Location Of Knee Puncture Point Based On Target Detection And Segmentation

Knee joint is the largest and most complex joint in the human body.In recent years,knee joint diseases appear frequently,which seriously affect people's normal life.At present,artificial puncture is still used for many diseases of the knee joint,which is timeconsuming and laborious,but also put forward higher requirements for clinicians.To solve the problems of long manual puncture time,high requirements for physician experience and subjective factors,a target location algorithm based on object detection and image segmentation is proposed,which provides a more reliable basis for the robot to achieve automatic puncture.Aiming at the problem that it is difficult to distinguish the lesion area from the background color in the ultrasound image of knee joint lesions,the original image is preprocessed by the object detection network,and a bbox containing the target area is obtained.Considering the real-time problem of the whole system,three fast single-phase detection networks,SSD,YOLOv3 and Mobilenet-SSD,are compared.Finally,a suitable detection network is selected.Meanwhile,considering the impact of dataset size,the original dataset is augmented by clipping,translation,random rotation and other methods.To segment the lesion area accurately,the semantic segmentation is used to classify the detected image at pixel level.The segmentation results of U-net and Deep Lab networks are compared,and the spatial pyramid pooling(SPP)is introduced into the VGG network for the bboxes images with different scale output from the detection network.A VGG-SPP segmentation network is proposed.Meanwhile,the original cross entropy loss function is changed to the loss function with the exponential weight factor,which improves the classification accuracy of the hard samples.Finally,experiments show that the segmentation detection image is more than 0.11 higher than the segmentation original image,and the Dice coefficient of the VGG-SPP network on the training set and the test set is 0.92 and 0.85.For the binary image of irregular shape outputted by the segmentation network,the concavity and convexity analysis of the contour is performed,and the sub-region segmentation point which is the best in the contour is calculated by the method based on the pit matching.Considering that part of the lesion area belongs to the convex polygon,it is not necessary to divide the sub areas.In this paper,the sub areas are divided according to the number of concave areas.The centroid of each sub region is calculated by the image moment,which is regarded as the best puncture target.Finally,the experiment proves that the proposed positioning algorithm achieves the expected target position.