Research On Real-time Pose Detection Of Surgical Instruments Based On Deep Learning

Pose detection of surgical instruments is an indispensable technology in robot assisted surgery,which enable doctors to obtain more comprehensive visual information and further improve the safety of surgery.Compared with conventional detection of surgical instruments,pose detection of surgical instruments need to delineate the contours of instruments.Therefore,the results of detection are more easily interfered by environmental factors,such as instrument occlusion,instrument shaking,displayed incompletely,insufficient light or reflection,and it has a high requirement for real-time and accuracy performance.To overcome those issues we proposed a high-precision and real-time pose detection method for surgical instruments.The details are as follows:(1)The development of object detection and object segmentation methods at home and abroad and their applications in detection of surgical instruments are described comprehensively by consulting information.We analyzed and compared the performance of surgical instruments’ detection and segmentation algorithms based on deep learning.We point out the problems in the current algorithms,and propose the main research content of this topic.(2)We introduced development of deep learning,the theory of convolutional neural network and some commonly used deep learning frameworks.The main evaluation criteria of object detection and semantic segmentation are described.The main research methods and technical routes of this topic are proposed.The m2cai16 tool locations public dataset is analyzed in detail,and the construction process of the pose detection of surgical instruments dataset is described in detail.Finally,4-Mosaic method and Albumations image enhancement library are used to enhance the dataset.(3)A pose detection algorithm for surgical instruments based on deep learning is proposed.Based on YOLOv5 algorithm,this paper analyzes it,aiming at its low detection accuracy and low real-time performance for surgical instruments.The GAM attention mechanism is introduced into the original backbone network of YOLOv5.It can make the network can focus on key information,thus further improving the feature extraction ability.In Bottennck CSP module,Si LU activation function is used to avoid gradient explosion and alleviate unstable training.The SIo U loss function is used in the prediction part.By introducing vector angle relationship between the true box and the prediction box,the degree of freedom of the regression is reduced,the network convergence is accelerated,and the accuracy of the regression is further improved.Finally,combined with PSPNet semantic segmentation head,the pose detection of surgical instruments is realized by paralleling detection and segmentation.The algorithm is trained and verified on the m2cai16 tool locations public dataset,and compared with other algorithms,the m AP and FPS achieve 97.9% and 133frames/second respectively,and the MIo U and Dice are 85.7% and 86.6% respectively.The results show that the algorithm proposed in this paper has a great improvement in accuracy and real-time performance,and its performance has reached a very high level.(4)A surgical simulation experiment platform is built and a dataset is built.The trained model is deployed on the experimental platform for testing.The experimental results show that the algorithm proposed in this topic still has good performance in the real environment.