Urinary incontinence (UI) in women is heterogeneous and poorly phenotyped, which may lead to poor treatment outcomes. Current diagnostic approaches can potentially be improved using advanced imaging techniques, such as Magnetic resonance imaging (MRI) alongside machine learning (ML) models [1]. This study aimed to evaluate whether quantitative MRI-based pelvic morphological features can differentiate continent and incontinent women using ML. To our knowledge, this is among the first studies to investigate quantitative MRI-derived pelvic morphological features as potential imaging biomarkers for objective identification of continence status in women.

We retrospectively reviewed the MRI studies in the Radiology department and selected the female patients who had a lower abdominal MRI and had been treated for either stress UI, urgency or mixed UI in our urology and gynaecology clinics. A total of 52 women were included. Continence status for each patient was determined based on a clinical diagnosis that involves history, symptom assessment, physical exam and a residual urine check with or without a urodynamic study. The continent group included patients without any symptoms or signs of UI.  Key pelvic anatomical structures, including the bladder, urethra, vagina, rectum, pubic bone, obturator internus muscle, and puborectalis muscle, were manually segmented by a urogyneacologist using ITK-SNAP software [2] to allow extraction of quantitative morphological features for downstream modelling.

We developed an anatomically guided automated method to select the most relevant MRI slice and extract quantitative pelvic measurements. Bladder shape features were computed from the slice with maximal bladder area, while multi-structure spatial and geometric features were derived from the slice with maximal urethral area. Extracted features were statistically compared between continent and incontinent groups and used to train and benchmark supervised ML models, including logistic regression, support vector machines (SVM), k-nearest neighbours, random forest, multilayer perceptron, and XGBoost, using nested stratified cross-validation. Performance was evaluated using well-established metrics including ROC-AUC, sensitivity, specificity, accuracy, and F1-score.

Additionally, a Statistical Shape Model (SSM) was developed to characterise geometric variability between continent and incontinent groups. For urethra–pubic bone analysis, the reference slice was defined as the middle slice among those where both structures were present (Figure 2A-B), while for urethra–vagina analysis the slice with maximal urethral area was selected (Figure 2C-D). Contours were extracted, aligned, and used to derive distance-based features.

ML models using bladder shape features distinguished incontinent from continent patients, with logistic regression achieving the highest ROC-AUC (0.768 ± 0.076) and SVM demonstrating the highest sensitivity (0.773 ± 0.352). Explainability analysis (SHAP) identified bladder solidity and eccentricity as key features contributing to continence status prediction. Models trained on combined pelvic structural features showed improved performance (Figure 1F), with Random Forest achieving the highest ROC-AUC (0.922 ± 0.077). 

Bladder shape analysis showed significantly lower eccentricity and aspect ratio in incontinent patients (both p = 0.0012) compared with continent women, together with higher solidity (p = 0.030). In addition, SSM of urethra–pubic bone and urethra–vagina relationships demonstrated distinct modes of urethral spatial variation relative to surrounding pelvic structures. SSM mode scores showed relationship with distance measures between continent and incontinent groups, indicating that principal modes captured meaningful spatial variation in pelvic anatomy.

Findings suggest that ML models showed improved performance when multi-structure spatial features were incorporated alongside bladder features, suggesting that pelvic anatomical relationships provide additional information for identifying continence status.
Associations between SSM mode scores and inter-structural distances suggest that variations in pelvic spatial relationships may contribute to continence status.

Although pelvic MRI is not routinely used in UI assessment, our findings suggest that quantitative MRI-derived pelvic morphological features may provide objective markers of continence status and support the development of future predictive and translational modelling approaches.

Wang, Q., Wang, X., Jiang, X. and Lin, C., 2024. Machine learning in female urinary incontinence: a scoping review. Digital Health, 10, p.20552076241281450.
Yushkevich, P.A., Piven, J., Hazlett, H.C., Smith, R.G., Ho, S., Gee, J.C. and Gerig, G., 2006. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage, 31(3), pp.1116-1128. www.itksnap.org

Continence 19S (2026) 102801
DOI: 10.1016/j.cont.2026.102801