CT-based Deep Learning Radiomics Nomogram for Differentiating pyelocaliceal Upper Tract Urothelial Carcinoma and Xanthogranulomatous Pyelonephritis

Abstract Objective This study aims to establish a CT-based clinical deep learning radiomics (CDLR) model for preoperative prediction of pyelocaliceal Upper Tract Urothelial Carcinoma (UTUC) and xanthogranulomatous pyelonephritis (XGP), providing scientific guidance for personalized treatment. Methods A retrospective analysis was conducted on 161 post-operative pathology-confirmed cases of pyelocaliceal UTUC and XGP patients, divided into training cohort (n = 112) and validation cohort (n = 49). Radiomics (Rad) and deep learning (LR) features were extracted from three-phase CT images, combined with clinical features, and after feature selection, a logistic regression (LR) classifier was used to construct clinical, radiomics (Rad), deep learning (LR), deep learning radiomics (DLR), and clinical deep learning radiomics (CDLR) models. The top-performing model was chosen utilizing receiver operating characteristic (ROC) curve analysis. Decision curve analysis (DCA) was used to evaluate the model's practicality. Results The evaluation of clinical, Rad, LR, DLR, and CDLR models revealed that the CDLR model exhibited superior diagnostic performance. The area under the receiver operating characteristic curve (AUC) of the CDLR model in the training and validation cohorts were 0.984 and 0.970, respectively, outperforming other models (clinical model, Rad model, DL model, DLR model). DCA results showed that the CDLR model provided a higher net benefit in preoperative prediction. Conclusion The CDLR model, combining clinical, Rad, and LR features, could serve as a non-invasive tool for differentiating pyelocaliceal UTUC and XGP, offering valuable guidance for clinical treatment.