Development and multi-institutional validation of the Swin transformer model for prediction of spontaneous passage of ureteral calculi: A retrospective multicentric study.

To improve the accuracy of our prediction of the spontaneous passage of urinary calculi, we developed a model based on the Swin Transformer model. In this retrospective multicentre study, computed tomography (CT) images of ureteral calculi from January 2019 to December 2023 from our hospital were used as internal training and validation datasets. The samples were split in a 7:3 ratio into an internal training set (N = 278) and internal testing set (N = 120). An external validation set (N = 142) was created using the CT datasets of patients from four different hospitals. Swin Transformer was used to create an ensemble learning-based model approach for predicting the passage of ureteral calculi. For both the internal and external validation sets, performance was assessed using the F1 score, accuracy, specificity, sensitivity, and area under the operating characteristic curve (AUC). A total of 540 urinary CT images were acquired (Internal dataset = 398; multicentre external dataset = 142). The AUC reached 0.873 (95% [CI]: 0.841-0.906), and the external validation set reached 0.784 (95% CI: 0.747-0.820),. The evaluation results of the Swin Transformer model, developed using CT images of urinary calculi, demonstrate that it outperforms models such as ResNet and ConvNeXt. Furthermore, the model outperforms both attending and resident doctors. The predictive power of attending doctors was significantly improved with the aid of the Swin Transformer model. The Swin Transformer model is a generalizable, objective and accurate prediction model for calculi passage based on urinary tract CT images and could help clinicians make better clinical judgements. Trial registration: It was also registered in the Chinese Clinical Trial Registation (registration number: ChiCTR2400086807). Informed consent was waived as this study used retrospectively collected anonymized data.