AI-Supported Prediction of Vesicoureteral Reflux in Children Based on Cystoscopic Configuration of the Ureteric...

Background: Vesicoureteral reflux (VUR) is a common pediatric urological disorder traditionally diagnosed by voiding cystourethrography (VCUG). Although ureteric orifice (UO) morphology during cystoscopy correlates with VUR severity, its assessment remains subjective. To our knowledge, artificial intelligence (AI) has not previously been applied to cystoscopic images for VUR prediction. We aimed to develop and evaluate the first AI model for predicting VUR from pediatric cystoscopic images and videos. Methods: In this retrospective single-center study, cystoscopic videos from children with and without VUR diagnosed by VCUG were analyzed. Individual frames were extracted, anonymized and annotated according to VUR grade. Four YOLOv12 object detection models were trained to classify UOs as no/grade I, low-grade (grades II-III), or high-grade (grades IV-V) VUR. To better understand sources of model failure, additional experiments evaluated binary image classification and class-agnostic UO localization. Finally, frame-level predictions were aggregated across complete cystoscopic sequences using majority voting to assess UO-level performance. Results: Three-class object detection demonstrated limited frame-level performance, with a maximum mAP@50 of 0.37 and mAP@50-95 of 0.17. Binary classification achieved a macro precision of 0.62, recall of 0.58, and F1 score of 0.49. Class-agnostic object detection improved localization performance (mAP@50 0.63). Aggregating predictions across complete video sequences substantially improved diagnostic performance, achieving UO-level accuracies of up to 77%, with leave-one-out cross-validation accuracies ranging from 0.69 to 0.77. Conclusions: This proof-of-concept study demonstrates the feasibility of AI-assisted VUR prediction from pediatric cystoscopic videos. While frame-level performance was limited, sequence-level aggregation markedly improved diagnostic accuracy, highlighting the importance of temporal information for future video-based AI models in pediatric endourology.