Classification Of Femoral Fractures Based On Fewshot Learning Method

Computer aided diagnosis that can reduce doctor’s workload and improve diagnosis efficiency is a new trend of medical development in the future.In the process of fracture diagnosis,it is necessary to judge the location of fracture and the specific type of fracture to take the best treatment.In clinic,thanks to computed tomography（CT）has the advantages of higher consistency and lower probability of misjudgment in the diagnosis of fracture types,CT images are generally used for analysis.But The classification results often depend on the observer’s experience and intuitive impression,which is easily to misdiagnosis.In addition,there are few samples and hard to obtain,because of the data involving patient privacy and uneven distribution of diseases.In order to deal with those problems,this thesis focuses on the key technologies such as3 D reconstruction,CT image registration,multi-view sampling and convolution neural network.Finally,the diagnostic efficiency of fracture classification is improved and the misdiagnosis rate of fracture classification is reduced by these studies.（1）Aiming at the problem of few samples,this thesis makes full use of the threedimensional characteristics of CT images,improves the fusion way of the original multi-view method,and proposes a fracture classification method based on multi-view attention fusion.Firstly,the VTK reconstruction library is used to reconstruct the original CT image and then Multi-Angle sampling of the 3D model to increase the amount of sample data.This method uses the way of channel concat to realize data fusion on the channel in classifying,and then puts the fused data into the 2D channel attention convolution network for classification.Experiments show that this method is superior to other methods in classification accuracy.Comparing with the results obtained based on three-dimensional convolution model,the accuracy is improved by10% in six-part standard and 5% in AO（Arbeitsgemeinschaftfür Osteosynthesefragen）standard.（2）To handle the problem of uncertain pose of CT data model after reconstruction,in this thesis,a fracture classification method based on local pose registration is proposed.Firstly,the CT data are locally registered to eliminate the feature occlusion caused by the tilt of pose during the reconstruction of the model,and then the registered3 D model is sampled from multiple-view.Then,the rotation matrix is introduced into the classification model to predict the rotation relative pose relationship of the model in space,so as to solve the interference of the initial pose uncertainty on the classification model and improve the classification performance.Experiments show that the accuracy of this method is improved by 8% in the six-part model compared with the attention perspective fusion method.（3）In order to verify the effect of the algorithm in the actual scene,this thesis designs a set of auxiliary classification system of femoral fracture based on the above femoral fracture classification algorithm.The system realizes graphical interactive display,including the display of cross-section,coronal plane and sagittal plane of the original fracture CT image,as well as the effect after 3D reconstruction.In addition,it also realizes the functions of predicting which display the classification results and displaying the basic information of patients.Experiments show that with the help of the classification system,the diagnostic detection rate has been improved.