Femur Segmentation Based On Deep Learning

With the acceleration of computer data processing capabilities and the maturity of artificial intelligence technology,doctors are increasingly using digital images for assisted diagnosis and treatment.X-ray examination is a traditional method for effective screening of diseases.It is used in different stages of diagnosis and treatment,including fracture diagnosis and treatment,evaluation of bone maturity,bone density measurement and pre-operative treatment planning.Segmentation of bone tissue in X-ray films is a major step in computer-assisted prognosis,surgery,and treatment.Because of the limitations of medical imaging techniques,the specificity of imaging objects,and inherent characteristics of medical images,such as uneven gray scale,overlapping images,loud noise and blurred borders,make the segmentation of medical images full of challenges.At present,the segmentation of X-ray bone tissue is mainly by manual labeling by doctors,which is not only time-consuming and labor-intensive,but also difficult to reproduce the segmentation result,and the accuracy is unstable.Based on this,the main work of this paper is as follows:1.Taking femur as the research object,this paper proposed a new deep neural network R-U-Net for the problem that the current X-ray films segmentation can not be automated and the U-Net network has the problem of network degeneration in femur segmentation.It combines the advantages of deep residual networks and U-Net architectures,using residual units instead of ordinary neural units as basic blocks.The jump connection inside the residual unit and the codec path of the R-U-Net network facilitate the forward propagation and backward calculation of the information,which not only simplifies the training,but also makes the network parameters only account for a quarter of U-Net.And it can maximize network performance.2.The automatic segmentation framework of the femur area was designed and implemented,and the proposed R-U-Net neural network was applied to the automatic segmentation of the femur area.The horizontal can be divided into two main parts: the neural network training phase and the batch automatic segmentation phase.Longitudinal can be divided into three main parts: data preprocessing,neural network training and testing,and image post processing.Firstly,after pre-processing of the original dataset,made the label images in the target area,and augmented the training dataset.Then,inputted it to the R-U-Net neural network for training and adjusting parameter,and saved the optimized network model.Finally inputted the test images to be segmented into the saved network model to obtain the outline of the femur area,and performed the mask operation after filling the connected area to obtain the femur area segmentation results.A fully automated segmentation of the end-to-end X-ray femur is achieved.3.Automatic segmentation of the femur area was achieved based on U-Net and R-U-Net neural networks,respectively.The pre-processed data sets are trained and tested on U-Net and R-U-Net neural networks respectively.Comparing the training convergence process,it can be seen that the R-U-Net neural network proposed in this paper improves the deficiency of the original U-Net.It is more excellent in the training of the femur segmentation network model.The optimal models are saved separately,and five metrics commonly used in medical image segmentation are used to quantitatively evaluate the automatic segmentation method of femur area based on U-Net and R-U-Net neural network.Taking Photo Shop artificial segmentation results as a reference,the experimental results show that the R-U-Net neural network based method guarantees the segmentation speed,and the segmentation result is very close to the reference result,which has high accuracy.Moreover,the automatic segmentation effect of femur area based on R-U-Net neural network is better than other two advanced image segmentation methods,and the execution efficiency is higher and the segmentation effect is better.