Deep learning body composition analysis on non-contrast CT accurately classifies urinary stone type and stratifies incident stone risk

CT-based AI body composition model classified urinary stone composition with AUC 0.90 (95% CI 0.78–0.98) and identified 75% of future incident stones (specificity 89%) in a retrospective multicenter study (n=781).

• Retrospective multicenter study with classification cohort (n=481) for model development and external validation, plus a longitudinal cohort (n=300 stone-free individuals, median follow-up 5.2 years) for incident risk assessment.
• The combined L1+L3 muscle segmentation-clinical model achieved the highest classification accuracy among tested approaches (external AUC 0.90).
• For incident risk, the model was based on only 8 events, and risk stratification requires prospective validation; the classification model was externally validated.

RadPigeon summaries are original and for information only. They are not clinical advice.