Aims: Congenital urinary tract dilatations are among the most common anomalies in pediatric urology and may lead to significant morbidity if not properly managed. Although magnetic resonance urography (MRU) enables detailed anatomic and functional assessment, conventional interpretation remains subjective and operator-dependent. Radiomics, coupled with machine learning (ML) and eXplainable Artificial Intelligence (XAI), has the potential to provide objective and reproducible diagnostic support.Methods: The dataset used in this study comprises a subset of data obtained from a previously completed medical residency thesis. From the computerized archives of this thesis, data from 13 patients could be retrieved. For radiological assessment and ML modeling, three-dimensional heavily T2-weighted images were utilized. Radiomic features were extracted, and ML–based classification models were developed to predict disease severity. To identify the most relevant imaging features contributing to model performance, XAI methods, including Shapley Additive Explanations (SHAP), were applied. No clinical variables were incorporated into the modeling pipeline; the analysis was based exclusively on imaging-derived radiomic features. Results: The radiomics-based ML model demonstrated preliminary classification performance in this small cohort, as assessed by cross-validation metrics. SHAP analysis revealed that texture and intensity-derived features were the most influential predictors of disease severity. Conclusion: Radiomics combined with ML and XAI represents a promising and technically feasible approach for the evaluation of congenital urinary tract dilatation in small exploratory cohorts. While the present findings are preliminary, this framework may support future development of decision-support tools following validation in larger, independent datasets.
Kılınç et al. (2026) studied this question.