Machine Learning-based Radiomic Model for Early Diagnosis of Male Urethral Injury in Pelvic Fracture Patients.

BACKGROUND AND OBJECTIVE: Pelvic fracture urethral injury (PFUI) is serious and requires prompt diagnosis. Traditional diagnostic methods, which rely on clinical evaluation and invasive imaging, are subjective and can cause delays. This study aims to utilize machine learning and radiomics to improve the accuracy and efficiency of PFUI diagnosis. METHODS: A retrospective analysis was conducted using computed tomography (CT) imaging data from male patients with pelvic fractures, including 100 cases in the PFUI group and 105 cases in the non-PFUI group. First, the key pelvic bony structures such as the iliac and pubic bones were segmented using the TotalSegmentator tool. Then, two physicians performed interactive correction of fracture areas using the 3D-Slicer tool. Radiomic features, including multidimensional data such as texture, shape, and wavelet transformations, were extracted via PyRadiomics. Core biomarkers were identified using three machine learning-based feature selection methods: random forest (using the Boruta algorithm), least absolute shrinkage and selection operator regression, and support vector machine recursive feature elimination. These consensus features were used to develop a radiomic nomogram (Nomo-score) prediction model. Additionally, clinical variables were combined with these key radiomic features to evaluate the changes in model performance. KEY FINDINGS AND LIMITATIONS: Patients in the PFUI group were significantly more likely to present with bilateral pubic ramus fractures (78% vs 61%, p = 0.013) and Tile B2 type fractures (15% vs 1.9%, p = 0.002) than those in the non-PFUI group. The combination of multiple models identified six core wavelet domain features, and the constructed nomogram demonstrated excellent performance in the validation cohort (C-index = 0.85). The Nomo-score of the nomogram was significantly higher in the PFUI group than in the non-PFUI group (p < 0.05 in both the training and the validation set), with a decision curve analysis confirming its clinical advantage. After combining clinical variables with radiomic features, the combined model demonstrated the best generalization capability (validation area under the curve: 0.94, 95% confidence interval: 0.86-1.00, accuracy: 91.97%, specificity: 90.24%). Indeed, the research was conducted as a single-center retrospective analysis, and the lack of multicenter external validation may restrict the generalizability of the model. CONCLUSIONS AND CLINICAL IMPLICATIONS: The machine learning model based on pelvic CT radiomics can predict the risk of PFUI. The combined model is the optimal approach for implementation. In the future, it can be integrated into emergency imaging workflows. PATIENT SUMMARY: This study developed a machine learning model using pelvic computed tomography images to diagnose urethral injury in men with pelvic fractures early. The model is noninvasive and accurate, which may help doctors make faster and better treatment decisions.