Evaluation and understanding of automated urinary stone recognition methods

Area of Science:

• Urology
• Medical Imaging
• Artificial Intelligence

Background:

• Urinary stone recognition during endoscopy is challenging.
• Accurate diagnosis often requires specialized expertise or complex analysis.
• Automated methods could improve diagnostic capabilities.

Purpose of the Study:

• To evaluate automated machine learning for recognizing urinary stones via endoscopy.
• To compare shallow classification and deep learning methods for stone identification.

Main Methods:

• Acquired surface and section images of 123 urinary calculi (pure stones).
• Applied shallow classification (texture, color) and deep learning methods.
• Evaluated sensitivity, specificity, and positive predictive value for six stone types.

Main Results:

• Shallow methods achieved high performance (e.g., 91% sensitivity, 90% specificity for whewellite).
• Deep learning methods demonstrated excellent accuracy (e.g., 99% sensitivity, 98% specificity for whewellite).
• Both methods showed high diagnostic values across multiple stone compositions.

Conclusions:

• Artificial intelligence shows significant potential for endoscopic urinary stone recognition.
• Promising results were achieved for pure stones, validating the proof of concept.
• Further research with diverse stone types is necessary to refine these AI methods.