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

Area of Science:

• Radiology and Medical Imaging
• Urology
• Artificial Intelligence in Medicine

Background:

• Urinary stones, particularly uric acid (UA) stones, require specific medical treatments.
• Computed tomography (CT) is crucial for diagnosing urinary stone disease.
• Predicting the chemical composition of urinary stones, specifically distinguishing UA from non-UA stones, remains challenging with conventional CT analysis.

Purpose of the Study:

• To develop and validate a method for distinguishing pure uric acid (UA) stones from non-uric acid (non-UA) stones.
• To utilize quantitative CT parameters from single-energy slices for chemical stone type prediction.
• To assess the efficacy of a deep learning model in classifying urinary stone composition.

Main Methods:

• Retrospective analysis of 918 non-enhanced thin-slice single-energy CT images and stone composition data.
• Application of a deep learning (DL) model based on a convolutional neural network (CNN).
• Validation of the DL model using receiver operating characteristic (ROC) curve analysis (Area Under the Curve - AUC) and confusion matrix.

Main Results:

• The DL model achieved an AUC of 0.83 for classifying UA versus non-UA stones.
• The model demonstrated high predictive performance with an accuracy of 97.01%, F1 score of 89.04%, sensitivity of 84.62%, and specificity of 82.28%.
• The model's predictions were compared against ex vivo infrared spectroscopy analysis.

Conclusions:

• A deep learning model utilizing convolutional neural network analysis of CT images can accurately predict pure UA and non-UA stone composition.
• This approach provides a simple, rapid, and highly accurate diagnostic method for urinary stone characterization.
• The developed model holds significant potential for improving the clinical management of urinary stone disease.