| 摘 要: 针对外科手术器械人工清点效率低、相似器械易出错及器械重叠等问题,提出一种引入NMS模块的改进YOLO双系统识别与计数方案。系统由PyQT端与Web端组成,分别面向手术台实时识别与计数与医院内普通场景。PyQT端采用固定机位实现实时无损传输与GPU推理,支持多版本YOLO模型部署；Web端基于Flask框架提供HTTPS加密与移动端访问,实现前后端交互与结果可视化。通过医院现场采集的404张图像并经扩增平衡处理,构建包含19类器械的4267张数据集,在YOLOv5m、YOLOv8m与YOLO11m模型的mAP50与mAP50-95均超过90%。系统可在手术全流程辅助器械清点,具有较高识别精度与实时性,为手术器械管理提供有效技术支撑,具备临床应用价值。 |
| 关键词: 手术器械清点 YOLO目标检测 实时识别 双系统架构 PyQT |
|
中图分类号:
文献标识码:
|
| 基金项目: 2024年广东省普通高校青年创新人才类项目（2024WQNCX104） |
|
| Research on Recognition and Counting of Surgical Instruments Based on Improved YOLO |
|
XI Juanxia, Guo Yiping, HUANG WeiJie
|
Neusoft Institute, Guangdong
|
| Abstract: To address the problems of low efficiency, high error rate and instrument overlapping in manual counting of surgical instruments, an improved YOLO dual-system recognition and counting scheme incorporating an NMS module is proposed. The system consists of a PyQT side and a Web side, which are designed for real-time recognition and counting on the operating table and general scenarios in hospitals respectively. The PyQT side adopts a fixed camera position to achieve real-time lossless transmission and GPU inference, supporting the deployment of multiple versions of YOLO models. The Web side, based on the Flask framework, provides HTTPS encryption and mobile access, enabling front-end and back-end interaction and result visualization. A dataset of 4,267 images covering 19 categories of instruments is constructed from 404 images collected on-site in hospitals and processed by augmentation and balancing. The mAP50 and mAP50-95 values all exceed 90% for the YOLOv5m, YOLOv8m and YOLO11m models. The system can assist in instrument counting throughout the entire surgical procedure, featuring high recognition accuracy and real-time performance. It provides effective technical support for surgical instrument management and possesses clinical application value. |
| Keywords: surgical instrument counting YOLO object detection real-time recognition dual-system architecture PyQT |
