Predicting stone composition via machine-learning models trained on intra-operative endoscopic digital images

Area of Science:

• Urology
• Medical Imaging
• Artificial Intelligence

Background:

• Urinary stone composition analysis is crucial for effective treatment planning.
• Current methods for intraoperative stone identification can be time-consuming and less accurate.
• Real-time prediction of stone composition can optimize laser lithotripsy parameters.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) model for predicting urinary stone composition using intraoperative endoscopic images.
• To assess the potential of DL in real-time stone analysis to guide surgical interventions.
• To improve surgical outcomes by enabling timely adjustments to laser settings.

Main Methods:

• A dataset of 1658 intraoperative images from 490 patients undergoing holmium laser surgery was collected.
• Images were analyzed using a deep convolutional neural network (CNN), specifically ResNet-101, for multiclass classification.
• The model was trained to identify eight common stone categories, including single and mixed compositions.

Main Results:

• The DL model achieved high prediction rates for individual stone components, with accuracy reaching 100% for several categories.
• Calcium oxalate monohydrate (99%), anhydrous uric acid (98%), and mixed stones (100%) were accurately identified.
• The overall weighted recall for component analysis across the entire cohort was 99%.

Conclusions:

• Deep learning demonstrates a promising capability for accurate and rapid identification of urinary stone components from intraoperative images.
• DL surpasses human visual identification in speed and accuracy for discriminating single and mixed stone types.
• This AI-driven approach can optimize holmium laser parameters in real-time, potentially reducing operative time, improving surgical efficiency, and minimizing postoperative complications.