Poster #3: Arielle Biron

Automated Classification of Kidney Stone Location on Non-Contrast CT Using Machine Learning

Arielle Biron¹ Joseph Logan, MS¹˒⁵ Gregory Tasian, MD, MSc, MSCE¹˒² Sri Sivalingam, MD, MSc, FRCSC³ Hunter Wessells, MD, FACS⁴ Jonathan Harper, MD⁴ Charles D. Scales Jr., MD, MSHS, FACS⁶ Affiliations 1. Division of Pediatric Urology, The Children’s Hospital of Philadelphia, Philadelphia, PA 2. Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA 3. Glickman Urological Institute, Cleveland Clinic, Cleveland, OH 4. University of Washington School of Medicine, Seattle, WA 5. Department of Biomedical and Health Informatics, The Children’s Hospital of Philadelphia, Philadelphia, PA 6. Duke Clinical Research Institute, Department of Urology, Duke University, Durham, NC

About one in 10 people will experience kidney stones in their lifetime. Stone location is an important characteristic that determines symptom likelihood and surgical success. Currently, radiologists or treating clinicians manually determine this location, a laborious process prone to inter-observer variability. The objective of this study was to develop a machine learning algorithm to automatically classify stone location on non-contrast computed tomography (CT) scans of the abdomen and pelvis. Methods: In a sample of 57 adults who participated in the Preventing Urinary Stones with Hydration randomized clinical trial, 71 CT scans were analyzed; the kidneys were segmented via deep learning, and each stone was detected and isolated by adaptive, kidney-constrained thresholding. To account for positional differences across patients and standardize kidney alignment, each kidney was reoriented along its principal long axis. A computer-vision pipeline was applied to localize renal calculi, with each stone’s centroid expressed as two normalized coordinates: a superior-inferior (SI) quantile, where 0 marked the inferior pole and 100 the superior pole, and a left-right (LR) quantile spanning the kidney’s width. Six urologists provided ground-truth classifications, with two experts evaluating each scan (33-34 per reviewer) and classifying stones as upper, middle, or lower pole. Inter-rater agreement was assessed using Cohen’s kappa (κ). Results: A random forest classifier achieved 85% accuracy (AUC 0.936, n = 101), with ground-truth comparable to expert raters. Classification of lower-pole versus non-lower-pole stones achieved 92% accuracy (AUC 0.95). Conclusion: A computer vision algorithm can automatically classify the location of kidney stones on CT scans with moderate to high accuracy. This automated approach offers a scalable method for early assessment of stone severity, supporting more efficient treatment decisions. By standardizing stone location reporting, these findings have important implications for both clinical care and research, including clinical trials.