Enhancing renal stone detection through artificial intelligence models

Objective Traditional ultrasonic detection for kidney stones is prone to variability due to the experience and skill of the operating physician, leading to inconsistent diagnostic outcomes. Factors such as stone size, location, and surrounding tissues can further compromise accuracy. Artificial intelligence (AI), with its advanced data processing and deep learning capabilities, offers a solution by automating the analysis of medical images to improve diagnostic accuracy and efficiency. Methods This cross-sectional study was conducted at our hospital between 01/06/2022 and 03/07/2024. After excluding 50 poor-quality ultrasound images, the final dataset included 1,130 images: 918 stone images and 212 non-stone images. At the patient level, 791 stone images were used for training/validation, and 127 stone images plus 212 non-stone images ( n = 339) were held out as an independent test cohort. The High-Resolution Network (HRNet) was employed for model training to detect kidney stones in ultrasound images. Detailed image annotations were created using the labelme tool and converted to COCO format for compatibility with AI algorithms. Results In the independent test cohort ( n = 339), the confusion matrix was: TN = 200, FP = 12, FN = 34, and TP = 93, corresponding to 212 actual negatives and 127 actual positives. The HRNet model achieved an accuracy of 86.4% (95% CI 82.4–89.7%), sensitivity of 73.2% (95% CI 64.9–82.2%), specificity of 94.3% (95% CI 90.4–96.7%), positive predictive value (PPV) of 88.6% (95% CI 81.1–93.3%), and negative predictive value (NPV) of 85.5% (95% CI 80.4–89.4%). The model showed high specificity but moderate sensitivity, missing approximately 26.8% of renal stones in the test cohort. Conclusions HRNet-based AI may improve specificity and reduce false-positive findings in ultrasound-based renal stone detection. However, the moderate sensitivity indicates a substantial risk of missed stones, and the model requires further optimization before clinical adoption. This retrospective study evaluated image-level performance only and did not directly measure patient outcomes, clinical workload, diagnostic time, or treatment decisions. Therefore, claims that this approach reduces diagnostic burden or improves patient outcomes are not supported by the present data.