Research On Liver Tumor Segmentation From CT Images For Precision Radiotherapy

Liver cancer is one of the major diseases that seriously threaten the life and health of Chinese people.Precision radiotherapy is one of the main means to treat liver cancer effectively.Tumor segmentation from 3-dimensional CT(3D-CT)or four-dimensional CT(4D-CT)is the prerequisite and basis of precision radiotherapy.Clinically,the tumors are delineated by specialist manually,which is time-consuming and laborious.In addition,the segmentation accuracy is susceptible to human factors,which affects the outcomes of precision radiotherapy.Therefore,it is urgent to conduct the research on semi-automatic or automatic liver tumor segmentation from CT images in precision radiotherapy for liver cancer.However,the features of liver tumors appearing on CT images are extremely complex,so the semi-automatic or automatic segmentation of liver tumors is a highly challenging task.Although many methods have been proposed,there still exhibit some problems such as low accuracy and efficiency,weak robustness,and difficulty in accurately quantifying segmentation uncertainty.To adress these problems,a systematic and comprehensive research has been carried out in this thesis from two dimensions,i.e.3D-CT and 4D-CT,and three levels,i.e.semi-automatic segmentation,multi-stage automatic segmentation and end-to-end automatic segmentation.The main research contents and innovations are as follows:(1)A semi-automatic liver tumor segmentation method with adaptive region growing and graph cuts was proposed.Firstly,to overcome the influence of noise and improve the robustness,an adaptive region growing method based on Kullback-Leibler divergence was proposed to initially segment the tumors and obtain the region of interest;Then,to improve the contrast between the tumor and non-tumor area,a non-linear mapping method based on Gaussian fitting was proposed to enhance the region of interest;Finally,the final segmentation results were obtained by building and minimization energy function of graph cuts based on the gray information in the enhanced region of interest and the gradient information in the original region of interest.This method can effectively improve the segmentation accuracy,and is robust.(2)An automatic liver tumor segmentation method based on cascaded Dense-UNet and graph cuts was proposed.First of all,a Dense-UNet network structure was proposed in view of the fact that deep network training often leads to gradient disappearance,which used cascade segmentation style to segment the liver first,then the tumor.After that the region of interest was obtained;Then,the gray model and probability model that can effectively distinguish tumor from non-tumor was constructed for the region of interest;Finally,based on these two models,a graph cut energy function was constructed to automatically and accurately extract liver tumors from CT images.The segmentation accuracy of this method is obviously superior to many existing automatic segmentation methods.It is also insensitive to manual annotation errors,so has good robustness.(3)An end-to-end liver tumor automatic segmentation method based on D-Net was proposed.First,we constructed a lightweight liver segmentation model,by using dilation convolution,residual connection and skip connection;Then,the network structure was deepened based on the lightweight model,and a tumor segmentation model was built by introducing a random dropout block;Finally,the two models were connected by a composite loss function to obtain an end-to-end tumor segmentation model,D-Net.This method not only improves the segmentation efficiency,but also maintains high segmentation accuracy.Furthermore,based on the dropout block,Monte Carlo sampling technology was used to realize the uncertainty evaluation of liver tumor segmentation.(4)A liver tumor segmentation method based on spatiotemporal dual path network from 4D-CT sequences was proposed.First,a spatiotemporal parallel decoding path was designed to extract spatiotemporal features synchronously;Then,channel and spatial attention mechanism were introduced to construct a spatio-temporal feature fusion module to achieve effective screening and fusion of spatio-temporal features;Finally,the fused features were decoded layer by layer,and the spatial details of the images were retained by using skip connections,thereby completing the construction of the entire segmentation network.This method realizes automatic and accurate segmentation of liver tumors in 4D-CT sequences first time,which can effectively improve the clinical work efficiency.The research in this thesis provides new methods and ideas for automatic and accurate tumor segmentation in precision radiotherapy for liver cancer,which is helpful to improve the outcomes of precision radiotherapy,reduce radiotherapy toxicity,save human resources,shorten the course of radiotherapy,speed up patient turnover,and benefit more patients with liver cancer.It has important clinical significance and practical value.