Spontaneous stone passage prediction in acute ureteric colic using radiomics and machine learning
Frederic Panthier, Zine‐Eddine Khene, Daniela Velinova, Laurent Berthe, Hugh Crawford‐Smith, Sian Allen, Olivier Traxer, Clare Allen, Daron SmithObjectives
To predict spontaneous stone passage (SSP) in uncomplicated acute ureteric colic (AC) using non‐contrast computed tomography (NCCT)‐based radiomics and machine learning (ML).
Patients and Methods
This retrospective single‐centre study included consecutive patients presenting with stone‐related uncomplicated AC between January 2022 and March 2024, confirmed by NCCT. Initial and 4‐week follow‐up clinical and imaging data were collected, including SSP status. After anonymisation, a semi‐automated density‐based stone segmentation was performed using 3DSlicer. Radiomic features were extracted (PyRadiomics_v3.1.1). Significant features were selected using the Least Absolute Shrinkage and Selection Operator (LASSO) method. After data splitting (training [80%]–validation [20%]), nine ML models were trained with hyperparameter fine‐tuning to predict SSP on the validation set. The primary performance metric was the area under the curve (AUC).
Results
A total of 428 patients were included, with a median (interquartile range [IQR]) age of 45 (32–56) years. Most ureteric stones were solitary, with a median (IQR) maximum stone diameter (MSD) of 5.3 (4.1–6) mm and stone volume of 68.8 (22–78.1) mm 3 . The majority of the stones were proximal rather than distal: 59.6%, were located in the upper ureter, 9.1%, in the middle, and 30.8% in the lower ureter. A pelvic phlebolith was present in 50.9% of cases. The overall SSP rate was 47.9%. Following LASSO selection, seven radiomic features were used for model training and validation: image‐original_mean (mean value of all voxels), image‐original_maximum (maximum voxel value), original_glrlm_RunEntropy (heterogeneity in grey level run lengths), original_gldm_DependenceVariance (structure irregularity), original_gldm_LargeDependenceEmphasis (voxel homogeneity), original_gldm_LargeDependenceHighGreyLevelEmphasis (homogenous regions with high voxel intensity), original_gldm_LargeDependenceLowGreyLevelEmphasis (homogenous regions with low voxel intensity), along with SSP status. The random forest algorithm achieved the best performance of all tested models with an AUC of 0.79 (95% confidence interval 0.68–0.88) on the validation set (accuracy = 0.72, F1‐score = 0.72, sensitivity = 0.69, specificity = 0.76).
Conclusion
Predicting SSP in uncomplicated AC is feasible using radiomics features extracted from NCCT. Hybrid models, including the stone location could improve their performance. External validation is required, and developing an automated ureteric stone detection algorithm would be necessary to create a fully stand‐alone clinical tool.