Artificial intelligence model based on CT imaging for predicting infected upper urinary tract calculi

Objective: To construct an artificial intelligence (AI) model based on Computed Tomography (CT) imaging and evaluate its efficacy in preoperatively predicting infected upper urinary tract calculi. Methods: Clinical data from December 2023 to February 2025 for patients diagnosed with urinary tract calculi at the Affiliated Hospital of Hebei University were collected. Postoperative analysis of stone composition defined stones containing more than 25% struvite and/or carbonate apatite as infectious stones, with the remainder being non-infectious stones. Labelimg software was utilized to annotate the stone locations in CT images by manually outlining the stone contours. Stratified random sampling was performed at the patient level to divide the 465 enrolled patients into training, validation, and test sets at a 7:1:2 ratio (326, 47 and 92 patients, respectively), with all CT images of each patient assigned to the corresponding dataset to avoid data overlap. We documented the model's Average Precision (AP), Mean Average Precision (mAP), and Mean Recall (mR). Additionally, CT images from patients diagnosed with urinary tract calculi from December 2021 to February 2023 at our hospital were randomly selected to evaluate the model's clinical efficacy. Results: Of the 465 patients enrolled, 134 were classified in the infectious stone group and 331 in the non-infectious stone group. The model's mAP for infectious stones in the training and validation sets was 95.3% and 95.0%, respectively. The mAP was lower at 62.4% for stones smaller than 32 × 32 pixels, and 81.3% for stones larger than this size. Of the 935 CT images analyzed from December 2021 to February 2023, the RetinaNet model achieved an accuracy of 85.17%, sensitivity of 72.78%, specificity of 93.09%, and positive and negative predictive values of 87.04% and 84.27%, respectively for predicting infectious stones. The kappa test demonstrated significant consistency between the model and infrared spectroscopy analysis (kappa value of 0.679). Conclusion: The RetinaNet model based on CT imaging shows high specificity for predicting infectious upper urinary tract calculi, supporting its clinical value in identifying suspected cases preoperatively. However, its moderate sensitivity precludes reliable standalone ruling-out of infectious stones. When combined with routine laboratory tests (e.g., urine routine and culture), this AI model acts as a valuable complementary preoperative tool, providing auxiliary guidance for treatment strategy formulation and surgical decision-making in patients with urinary tract calculi.