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

Area of Science:

• Radiology and Medical Imaging
• Machine Learning in Healthcare
• Urology and Trauma Surgery

Background:

• Pelvic fracture urethral injury (PFUI) diagnosis is critical but traditionally relies on subjective clinical evaluation and invasive imaging, leading to potential delays.
• There is a need for more accurate and efficient diagnostic methods for PFUI to improve patient outcomes.
• Machine learning and radiomics offer promising avenues for objective and rapid analysis of medical imaging data.

Purpose of the Study:

• To develop and validate a machine learning model utilizing radiomic features from pelvic CT scans for the diagnosis of PFUI.
• To assess the diagnostic performance of a radiomic nomogram and a combined model incorporating clinical variables.
• To evaluate the potential of this noninvasive approach for improving the accuracy and efficiency of PFUI diagnosis in male patients.

Main Methods:

• Retrospective analysis of CT imaging data from 205 male patients with pelvic fractures (100 PFUI, 105 non-PFUI).
• Segmentation of pelvic bony structures and interactive correction of fracture areas using TotalSegmentator and 3D-Slicer.
• Extraction of radiomic features (texture, shape, wavelet transformations) using PyRadiomics, followed by feature selection and model development (radiomic nomogram, combined model).

Main Results:

• The radiomic nomogram showed excellent performance (C-index = 0.85) in the validation cohort.
• The combined model integrating clinical variables and radiomic features achieved the highest generalization capability (validation AUC: 0.94, accuracy: 91.97%, specificity: 90.24%).
• The Nomo-score was significantly higher in the PFUI group, and decision curve analysis confirmed clinical utility.

Conclusions:

• A machine learning model based on pelvic CT radiomics can effectively predict the risk of PFUI in male patients with pelvic fractures.
• The combined model incorporating clinical variables and radiomic features represents the optimal approach for clinical implementation.
• This noninvasive, accurate model has the potential to be integrated into emergency imaging workflows for faster and better treatment decisions.