Liver Tumor Segmentation Based On Concatenated Fully Convolutional Network

In recent years,Computed Tomography(CT)has become a commonly used examination for liver diseases.The treatment of liver tumors requires accurate understanding of the size,location,and number of tumors before surgery.Therefore,segmentation of liver tumors from liver CT is the primary task of treatment.The difficulty of tumor segmentation mainly lies in the low contrast of CT image,complex size and shape of tumors,adjacent to other organs,and blurred boundaries between normal liver region and tumor region.Therefore,segmentation of liver tumors is still a difficult and hot spot in the field of medical imaging.This thesis studies the method of liver tumor segmentation based on fully convolutional networks.On the basis of fully convolutional networks framework,the thesis conducts in-depth research on the problem of segmentation inaccuracy caused by liver segmentation,tumor segmentation and edge blurring.Liver and tumor segmentation results are obtained by concatenated fully convolutional networks.Introduction of variable pooling kernels in the first layer of liver segmentation network,retains more information about liver and improves liver segmentation accuracy.After obtained the rough segmentation results through the fully convolutional networks,fully connected conditional random field is used to optimize the rough segmented results by concatenated fully convolutional networks,constrains the segmentation edge,and improves the segmentation precision.The main research work and innovations of the thesis are as follows:(1)A liver tumor segmentation method based on concatenated convolutional networks is proposed.The CT images are preprocessed by Gaussian filter,sharpen operating,histogram equalization,respectively to remove noise,highlight details and edges,improve image contrast.The liver dataset is used to train the two networks,the first layer network realizes liver segmentation,and the liver segmentation results are used as the input of the second layer network to train the tumor segmentation network.Fully convolutional networks extract features from original image by the convolutional layers and the pooling layers.The networks automatically learn higher order features by using multiconvolutional layers,classify the pixel points one by one,and restores the feature map to the original size by upsampling.(2)A liver tumor segmentation method based on variable pooling kernals networks and dilated convolutional networks is proposed.During the downsampling process of fully convolutional network,the pooling layers are used to reduce the image size to extract higher order features,but the traditional network uses the same pooling kernels,which ignores the liver’s position information.Because the human’s liver is generally located on the left-front side in abdominal cavity,which is equal to the upper-left of the CT image,smaller pooling kernels are used in the upper left region of the image to retain more liver information,while larger pooling kernels are used in region which has nothing to do with liver.Thus,variable pooling kernels retain more liver features and improve the segmentation accuracy of the liver.(3)A liver tumor segmentation method based on improved concatenated fully convolutional networks and fully connected conditional random field is proposed.Because the segmentation results of fully convolutional networks have the problems that the receptive field is too large and the edge constraint is insufficient.Therefore,after the whole convolutional networks,the fully connected conditional random field is used to finely segment the result again and enhance the ability of edge constraint.The thesis experiment and verify the proposed methods above through the published liver dataset 3D-IRCADb-01.The experimental results show that the concatenated fully convolutional networks segment tumor more precisely than the single-layer fully convolutional network.The variablepooling kernel method used in convolutional network can improve the segmentation accuracy of the liver.It can perform fine segmentation on the rough results of the network and enhance the edge constraint ability.