Research And Application Of Multi-stage Pancreas Segmentation Based On Convolutional Neural Network

Pancreatic cancer is a more serious malignant tumor and is called the king of cancer.In 2020,pancreatic cancer ranks 8th in the number of new cancer cases in China,and 7th in mortality.Its 5-year survival rate is only about 9%,which has brought great harm to the lives and health of the people.The key to reducing the incidence and mortality of pancreatic cancer is to achieve early diagnosis and early treatment.Pancreas segmentation is the basis of auxiliary diagnosis and treatment of pancreatic diseases.Segmentation of the pancreas in the CT image can not only extract the region of interest of the pancreas,but also facilitate the analysis and identification of pancreatic tissues and lesions;It can also measure the size and volume of pancreatic organs,which is convenient for doctors to carry out quantitative treatment plans;At the same time,it is also a prerequisite for the 3-dimensional reconstruction of pancreatic organs,which can assist in the development of surgical procedures.In summary,the work of pancreas segmentation based on CT images has theoretical research and practical clinical application value.This thesis focuses on the pancreas CT image segmentation method based on deep learning,combined with the anatomical characteristics of the pancreas itself,and analyzes the problems of current segmentation techniques and models.The research content is as follows:1)Aiming at the problem that the traditional U-Net segmentation method is easily affected by irrelevant background information and ignores the sequence information between slices,a multi-stage pancreas segmentation method based on DASU-Net is proposed.The entire pancreas segmentation process is divided into three stages: The first stage uses anatomical prior information to roughly locate the pancreas,thereby roughly narrowing the background;The second stage uses the designed DASU-Net and slices context information extraction module to roughly segment the pancreas,and then performs fine locating based on the segmentation results of a single pancreatic slice,so as to further reduce the irrelevant background;the third stage uses DASU-Net to complete the fine segmentation.The verification results on the NIH public data set show that the proposed method improves the under-segmentation and mis-segmentation,and the final DSC coefficient reaches 84.26%,and the JI index reaches 72.63%.2)Aiming at the problem that the multi-stage pancreas segmentation method based on DASU-Net with complicated operation,insufficient feature extraction ability of segmentation network model and large amount of parameters,a multi-stage pancreatic organ segmentation method based on ADT-YOLO is proposed.The whole segmentation process consists of three stages: the first stage uses pancreatic anatomy prior locating to roughly remove irrelevant background;the second stage uses the designed ADT-YOLO detection model to achieve end-to-end pancreatic detection,thereby further reducing the pancreatic area.The third stage uses the designed PCSHRNet model to complete the fine segmentation of the pancreas.The test results on the public data set show that the method proposed in this thesis can simplify the locating steps while maintaining a high accuracy of pancreatic segmentation.In addition,the designed PCS-HRNet segmentation model not only reduces the amount of parameters but also improves the feature extraction ability.3)This article designs a pancreatic CT image segmentation system based on actual clinical needs in the process of differential diagnosis and treatment of pancreatic diseases,and combined with doctors’ habits.Based on the above segmentation algorithm and related technologies,the system provides pancreatic CT image preprocessing,reading,segmentation and result optimization,which can improve the accuracy and efficiency of the diagnosis of pancreatic diseases,thereby reducing the burden on doctors’ work,and has a broad clinical application prospect.