Development and validation of a real-time AI model for ureteral orifice recognition during cystoscopy for fluorescence ureteral navigation

Fluorescence ureteral navigation (FUN) can improve ureteral visualization during minimally invasive pelvic surgery, but its adoption is limited by the need for cystoscopic ureteral catheter placement by urologists. We developed and validated an artificial intelligence (AI) model for real-time ureteral orifice recognition during cystoscopy to support FUN. In this single-center retrospective study, 79 cystoscopy videos recorded during ureteral stent or catheter placement were analyzed. Ground truth was established using an anchor-frame reverse playback method. Videos were divided into training, validation, development, and test sets, and a urologist baseline was determined using 200 still images annotated by 10 urologists. Success was defined by the center-to-center distance between the predicted point and ground truth, normalized to image height, with a threshold of ≤ 15%. The urologist reference baseline showed an F1 score of 0.64 ± 0.08 and an accuracy of 0.70 ± 0.03. The AI model achieved an F1 score of 0.71 ± 0.19, an accuracy of 0.81 ± 0.11, and real-time inference at 30 fps with 0.033–0.034 s latency per frame. This study provides a clinically oriented evaluation framework for AI-assisted FUN support.