The review summarizes evidence from studies identified through searches of PubMed, Web of Science, and Scopus, focusing on radiomics, machine learning, and deep learning applications in UTUC. Rather than relying on visual image interpretation alone, radiomics converts CT or MRI scans into hundreds or thousands of quantitative features that capture tumor heterogeneity. Deep learning approaches, especially convolutional neural networks, learn imaging patterns directly from the data without handcrafted feature engineering. Together, these methods aim to provide noninvasive preoperative risk stratification by identifying imaging features that may reflect underlying tumor biology.

The review also outlines the typical workflow, from image acquisition and tumor segmentation through feature extraction and selection, model development, and validation. Importantly, the authors emphasize that variability in imaging protocols, segmentation methods, and feature selection can substantially influence model performance, making methodological rigor essential for reproducibility.

Strong Early Performance Across Several Clinical Questions
Among the most mature applications is preoperative tumor grading, a key determinant of whether patients undergo kidney-sparing surgery or radical nephroureterectomy. Several CT- and MRI-based radiomics models achieved excellent discrimination, with reported area under the curve (AUC) values frequently exceeding 0.90 in development cohorts. One model incorporating both tumor and perirenal fat radiomic features reached an AUC of 0.961, suggesting that the surrounding microenvironment may add predictive value beyond tumor characteristics alone. Deep learning models have also demonstrated high sensitivity and specificity for automated grade prediction.

Radiomics has also shown promise in distinguishing UTUC from renal cell carcinoma, an important diagnostic challenge because surgical management differs substantially between the two diseases. Reported validation AUCs of approximately 0.87 to 0.90 indicate that quantitative imaging may outperform conventional visual assessment in selected settings. Similarly, CT-based radiomics models have predicted muscle invasion with validation AUCs exceeding 0.80, potentially improving preoperative assessment where standard imaging remains limited.

Prognostic modeling represents another emerging application. Studies integrating radiomic features with clinical variables reported improved prediction of recurrence and overall survival, while multimodal deep learning models combining imaging and clinical data achieved AUCs approaching 0.90 for prognostic stratification. These findings suggest imaging-derived biomarkers could eventually support more individualized treatment planning before surgery, although prospective validation and demonstration of clinical benefit are still needed before routine clinical implementation.

Methodological Hurdles Remain Substantial
Despite encouraging performance metrics, the review repeatedly cautions against overinterpreting current evidence. Most published studies are retrospective, single-center analyses with relatively small cohorts, creating a high risk of overfitting. External validation remains uncommon, and model performance often declines when tested outside the original development dataset. In some studies, ureteroscopic biopsy rather than final surgical pathology served as the reference standard, introducing the possibility of misclassification because biopsy frequently underestimates tumor grade.

The authors also highlight broader methodological concerns, including inconsistent imaging protocols, manual segmentation variability, inadequate reporting of data leakage, and limited adherence to quality frameworks such as the Radiomics Quality Score (RQS), TRIPOD-AI, and CLAIM. Perhaps most importantly, few studies demonstrate that AI provides clinically meaningful improvements over expert radiologist interpretation or influences patient outcomes.

Looking ahead, the review argues that prospective multicenter validation, standardized imaging and reporting protocols, improved model interpretability, and integration with molecular and genomic data will be essential before radiomics and deep learning can become routine components of UTUC management. At present, these technologies offer a promising framework for noninvasive risk stratification, but they remain largely investigational pending stronger evidence of clinical utility.

Reference:
Zhang Y, Wu G, Sun F, Wang B, Guo Y, Wu J. Radiomics and deep learning in upper tract urothelial carcinoma: advancing preoperative risk stratification and clinical decision-making. Front Oncol. 2026;16:1838755. Published 2026 Jun 19. doi:10.3389/fonc.2026.1838755