Objective: To enhance the information modeling capacity of large-kernel convolutional neural networks and to build a ureteral orifice detection framework for ureteroscopic imaging. Methods: A retrospective dataset of ureteroscopic images from 222 patients was collected. The patients were randomly divided into training and testing sets at a ratio of 7:3. Initially, video files were converted into image frames, and feature-relevant images were manually labeled by physicians. Subsequently, a ConvNeXt-based backbone augmented with squeeze-and-excitation (SE) modules was employed to extract diverse deep features. SCConv modules were incorporated across stages to strengthen the network’s feature extraction performance. Lastly, enhanced spatial excitation attention mechanisms were cascaded to achieve superior feature fusion and detection accuracy. Comparative experiments were conducted against baseline models, including ConvNeXt, assessing accuracy, computational overhead, and inference latency. Results: On a test set of 491 ureteroscopic images, all models achieved mAP@50 values above 0.75, whereas the proposed network achieved 0.890, markedly exceeding baseline performance. The model operated at 20 ms per frame, achieving a frame rate of 50 FPS. Conclusions: We developed an improved deep learning framework based on large-kernel convolutional networks for real-time ureteral orifice detection in endoscopic scenarios. This system achieves a favorable balance between detection accuracy and real-time efficiency. The method demonstrates significant potential as a training and feedback tool for residents and junior urologists in clinical environments.
Li et al. (2026) studied this question.