Computer-aided diagnosis of urodynamic stress incontinence with vector-based perineal ultrasound using neural networks.

OBJECTIVES

Adequately defining the anatomical and functional defects in women with stress urinary incontinence prior to surgical repair is essential. Differences in dynamic changes of the bladder neck on perineal ultrasound have been shown to be effective in differentiating between women with and those without urodynamic stress incontinence (USI). This study evaluated a system to diagnose USI with computer-aided vector-based perineal ultrasound.

METHODS

We evaluated 48 women with lower urinary tract symptoms, of whom 36 were found to have USI, by urodynamic study and computer-aided vector-based perineal ultrasound. The function and morphology of the lower urinary tract were used as features in a multilayer perception (MLP) neural network classifier to distinguish between women with and those without USI. The k-fold cross-validation method was used to estimate the performance of the proposed computer-aided diagnosis (CAD) system.

RESULTS

The accuracy of the proposed CAD system for classifying USI was 91.7% (44/48), the sensitivity was 94.4% (34/36), the specificity was 83.3% (10/12), the positive predictive value was 94.4% (34/36) and the negative predictive value was 83.3% (10/12). The receiver-operating characteristics area index for the proposed CAD system was 0.941 +/- 0.034. The proposed CAD system differentiated USI effectively with perineal ultrasound.

CONCLUSION

Computer-aided vector-based perineal ultrasound is valuable in assessing anatomical changes of the bladder neck. Women with and those without USI can be distinguished using perineal sonographic analysis with the proposed CAD system. Because the MLP model is trainable, it could be optimized if a larger set of perineal ultrasound images and other useful features were supplied.