Research On Surgical Instrument Image Detection Method Based On Deep Learning

With the advancement of human medical level,the application of surgery in clinical medicine is becoming more and more extensive,and it is developing in the direction of minimal trauma,low risk and high efficiency.The disinfection supply center of the hospital is responsible for cleaning,sterilizing and packaging the surgical instruments used in surgical operations.In this process,it is necessary to manually count the surgical instruments many times,which often brings about low efficiency and high error rate.Therefore,it is a trend to use machine vision technology to replace manual inventory of surgical instruments.However,in the actual inventory link,some surgical instruments have the characteristics of densely arranged,severely reflective,and highly similar to the background,which makes it difficult for traditional object detection technology to perform the inventory task.In order to solve the above problems,an image detection method of surgical instruments based on deep learning is proposed in this paper.This paper takes the image detection of surgical instruments as the research object and uses the YOLOv4 algorithm to study it.The YOLOv4 algorithm has the problems of low detection accuracy,high false detection rate and low Io U(Intersection over Union)value for surgical instrument data sets.The YOLOv4 algorithm has been improved while combining the characteristics of surgical instruments in this paper.The specific work completed is as follows:1.As a special surgical instrument,hemostatic forceps are placed in a highly overlapping manner during the inventory process.The YOLOv4 algorithm has low detection accuracy for this target.In this paper,the normalized area constant of the hemostatic forceps is proposed for the actual inventory working conditions of the hemostatic forceps,and it is applied in the DIo U-NMS algorithm.Before the DIo U-NMS algorithm executes the loop deletion operation,the candidate frame of the hemostatic forceps whose area is much larger than the constant area of the hemostatic forceps is traversed and deleted first.Furthermore,a penalty factor associated with the hemostatic forceps area constant is added to the selection formula of the DIo U-NMS algorithm.Experimental results show that the AP(Average Precision)value of the hemostatic forceps with this improved YOLOv4-NMS algorithm reaches 82.1%,which is 19.4% higher than the YOLOv4 algorithm.The effectiveness of the improvement has been fully proven.2.Aiming at the problem of serious reflection of surgical instruments,the correlation between the average gray value of surgical instrument images and the recognition accuracy of YOLOv4 algorithm is analyzed in this paper.Therefore,the adaptive Gamma correction algorithm is introduced.Images with an average gray value higher than 145 are preprocessed.The gray value of the image is reduced to the interval of 135-145 where the recognition accuracy is better.Experimental results show that the m AP(mean Average Precision)of this algorithm is increased to 92.73%,which is 7.74% higher than that of YOLOv4.The missed detection rate and false detection rate of surgical instruments are effectively reduced.3.Aiming at the problem of the low Io U value of the YOLOv4 algorithm for detecting surgical instrument data sets,the better-performing k-means++ clustering algorithm is used in this paper to replace the k-means algorithm used by YOLOv4.The k-means++ clustering algorithm is used to cluster the surgical instrument data set.The experimental results show that,compared with the YOLOv4 algorithm,the Io U value has increased to 0.91 by the algorithm of this paper,and the detection result shows that the positioning of the label frame is more accurate.4.Seven representative surgical instruments were selected in this paper,and the stainless steel grid basket which used in actual working conditions was used as a background to establish a surgical instrument data set.In addition,the Cut Mix image enhancement method is used in this paper to expand the surgical instrument data set during training.This thesis draws a conclusion by comparing the experimental data on the surgical instrument data set and the actual application performance of the algorithm.Due to the above improved design,compared with the YOLOv4,YOLOv3 and Faster RCNN algorithms,the m AP value and Io U value of the improved algorithm in this paper are significantly improved,which effectively reduces the missed detection rate and false detection rate of surgical instruments.