Deep learning-based automatic segmentation and quantitative analysis of upper urinary tract stones on non-enhanced CT in stone-positive patients

To construct a deep learning-based automatic segmentation model for upper urinary tract stones on non-enhanced CT, and to evaluate its segmentation performance, automatic quantitative consistency, and analysis efficiency in an independent test set. A total of 620 patients with upper urinary tract stones were prospectively and consecutively enrolled, including 934 lesions. They were divided at the patient level into a training set of 400 cases, a validation set of 100 cases, and an independent test set of 120 cases. Two radiologists performed independent annotations, and a manual reference standard was established after expert arbitration. A three-dimensional U-Net model was used for automatic segmentation, and segmentation performance, automatic quantitative consistency, and analysis time were evaluated in the independent test set. In the independent test set, the patient-level Dice coefficient of the model was 0.87 ± 0.13, the IoU was 0.79 ± 0.15, and the lesion-level sensitivity, precision, and F1 value were 0.931, 0.920, and 0.925, respectively. The HD95 was 1.63 ± 0.21 mm. The ICC(2,1) values for patient-level total stone volume, maximum stone diameter, and number of stones were 0.964, 0.951, and 0.982, respectively, and the complete agreement rate for the number of stones was 93.33%. The automatic analysis time was significantly shorter than the manual analysis time, which were 0.37 ± 0.04 min/case and 13.84 ± 1.86 min/case, respectively. The time-saving proportion was 97.33 ± 0.31% ( P < 0.001). In stone-positive patients under standardized imaging conditions at a single center, the present findings support the feasibility of deep learning-based automatic segmentation and quantitative analysis of upper urinary tract stones on non-enhanced CT. The quantitative results showed high agreement with the manual reference standard, and the analysis time was shorter than that required for manual slice-by-slice segmentation and quantitative calculation.