Deep Learning for the Study of Urinary Stone Composition from Computed Tomography Images.

Abstract

Urinary stones composed of uric acid can be treated with medicine. Computed tomography (CT) can diagnose urinary stone disease, but it is difficult to predict the type of uric stones. This study aims to develop a method to distinguish pure uric acid (UA) stones from non-uric acid (non-UA) stones by describing quantitative CT parameters of single-energy slices of urinary stones related to chemical stone types. Clinical data, CT images, and stone composition analysis results of patients with urinary stones clinically diagnosed at The Department of Urology, Affiliated Hospital of Qingdao University between 1 January 2018 and 31 December 2020 were collected and retrospectively analyzed. The above data were preprocessed and fed into a convolutional neural network to perform deep learning (DL) of the model, and the dataset was validated at a ratio of 4:1. The area under the curve (AUC) value of the receiver operating characteristic (ROC) curve and the confusion matrix were utilized to evaluate the predictive effect of the model. A retrospective analysis of 918 non-enhanced thin-slice single-energy CT images of known chemical stone types (124 with UA stones and 794 with non-UA stones) was conducted using a DL model. Compared with the results of ex vivo analysis by infrared spectroscopy, the prediction model obtained an AUC of 0.83 for the dichotomous classification of UA stones and non-UA stones. The accuracy of the model was 97.01%, with an F1 score of 89.04%, sensitivity of 84.62%, and specificity of 82.28%. This DL model constructed based on convolutional neural network analysis of thin-slice single-energy CT images is highly accurate in predicting the composition of pure UA and non-UA stones, providing a simple and rapid diagnosis method.