Research On Automatic Segmentation Method Of Liver And Liver Tumor

China is a country with a large population.With the aggravation of the aging population,the burden of cancer has become increasingly heavy.Liver cancer,as the second largest cancer in China,poses a huge threat to people’s lives and health.Due to the inconspicuous early pathological characteristics of liver cancer and the high mortality rate in the late stage,early prevention and intermediate treatment are relatively good coping strategies for liver cancer.Computed Tomography(CT)is the most commonly used medical imaging technology in clinical diagnosis and treatment.Accurate and fast automatic segmentation methods for medical images of liver and liver tumors are of great significance for early screening and diagnosis of liver diseases,improving doctors’ work efficiency,and assisting clinical diagnosis.Accurate and automatic segmentation of liver organs is the basis for the diagnosis of liver cancer and other liver diseases.However,in CT images,the liver edge is blurred and difficult to distinguish from adjacent organs and tissues.In addition,compared to the number of liver organ and background pixels,the liver tumor pixels are very unbalanced,and the size,number,and shape of liver tumors have strong specificity in different patients,greatly increasing the difficulty of segmentation.In response to the above issues,the research content of this article mainly includes the following parts:(1)To solve the problem of blurred liver edges and difficulty in distinguishing them from adjacent organs and tissues,an automatic liver segmentation method based on diffusion probability model and edge enhancement(EEDiff)was proposed.Firstly,EEDiff gradually adds noise to the edge probability map through a diffusion process and converts it into latent variables.Through the learning inverse diffusion process,the latent variables are iteratively converted into accurate prediction of the liver edge.Secondly,EEDiff includes a semantic segmentation branch and an edge noise prediction branch,enabling the model to simultaneously focus on the discriminative features and details of input data through the form of multi-tasking learning.At the same time,we design a multi-scale feature fusion module to interactively fuse the potential target information of the original slice and inaccurate edge information in the noisy edge map.Finally,the deformable Transformer module is used to learn the global information association between pixels.(2)Aiming at the problem that existing convolution-based methods are difficult to establish long-distance dependencies due to the locality of convolution.A multi-scale liver tumor segmentation algorithm(Conv Trans Net)combining convolution and Transformer is proposed.The algorithm uses a hybrid structure of convolution and transformer at the same feature layer level to simultaneously explore information associations at different scales.In addition,in the process of Transformer self-attention computing,the fusion representation of coarse grained vectors and fine grained feature vectors is explored through mix embedding.In order to reduce semantic loss during the down-sampling process,a multi-level feature fusion module is designed at the skip connection.This module enriches the semantic information of the current level by fusing the front and rear features of the current scale feature map.Finally,in order to cope with different scale tumor pixels,a multi-scale module combined with deformable convolution is also introduced.(3)A set of liver tumor assistant diagnosis system based on deep learning was designed for the increasing number of patient CT images.The system can quickly locate the liver region,segment liver tumors in the liver,and timely analyze tumor data of different sizes and diameters based on the input CT images,to help clinicians make more accurate and scientific conclusions.In order to evaluate the algorithm proposed in this article more fairly and objectively,we used 131 CT images included in the MICCAI Li TS17 dataset to train and test the algorithm proposed in this article.The experiment proved the effectiveness of the algorithm proposed in this article,which can achieve accurate segmentation of liver and liver tumors and play an important role in clinical adjuvant therapy.