Automatic Measurement And Correlation Analysis Of Femur Proximal Parameters Based On Deep Learning

China has entered a deep aging society and will enter a super aging society in 2033.The prevalence of diseases in old age is also bound to increase.In elderly patients with bone disease,most of the diseases are related to femur and femur fracture accounts for a large proportion.At present,the main treatment method for proximal femoral fracture in China is hip prosthesis replacement.Most of these prostheses are imported.According to the clinical practice of the Second Affiliated Hospital of Xi ’an Jiaotong University and the General Hospital of the Chinese People’s Liberation Army,the current imported prostheses do not match the shape of Chinese femur well,resulting in poor postoperative recovery effect.The measurement of the parameters of the proximal femur of Chinese people can assist doctors to design artificial prostheses suitable for Chinese people,and also assist doctors to choose surgical programs and instruments.In addition,using the data obtained from measurement to establish a database of femur parameters is conducive to the related research of orthopedic disease prevention and treatment.Currently,artificial methods are used to measure the parameters of proximal femur.Manual measurement is time-consuming and laborious,with low repeatability and human bias.In recent years,with the development of deep learning technology,automatic measurement using neural network has become a research hotspot.In this context,in cooperation with the Second Affiliated Hospital of Xi ’an Jiaotong University,this thesis studied the measurement method of femur proximal parameters based on computer scan tomography data.The main work of this thesis is as follows:(1)For the problem of three-dimensional reconstruction of femur data,Res Net was used to divide CT scan data of hip joint converted into pictures into femur,tibia and other three categories.Then the improved U-Net network was used to segment the internal and external contours of femur,and the effectiveness of the method was verified by comparative experiments.At the same time,the three-dimensional point cloud model of femur was reconstructed and the thickness of femur cortex was measured.(2)In order to improve the segmentation accuracy,this thesis proposes an improved model of point cloud segmentation network Pointnet ++.The improved model was used to segment the femur into femoral head,femoral neck,and femoral shaft.Compared with other models,the improved network model in this thesis can achieve high precision segmentation of the femur and prepare for subsequent measurement.(3)On the basis of femur segmentation,this thesis calculated corresponding marks by fitting method based on morphological features of each part.Six parameters of femoral head diameter,femoral shaft diameter,cervical trunk Angle,femur length,eccentricity and anterior inclination Angle were calculated by using the marker points and corresponding mathematical formulas,and then the correlation was analyzed.The accuracy and reliability of the experimental results are verified.(4)To compare the difference between two-dimensional and three-dimensional measurements of cervical stem Angle.In this thesis,the optimal two-dimensional projection plane of cervical trunk Angle was found and the femur was projected onto this plane.On this basis,a two-dimensional measurement method of cervical stem Angle is proposed and compared with the results of three-dimensional and manual measurement.The results show that there is no statistical difference among the three methods.