Background: Noncontrast CT (NCCT) relies on labor-intensive examinations of CT slices to identify urolithiasis in the urinary tract, and, despite the use of deep-learning algorithms, false positives remain. Materials and Methods: A total of 410 NCCT axial scans from patients undergoing surgical treatment for urolithiasis were used for model development. The deep learning model was customized to combine a urolithiasis segmentation with per-slice classification for screening. Prediction models of the axial, coronal, and sagittal views were trained, and an additive model with an intersection of the coronal and sagittal predictions added to the axial outcome was introduced. Automated quantification of clinical metrics was evaluated in three-dimensional models of urinary stones. Results: The axial model detected 88.92% of urinary stones and produced a dice similarity coefficient of 87.56% in the urolithiasis segmentation. For urolithiasis (>5 mm), the sensitivity of the axial model reached 95.10%. False positives were reduced to 0.34 per patient using an ensemble of individual models. The additive model improved the sensitivity to 90.97% by detecting more small urolithiasis (<5 mm). All clinical metrics of size, long-axis diameter, volume, mean stone density, stone heterogeneity index, and skin-to-stone distance showed a strong correlation of R2 > 0.964. Conclusions: The proposed system could reduce the burden on the physician for imaging diagnosis and help determine treatment strategies for urinary stones through automated quantification of clinical metrics with high accuracy and reproducibility.
Keywords: deep learning; multiview ensemble; segmentation; urinary stone.