CT Liver Tumor Segmentation Method Based On Improved Attention Mechanism And Cascade Mode

More than half of the world’s deaths from liver cancer have been in China,and the number of deaths is still increasing year by year.The main manifestation of liver cancer is liver tumor.Enhanced computed tomography(CT)images are an important means of detecting liver tumors.Segmenting liver tumors from abdominal CT images is great significance for subsequent clinical diagnosis and treatment.With the number of patients grows,a large amount of image data needs to be processed,but the manual method has problems such as subjective interference,different standards,complicated procedures,time-consuming and labor-intensive.The automatic segmentation method based on deep learning can effectively assist doctors to complete the segmentation and diagnosis,and reduce the missed diagnosis and misdiagnosis caused by subjective errors.The accuracy of the automatic segmentation algorithm affects the performance of the computer aided diagnosis system.Therefore,this paper studies the automatic segmentation method of liver and liver tumors in abdominal CT images.This paper discusses the related research of traditional image processing methods,machine learning methods and deep learning methods in liver tumor segmentation,and summarizes the current research status.Traditional image processing methods can only deal with simple image features such as grayscale and texture,and usually have poor results.Most of the traditional machine learning methods need to carefully design artificial features,which often have limitations such as difficulty in describing global information and high algorithm complexity.Due to the advantages in feature extraction and other aspects,deep learning methods occupy the mainstream in liver tumor segmentation.The main method of this paper is based on the convolutional neural network in deep learning.The Li TS public dataset is used to train our methods.Combined with the needs of practical applications and analyzing the characteristics of liver and liver tumor segmentation,we propose a cascade segmentation model based on an improved attention mechanism.The cascade model is divided into two main stages.The first stage is liver segmentation,and the second stage is liver tumor segmentation.The main research contents are as follows:(1)We propose a liver segmentation model based on self-attention and multi-scale feature fusion.For the problem of blurred boundaries,the multi-scale feature fusion method is used to extract spatial information,and the self-attention is used based on the relationship between the liver and surrounding organs.The force mechanism captures the association between multi-scale fused features,combined with the improved attention convolution module to further improve the segmentation accuracy of liver organs.The average Dice value of the overlapping area between the predicted segmentation results and the ground truth segmentation can reach 96.4%,which is 4.4% better than that U-Net,and reaches a level similar to the advanced 3D model through the 2D model method.(2)We propose a liver tumor segmentation model based on UNet bottleneck feature selection and skip connection optimization.In view of the characteristics of liver tumors with unclear boundaries and variable tumor size and location,combined with the model structural characteristics of UNet,the improved attention method in this paper is used to process the bottleneck features of UNet.The improved attention method we proposed suppresses the expression of non-main features,enhances the proportion of main features,and enhances the model’s ability to locate the liver tumor area.And through the spatial attention module and the residual fusion module in skip conncetion to deal with the problem of feature semantic gap and segmentation boundary.The model has performed segmentation experiments on both liver and liver tumor.The average Dice of liver prediction segmentation is 96.2%,the average Dice of liver tumor prediction segmentation is 68.4%,and the Global Dice is 80.5%.A cascaded segmentation model is designed by the liver tumor segmentation model and the first-stage liver segmentation model,which can further improve the segmentation accuracy while obtaining additional liver segmentation results.And the average Dice of liver tumors reached 70.1%,and the Global Dice reached 81.3%,reaching the leading level in 2D models.Finally,based on the cascaded segmentation model,this paper designs and implements a prototype system for automatic segmentation and auxiliary staging of liver and liver tumors.The system provides reference results for automatic segmentation of CT images,quantification of tumor characteristics and auxiliary diagnosis for staging diagnosis.