Prediction of residual stones after ESWL for ureteral calculi using deep learning-radiomics-based stone heterogeneity and body composition: an interaction analysis

ObjectiveTo investigate the association between deep learning-radiomics-based stone heterogeneity, body composition, and residual stones after extracorporeal shock wave lithotripsy (ESWL) for ureteral calculi.MethodsThis retrospective study included 200 patients with ureteral stones who underwent ESWL between January 2024 and December 2025, comprising the development cohort. An additional 57 patients treated between January 2026 and June 2026 were included as the temporal testing cohort. Baseline clinical data and imaging data were collected. Deep learning (DL) and radiomics features of ureteral stones were extracted, and feature selection was performed to construct a DL-radiomics model (Radscore). TotalSegmentator was used to quantify visceral fat, skeletal muscle, and subcutaneous fat volumes within the T12-S4 region. The predictive performance of the clinical-body composition model, DL-radiomics model, and nomogram model was compared. Subgroup analyses were conducted to evaluate the stability of the nomogram model. The interaction between the Radscore and body composition in relation to residual stones was further explored using the optimal cutoff values derived from the training set.ResultsThe development cohort was randomly divided into a training set and a validation set in a 7:3 ratio. Visceral fat and skeletal muscle volumes were included in the clinical-body composition model, yielding AUCs of 0.792 (0.706, 0.878), 0.821 (0.577, 1.000) and 0.827 (0.710, 0.943), respectively. The DL radiomics model achieved AUCs of 0.778 (0.679−0.876), 0.704 (0.363−1.000) and 0.742 (0.547−0.937). The nomogram model showed AUCs of 0.893 (0.835, 0.952), 0.824 (0.630, 1.000) and 0.898 (0.807, 0.989). Subgroup analyses of stone location, side of the stone, and perinephritis revealed no significant interaction in the nomogram model (P > 0.05). Calibration and decision curve analyses demonstrated good agreement and clinical utility of the nomogram model. In the interaction analysis between Radscore and body composition, visceral fat volume significantly attenuated the risk of residual stones associated with a high stone Radscore.ConclusionDeep learning-radiomics-based stone heterogeneity demonstrates good predictive performance for residual stones after ESWL. Moreover, lower visceral fat volume significantly attenuates the impact of stone heterogeneity on residual stone formation.