Deep Learning of Videourodynamics to Classify Bladder Dysfunction Severity
Updated: Sep 29
Weaver JK, Martin-olenski M, Logan J, Broms R, Antony M, Van Batavia J, Weiss D, Long CJ, Smith AL, Zderic SA, Yong F, Tasian GE
Introduction: A videourodynamics (VUDS) study is rich in detail, but laborious for an individual to understand and its interpretation has high interobserver variability. We applied deep learning models of VUDS pressure tracings and fluoroscopic images to identify informative features associated with severity of bladder decompensation. We hypothesized that these deep learning features would correctly classify children with preserved bladder function from those with decompensated bladders. Methods: We performed a cross-sectional study of 306 VUDS studies of children with spina bifida (SB) evaluated at our institution from 2019-2021. We excluded children with a history of a bladder augmentation or imperforate anus. The outcome was degree of bladder decompensation, defined as an ordinal variable: mild, moderate, and severe. Degree of bladder decompensation, which was the ground truth for our models, was defined by a panel of five expert reviewers (four pediatric urologists who regularly care for SB patients and an adult urologist fellowship trained in female pelvic medicine and reconstructive surgery). Factors considered in determining bladder dysfunction severity were those that increase the risk of upper tract injury, such as low compliance, detrusor overactivity, high detrusor leak point pressure, or detrusor sphincter dyssynergia (DSD). We built a random forest model to predict severity of bladder decompensation using prospectively collected clinical data (e.g. presence of leak, reflux, DSD, pressure at expected bladder capacity (EBC), bladder shape, percent EBC achieved). We also built a 1-dimensional convolutional neural network of raw data from the volume-pressure recordings and a deep learning imaging model of fluoroscopic images to predict severity of bladder decompensation. An ensemble model was generated by averaging the risk probabilities of these two deep learning models.