Deep-Learning Segmentation of Urinary Stones in Noncontrast Computed Tomography

Area of Science:

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

Background:

• Noncontrast CT (NCCT) is labor-intensive for identifying urinary tract stones (urolithiasis).
• Existing deep learning algorithms still produce false positives in urolithiasis detection.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated urolithiasis detection and quantification.
• To improve diagnostic accuracy and reduce physician workload in imaging urolithiasis.

Main Methods:

• Developed a deep learning model combining urolithiasis segmentation and per-slice classification using 410 NCCT scans.
• Trained axial, coronal, and sagittal prediction models, introducing an additive model for improved sensitivity.
• Evaluated automated quantification of clinical metrics (size, volume, density, etc.) in 3D stone models.

Main Results:

• The axial model achieved 88.92% detection rate and 87.56% Dice similarity coefficient for segmentation.
• High sensitivity (95.10%) for stones >5 mm; additive model improved overall sensitivity to 90.97% for smaller stones.
• Reduced false positives to 0.34 per patient; clinical metrics correlated strongly (R² > 0.964) with automated measurements.

Conclusions:

• The proposed deep learning system significantly reduces the burden of imaging diagnosis for urolithiasis.
• Automated quantification of clinical metrics provides high accuracy and reproducibility, aiding treatment strategy determination.