Research And Application Of Liver CT Image Segmentation Based On U-Net

The liver plays an indispensable role in the body’s internal metabolism and detoxification.However,there are many types of liver diseases,and the incidence rate is also high.Among them,the lethality rate of primary liver cancer ranks third among cancers.For liver tumors,which have prominent morbidity and mortality,prevention is the primary treatment in the early stage,and partial liver resection is often used as a treatment method in the middle and late stages.In the process of diagnosis and treatment of liver diseases,doctors need to observe and analyze the condition of the liver with the help of medical images.With the continuous improvement of medical standards,the amount of medical image data has increased sharply,and manual segmentation is no longer competent for medical image analysis.Therefore,automated medical image segmentation is more helpful to improve the work efficiency of medical personnel,and can effectively avoid evaluation differences caused by the subjective analysis of doctors.Deep learning methods have advantages in acquiring global image information and contextual information,and have been favored in medical image segmentation research in recent years.Combined with the analysis and summary of the current liver and liver tumor segmentation research,this thesis,based on U-Net in deep learning,successively carried out the research on liver segmentation and liver tumor automatic segmentation.The main content of this thesis is as follows:1)In the study of liver segmentation,in order to solve the problem that the background area composed of various organs and tissues in abdominal CT is too complicated,and the conventional neural network is not sensitive to the local information of the liver,a cyclic dense connection based on local feature enhancement is proposed.U-Net liver segmentation method.This method uses the difference between the attributes of the liver and the background area in the CT image,and performs data preprocessing by constraining the CT value interval,making the characteristic information of the liver area more significant,and to a certain extent solves the problem of complex background information affecting the efficiency of network training.Problem: In order to obtain the global and local information of CT images in the network,a densely connected module that mixes more local information is designed,so that the segmentation model can learn more sufficient feature information,and solve the problem that the network is not sensitive to local information.Experiments show that this method has higher liver segmentation accuracy,and the segmentation model has better generalization ability.2)In the study of liver tumor segmentation,a method of liver tumor segmentation based on channel attention and multi-scale U-Net is proposed for the problem that it is easy to ignore the local key feature information in the network learning process.This method analyzes the location of unrelated tumor regions and preprocesses the input data of the liver tumor segmentation network.Only the region where the liver is located in the CT image is retained,which solves the redundancy caused by the presence of tumors in other organs and tissues of the abdomen with too high characteristics The problem of segmentation: In order to learn more effectively the feature information of the tumor region,a liver tumor segmentation network(HCAM-Net)based on channel attention and multi-scale U-Net is designed.Experiments show that this method can achieve better liver tumor segmentation performance.3)An auxiliary diagnosis system for liver and its tumor segmentation based on abdominal CT image was designed and developed.The system realizes functions such as routine processing of abdominal CT images,liver and liver tumor segmentation and segmentation optimization.This not only reduces the labor and time investment of medical institutions in image annotation,but also provides doctors with fast and effective segmentation prediction.This has a great effect on the efficiency of doctors.