基于YOLOv5的手术器械末端识别研究

Research on the Identification of Surgical Instrument Ends Based on YOLOv5

摘要: 目的：在手术过程中，医生需要实时掌握手术器械的位置。传统机器学习方法与深度学习模型受遮挡等因素影响，其场景适应性和漏检率仍有优化的空间。本研究提出一种基于YOLOv5的手术器械末端识别方法。方法：通过光学相机采集不同背景、光照条件下的手术器械图像，利用Roboflow平台标注器械末端，构建符合条件的数据集进行训练。其次，将数据集按7:2:1比例划分训练集、验证集和测试集，采用精确度、召回率、mAP等指标对比识别器械整体与末端的效果。结果：在器械被遮挡的场景下，末端识别置信度较高于整体识别，目标检测的mAP@0.5为92.7%。结论：在手持器械导致的器械被遮挡情况下，通过识别器械末端辅助医生找到手术器械的位置，为解决手术环境中的遮挡问题提供了初步探索和思路。

Abstract: Objective: During surgery, doctors need to know the position of surgical instruments in real time. Traditional machine learning methods and deep learning models are affected by factors such as occlusion, and their scene adaptability and missed detection rate still need to be optimized. This study proposes a method for identifying the end of surgical instruments based on YOLOv5. Methods: An optical camera was used to collect images of surgical instruments under different background and illumination conditions. The Roboflow platform was used to mark the end of the surgical instruments to construct a qualified dataset for training. Secondly, the data set was divided into training set, validation set and test set according to the ratio of 7:2:1, and the accuracy, recall rate, mAP and other indicators were used to compare the effect of identifying the whole and the end of the device. Results: In the scene where the device was occluded, the confidence of end recognition was higher than that of overall recognition, and the mAP@0.5 of object detection was 92.7%. Conclusions: In the case of instrument occlusion caused by hand-held instruments, it assists the doctor to find the position of the surgical instrument by identifying the end of the instrument, which provides a preliminary exploration and idea for solving the occlusion problem in the surgical environment.