Two novel deep-learning models to predict spontaneous ureteral calculi passage: Model development and validation

Abstract

Objective The aim of this study was to develop and evaluate 2 deep-learning (DL) models for predicting spontaneous ureteral stone passage (SSP). Materials and methods A total of 1217 patients with thin-layer computed tomography–confirmed ureteral stones in our hospital from January 2019 to December 2022 were retrospectively examined. These patients were grouped into 3 data sets: the training set (n = 1000), the validation set (n = 100), and the test set (n = 117). Two DL models based on residual neural network (ResNet)—2-dimensional (2D) ResNet29 and 3-dimensional (3D) ResNet29—were separately developed, trained, and assessed. The predictive ability of a conventional approach using a stone diameter of <5 mm on computed tomography was investigated, and the results were compared with those of the 2 DL models. Results Of the 1217 patients, SSP was reported in 446 (36.6%). The total accuracy, sensitivity, and specificity were 76.9%, 56.1%, and 90.8% for the stone diameter approach; 87.1%, 84.2%, and 92.7% for the 2D ResNet29 model; and 90.6%, 88.2%, and 95.1% for the 3D ResNet29 model, respectively. Both the 2D and 3D ResNet29 models showed significantly higher accuracy than the stone diameter approach. Receiver operating characteristic curve analysis showed that both DL models had a significantly higher area under the curve than the stone diameter–based classification. Conclusions The DL models, particularly the 3D model, are novel and effective methods for predicting SSP rates. Using such models may help determine whether a patient should receive surgical intervention or expect a long interval before stone passage.