UO-YOLO: Ureteral Orifice Detection Network Based on YOLO and Biformer Attention Mechanism

• Published in: Applied sciences Vol. 14; no. 12; p. 5124
• Main Authors: Liang, Li, Yuanjun, Wang
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.06.2024
• Subjects: Accuracy; Algorithms; Classification; Computational linguistics; Deep learning; Detectors; Language processing; Natural language interfaces; Neural networks; object detection; Polyps; Prostate cancer; Sensors; Surgery; ureteral orifice detection; Urinary tract diseases; Urogenital system; Urology
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app14125124

Background and Purpose: In urological surgery, accurate localization of the ureteral orifice is crucial for procedures such as ureteral stent insertion, assessment of ureteral orifice lesions, and prostate tumor resection. Consequently, we have developed and validated a computer-assisted ureteral orifice detection system that combines the YOLO deep convolutional neural network and the attention mechanism. Data: The cases were partitioned into a training set and a validation set at a 4:1 ratio, with 84 cases comprising 820 images in the training set and 20 cases containing 223 images in the validation set. Method: We improved the YOLO network structure to accomplish the detection task. Based on the one-stage strategy, we replaced the backbone of YOLOv5 with a structure composed of ConvNeXt blocks. Additionally, we introduced GRN (Global Response Normalization) modules and SE blocks into the blocks to enhance deep feature diversity. In the feature enhancement section, we incorporated the BiFormer attention structure, which provides long-distance context dependencies without adding excessive computational costs. Finally, we improved the prediction box loss function to WIoU (Wise-IoU), enhancing the accuracy of the prediction boxes. Results: Testing on 223 cystoscopy images demonstrated a precision of 0.928 and recall of 0.756 for our proposed ureteral orifice detection network. With an overlap threshold of 0.5, the mAP of our proposed image detection system reached 0.896. The entire model achieved a single-frame detection speed of 5.7 ms on the platform, with a frame rate of 175FPS. Conclusion: We have enhanced a deep learning framework based on the one-stage YOLO strategy, suitable for real-time detection of the ureteral orifice in endoscopic scenarios. The system simultaneously maintains high accuracy and good real-time performance. This method holds substantial potential as an excellent learning and feedback system for trainees and new urologists in clinical settings.