Automated body composition quantification from non-contrast CT for urolithiasis classification and exploratory...

To develop a deep learning-based body composition quantification framework from non-contrast CT for urolithiasis classification (calcium, non-calcium, stone-free individuals) and incident risk stratification for de novo radiologically-confirmed urinary stone formation in baseline stone-free individuals. This retrospective multicenter study included 781 participants. A classification cohort (n = 481; 246 calcium, 104 non-calcium, 131 stone-free individuals) was divided into training, internal test and external validation sets. A separate longitudinal cohort (n = 300) of stone-free individuals who underwent abdominal CT for non-urological indications was followed for a median of 5.2 years. Automated L1/L3 muscle/fat segmentation informed clinical, radiomics, and combined models, which were evaluated via area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. The combined muscle (L1 + L3)-clinical model achieved the highest classification performance (external AUC: 0.90; 95% CI: 0.78-0.98). In the longitudinal cohort (median follow-up, 5.2 years), 8 incident urolithiasis events occurred. Using a prespecified threshold, high model-derived risk status was observed in 6 of 8 participants who later developed incident urolithiasis (sensitivity of 75%; 95% CI: 0.35-0.97), demonstrating a preliminary association with subsequent stone events, alongside a specificity of 89% (260 of 292 non-events). This automated body composition model supports accurate urolithiasis classification and shows a preliminary association with incident stone events, warranting prospective validation.