| Lung cancer is one of the most common malignant tumors and a serious threat to human health.Radiation therapy is one of the ways to cure lung cancer.Lung four dimensional computed tomography(lung 4D-CT)plays an important role in lung cancer radiotherapy.Lung 4D-CT can reflect the characterization of organs,tissues and tumors with respiratory motion and the respiratory motion information captured by lung 4D-CT is tremendously valuable for precise radiation therapy.We can make use of lung 4D-CT images to implement individual and precise radiotherapy according to the motion character of patients,which can reduce tumor margin,guide high dose radiation to tumor,and at the same time decrease the radiation dose to normal tissue.But a lung 4D-CT dataset includes images of multi phases and contains thousands of images,which makes manual target delineation a heavy workload for doctors and affects doctor’s judgement about the target.The target delineation of lung 4D-CT is a heavy and time-consuming work,but the accurate radiation target definition is very crucial for cancer radiation therapy.The problem can be solved by using image segmentation techniques to help doctors delineate target.Image segmentation is to divide image,according to people’s needs,into several regions which have different characteristics.Medical image segmentation is to extract the interested organs and tissues from image.This paper aims to segment lung 4D-CT images and extract tumor in different phases to help doctors delineate target for cancer radiation therapy.Although many image segmentation algorithms came up recently,lung tumor is variable in size,shape and location and characterized by weak edges and low contrast with surrounding normal structures and prone to adhere to its surroundings,which makes accurate segmentation challenging.In addition,the segmentation algorithms should be highly automated in order to relieve the burden of doctors.So this paper proposed two novel lung 4D-CT image segmentation algorithms based on multiple constraints graph cuts.They both improved the segmentation accuracy and automation degree of graph cuts to assist doctors outline target and reduce their workload,which makes lung 4D-CT more widely applied in lung tumor precise radiotherapy.Firstly,we put forward automatic segmentation of lung 4D-CT tumor based on graph cuts with star shape prior.Doctors first select object seeds in the initial phase of 4D-CT images.An initial target block which contains the entire tumor with the size of N×N×N is formed centering this seed.N is defined by doctors’ estimation about tumor size.Then the full search block-matching algorithm is adopted to obtain the most similar target block in the next phase and compute the motion field between them,and so on.We can obtain matching blocks of all phases as well as the corresponding block movement.Next using the displacement and the position of object seeds in initial phase,we can calculate the position of object seeds in other phases,which will be set to the center point of star shape prior.Finally,tumors can be automatically segmented within the blocks via the graph cuts algorithm with star shape prior.Both qualitative and quantitative evaluation results show that this approach can not only guarantee the accuracy of segmentation but also increase automation,compared with the original graph cuts algorithm.Secondly,we put forward multi-phase simultaneous tumor segmentation based on graph cuts in lung 4D-CT data with context information.The images of each phase in lung 4D-CT dataset are jointly constructed into a global graph and the image of each phase is a sub-graph.The context information constraint is added between neighboring sub-graphs,which means edges are added between correspondence nodes.Based on the built global graph of lung 4D-CT data,we construct a new global energy function,which contains regional and boundary terms of all phases,and new energy terms which make use of context information.Region and boundary terms are the same as the original graph cuts algorithm,and the context information constraint comes from the Potts model to penalize the disagreement between the labels of the corresponding voxels.Optimizing the new energy function can accomplish multi-phase simultaneous tumor segmentation.This method further reduces the user interaction.The user only needs to select object and background seeds in the image of one phase and the tumors in all phases can be automatically segmented.The results of experiment on ten different lung 4D-CT datasets,no matter visual or quantitative,both show our method overcomes the graph cuts algorithm without context constraint,and the graph cuts with star shape prior. |
