Research On Detection And 3D Segmentation Of Rib Fracture

Accidents such as car accidents,falling from height,criminal cases and other emergencies often cause chest rib fractures,resulting in human injury.As different types of rib fractures require different treatment methods,it is significance for disease treatment and injury grade identification to accurately detect the location of rib fractures and judge the type of rib fractures.The detection of rib fracture belongs to small target detection,the detection accuracy is difficult to improve,and it has high similarity with the surrounding area.The traditional artificial method of rib fracture will decline significantly in accuracy with the growth of working time and mental declineIn recent years,deep learning methods have been developed rapidly,the object detection technology and the object segmentation method have brought new ideas to obtain accurately rib fractures.The method of deep learning can not only quickly detect the area and type of rib fracture,help doctor to get the fracture type quickly,but also effectively conduct modeling of rib fracture in space,help doctor get the rib fracture area more intuitively and carry out effective treatment.However,due to the characteristics of rib fracture,the existing deep learning algorithm is not suitable for rib fracture detection and segmentation.Aiming at the problems,this paper designs the corresponding method of detection and segmentation,which improve the detection rate of rib fractures.The main work of this paper is summarized as follows:(1)On the basis of Center Net network,a method suitable for rib fracture object detection is designed.Firstly,according to the characteristics of rib fracture,this paper designed a hierarchical fusion hourglass network as the feature extraction module,which can capture the features of rib fracture more effectively.Secondly,in view of the problem that the heatmaps generated by the original network of Center Net contains less information and the misclassification is easy to occur when the corner points of two group bounding boxes have the same center by coincidence,this paper designed a Heatmap Pyramid Structure to make the heatmaps generated by the prediction module contain more information;Designed a Non-Local Double Spatial Attention module to supplement the global information for the center of each group of corner points,which improves the detection accuracy.The results show that the above strategies can improve the accuracy of the rib fracture detection.(2)The two-dimensional detection task was extended to the three-dimensional segmentation task of rib fracture.Based on Frac Net,on the one hand,in the feature extraction stage,this paper designed a Deep Encoding Block based on the experience of two-dimensional rib fracture detection task,so that the feature extraction network of U-Net can capture more rib fracture features and reduce the number of false positive samples;AAuxiliary Encoder is designed to supplement the lost information for the decoder of different scales,which makes all points improved compared with the original U-Net.On the other hand,in the post-processing stage,inspired by the spatio-temporal LSTM,this paper designed a Disspatiotemporal Correlation Module for the characteristics of the rib fracture task,which further reduced the false positive samples in the segmentation results.The experiment result shows that the proposed methods are beneficial to improve the final 3D rib fracture segmentation results.