Artificial Intelligence–Driven Volumetric Quantification of Urinary Stone Burden and Hydronephrosis From CT Images

Abstract Purpose Given the limitations of conventional stone burden assessment based on maximum diameter or ellipsoid-derived volume estimates, we aimed to develop an artificial intelligence (AI)–based method for automated quantification of stone and hydronephrosis volumes from CT images. Materials and Methods A 2-step algorithm was developed. First, a U-Net incorporating residual blocks, dense blocks, a morphology-aware block, and attention gates (RDAM U-Net) segmented the kidneys; hyperparameters were optimized using Complex System Response (CSR) experimental design with 5-fold cross-validation. Second, within the segmented kidney, hydronephrosis was identified from low–Hounsfield unit (HU) regions, and renal stones from high-HU regions using morphological processing and 3-dimensional (3D) connected-component analysis. AI-derived stone volumes were compared with ellipsoid-based estimates using Pearson and Spearman correlation, Lin's concordance correlation coefficient (CCC), intraclass correlation coefficient (ICC), and Bland–Altman analysis. Results Eighty-two kidney CT cases were used for model development. RDAM U-Net achieved a mean intersection over union of 0.9432 and Dice coefficient of 0.9681, with fewer parameters and a 32.58% shorter parameter-update time than reference attention-based U-Nets. In 39 paired stone measurements, AI-derived and ellipsoid-estimated volumes showed a near-perfect correlation (Pearson r = 0.977; Spearman ρ = 0.966; R² = 0.954; all p < 0.001) and high agreement (Lin's CCC = 0.959; ICC = 0.960). AI volumes were systematically lower than ellipsoid estimates (mean difference, − 51.21 mm³; p = 0.001), with the difference increasing with stone volume. HU-based composition analysis agreed with the clinical record in 12 of 13 classifiable stones (92.3%). Conclusions AI-based CT segmentation enabled automated volumetric quantification of both urinary stone burden and hydronephrosis, providing direct 3D assessment without reliance on geometric assumptions. Integrating stone volume, hydronephrosis volume, and stone composition into a single automated workflow may support more individualized surgical and metabolic management of urolithiasis. Further validation in larger external cohorts is warranted.