Research On Segmentation Method Of Vertebral CT Image Based On Improved U-Net Network

Medical image segmentation is not only an important auxiliary diagnosis and treatment technique in clinical practice,but also a hot research field.In the treatment of orthopedic patients,doctors with theoretical basis and clinical experience manual segmentation of patients' vertebral CT images,easy to cause diagnostic errors.Although manual segmentation is accurate,it requires high basic knowledge and experience of doctors,and the segmentation efficiency is very low.In view of the above problems,this paper proposes an improved method of vertebral CT image segmentation.The main research work is carried out in the following two aspects:1.Firstly,in view of the multi-organ interference and segmentation of vertebrae image noise,an appropriate preprocessing method was selected to reduce the noise in the vertebral body image through anisotropic filtering.The gray linear transformation increase the contrast to make the vertebral body part in the CT image more clear.Secondly,the data set is expanded by rotation,translation,scaling and other operations to provide data conditions for the training network and prevent network overfitting.2.Making improvements based on U-Net model,and use it to segment the vertebrae CT images.In order to improve the segmentation speed,the problem is solved from two aspects.First,the convergence speed of the network is improved.The activation value of the previous layer is re-normalized by using the specification layer,and the activation function is re-enlarged.Secondly,the basic convolutional layer is replaced by deep convolution or point-by-point convolution while minimizing the network structure while ensuring the segmentation accuracy.In order to refine the segmentation results of vertebral margins,the feature maps obtained based on the DB-U-Net segmentation results were transformed into probability images.3.Combining the existing segmentation precision with the Graph Cutalgorithm,the boundary term energy function and the region term energy function are introduced.The segmentation problem is transformed into an extremum problem of energy function.The value of the energy function is experimentally analyzed to improve the degree of edge segmentation.After experiments,the proposed segmentation network can accurately segment the vertebral region in CT images.Among the indexes of vertebrae segmentation,compared with the original FCN segmentation results,the segmentation accuracy of the algorithm in this paper is significantly improved.Compared with the gold standard,the degree of segmentation is highly consistent,which can play an auxiliary guiding role in the diagnosis of clinical vertebrae diseases.