Research On Liver CT Image Segmentation Based On U-Net Network And Graph Model

Computed tomography(Computed Tomography,CT)has become an important method for surgeons to assist diagnosis and treatment with its advantages of high signal-to-noise ratio,clear image,and fast imaging speed.Liver diseases have increasingly attracted the attention of researchers.The shape and size of the diseased liver are significantly different from those of normal people.Therefore,doctors often need to segment the CT image of the patient’s liver by Observe the segmentation results and make judgments on the patient’s condition.Although manual segmentation of the liver is high,the segmentation accuracy is high,but the segmentation result lacks objectivity,and the time cost is relatively high.The research in this paper is carried out to improve the efficiency of doctors.In this paper,we have adopted the U-Net network that has developed rapidly in recent years and improved it on the original network:In order to increase the training speed of the network,the BN algorithm is introduced;In order to optimize the parameters quickly without causing the gradient to disappear,add Leaky Relu function.The U-Net network is a fully automatic segmentation algorithm,which belongs to the category of deep learning algorithms.The U-Net network that has been trained can independently learn the imagery and abstract features of the image through its own layer-by-layer convolution,and finally complete the automatic image segmentation.On the basis of the improved U-Net network,two graph models are added for optimization.One of them is the Ncut algorithm that adds spatial nodes to fuse the segmentation results to increase the spatial feature extraction ability of the segmentation nodes.The other one uses the fully connected conditional random field model to optimize the edge detail segmentation results of the segmentation results.The main work of this article is:1.The composition of the U-Net network is introduced in detail.In order to improve the segmentation performance of the U-Net network,the original U-Net network is improved:drawing on the fully convolutional neural network model(Fully convolutional neural network,FCN),adding the BN algorithm,in addition The activation function is changed to Leaky ReLU function to reduce the appearance of silent neurons,which can better complete the initial segmentation of the image.2.In this paper,the Ncut algorithm with spatial nodes is added to achieve the segmentation of liver CT images.The deep optimization of the deeplab-v2 model is used for reference,and the fully connected conditional random field model is introduced to optimize the edge details to obtain the final segmentation results and assist the clinician.Diagnosis and treatment of diseases.3.Introduced the 3Dircadb and Chaos 2019 two data sets,and used the median filter to denoise and normalize the related images in the data set as the preprocessing,and expanded the experimental samples with the data enhancement technology of the training set.In this paper,the segmentation model used to segment the liver CT image is compared with the region growth algorithm,the original U-Net segmentation algorithm,the improved U-Net segmentation algorithm,and the effect is obtained by using only the improved U-Net segmentation algorithm and the Ncut algorithm segmentation method.To further improve.Finally,the joint model adopted in this paper is further summarized,and the development direction of the model is prospected.