Development of UroSAM: A Machine Learning Model to Automatically Identify Kidney Stone Composition from Endoscopic...

Area of Science:

• Nephrology
• Artificial Intelligence
• Medical Imaging

Background:

• Chemical analysis of kidney stones is crucial for effective prevention counseling.
• Current laser dusting techniques limit material collection for analysis.
• A need exists for non-invasive methods to determine stone composition intraoperatively.

Purpose of the Study:

• To develop and validate a novel machine learning model (UroSAM) for predicting kidney stone composition using intraoperative endoscopic video data.
• To assess the accuracy of UroSAM in identifying and classifying common kidney stone types: calcium oxalate monohydrate (COM), calcium oxalate dihydrate (COD), calcium phosphate (CAP), and uric acid (UA).

Main Methods:

• Two endourologists collected ureteroscopy videos of kidney stones (≥ 10 mm).
• A machine learning model, UroSAM, was developed using the Segment Anything Model (SAM) and a U-Net convolutional neural network (CNN).
• The model was trained on extracted video frames, with stones manually outlined for training data, and then evaluated for segmentation and classification performance.

Main Results:

• UroSAM achieved high segmentation precision (94.77%) and performance metrics (Dice: 0.9135, IoU: 0.8496).
• Initial video-wise classification accuracy was 60%, with 84.4% for COM stones.
• A post hoc adaptive threshold improved overall classification to 62% and enhanced accuracy for COD, CAP, and UA stones.

Conclusions:

• UroSAM effectively identifies kidney stones from endoscopic videos, demonstrating potential for intraoperative composition assessment.
• Further refinement with more diverse, high-quality video data is expected to improve the model's classification accuracy for all stone types.
• This AI-driven approach offers a promising avenue for enhancing kidney stone disease management and prevention strategies.