AI in Pediatric Urology: Deep Learning-Based Approach Supporting Posterior Urethral Valves Diagnosis on VCUG Imaging

Posterior Urethral Valves (PUV) are the leading cause of lower urinary tract obstruction in male infants. Diagnosis relies on Voiding CystoUrethroGraphy (VCUG) and invasive cystoscopy, which, despite being the gold standard, poses anesthesia and procedural risks, underscoring the need for a non-invasive, imaging-based screening tool. In this study, we investigate whether recent Deep Learning (DL) architectures can identify PUV directly from routine VCUG images, offering a non-invasive alternative. We present the design and implementation of a novel DL-based framework, trained and validated on a real-world multicenter dataset of 403 VCUG acquisitions, previously annotated by pediatric urologists. In the binary classification task (PUV vs. non-PUV), EfficientNet-B0 achieved the highest overall accuracy of . To promote clinical interpretability, we integrated Grad-CAM to generate attention maps, verifying that the model was indeed focusing on relevant discriminative features. This enhances the reliability of the system and supports its potential clinical adoption. The results demonstrate the feasibility of automated PUV classification from VCUG using DL, laying a solid groundwork for future development toward non-invasive screening tools in real-world clinical applications.