Research On Segmentation Of Head And Neck Organs In CT Images

The incidence and mortality of head and neck(H&N)cancer are among the highest in the world.Radiotherapy is an important and effective treatment for H&N cancer.In the course of radiotherapy,the radiation dose received by the organs at risk(OARs)around the tumor volume must be strictly controlled,so as to avoid radiation damage and affecting the quality of life in patients after radiotherapy.Intensity-modulated radiation therapy(IMRT)can effectively protect OARs and achieve a three-dimensional dose distribution highly conformal to planned target volume and therefore becomes the most advanced treatment for H&N cancer.In order to give full play to the advantages of IMRT,it is essential to accurately delineate OARs based on planning CT images.However,due to the concentration of important organs in the H&N,as well as the complex anatomical relationship,manual delineation,which is still widely used in clinic,is time-consuming,exhausting and inaccurate.Due to the drawbacks of manual delineation,people hope that automatic delineation by computer will replace the manual delineation by physicians,and a lot of research work has been carried out.However,automatic segmentation of H&N tissues and organs in CT images is very difficult.Firstly,because of the low contrast of soft tissue in CT images,the boundary between tissues is very blurred or even there is without any boundary information.Secondly,due to the concentration of important organs and their complex anatomical relationship,as well as great variations of OARs across individuals,the background for segmentation in H&N CT images is very complex.Furthermore,metal implants in the head and neck of patients,such as dentures,can cause serious metal artifacts in CT images.The above characteristics of H&N CT images make the automatic segmentation of tissues and organs very challenging and still unable to meet the actual clinical demands.In order to improve the accuracy and robustness of automatic segmentation of tissue and organs in H&N CT images,a method based on random forest machine learning and deformable model is studied in this paper.The main contents and innovation include the following three aspects:(1)A novel atlases alignment mthod based on anatomical location prior constraints of tissues and organs.The accuracy of atlases alignment is very important for image segmentation based on machine learning.Traditional atlases alignment is achieved through the registration between intensity images in atlases.Because of the low contrast of soft tissue in CT images,traditional methods can not get ideal atlases alignment results.To solve this problem,the proposed method makes full use of the anatomical location prior information of tissue and organ in label image of atlases.On the one hand,anatomical location prior is combined with gray image information to form a hybrid dissimilarity measure,which serve as an optimization objective function to drive the evolution of spatial transformation model.on the other hand,anatomical location prior is used to construct an initial B-splines transformation to obtain better initial optimization parameters and enhance robustness of alignment algorithm.Comparison experiments show that the proposed method significiantly improves the accuracy of atlases alignment.(2)Hierarchical vertex regression-based segmentation of H&N CT images.The segmentation method based on deformable model can use shape prior to guide and constrain the segmentation process to overcome the problem of low contrast of soft tissue in CT images,and therefore has been widely used in tissue and organ segmentation of CT images.However,it is sensitive to initialization,weak adaptability to individual differences of target organs and ineffective in segmenting organs with complex shapes.In order to solve above problems,this paper presents a hierarchical iterative regression strategy based on the significance of vertices in shape model.The proposed implicit deformable model based on vertex displacement regression avoids determining directions and distances to move vertices,which is common in a traditional explicit deformable model.Thus,the flexibility and fitting accuracy of the deformation model are improved.The proposed machine learning-based critical vertices recognition and location mechanism can find out the critical model vertices and locate them according to the uniqueness of vertices and the consistency between individuals.The proposed learning framework of joint shape and appearance features can simultaneously capture salient shape prior and appearance features into a machine learning model based on random forest.Experimental results demonstrated the improvement in accuracy and robustness in segmenting H&N OARs by the presented algorithm,as compared to the state-of-the-art methods.(3)Non-blind learning-based segmentation of CT images with metal artifacts.In traditional machine learning-based segmentation methods,knowledge only from training atlases is uesd to train learning model.The process of image segmentation blindly apply knowledge model to a new image.Thus,there is only one-way relationship between training and application of model.However,image appearance in artifact area is different to that in normal CT images.Because the knowledge model does not match the actual situation in the target image to be segmented,the forced knowledge application will lead to segmentation errors.In order to solve this problem,this paper proposed a non-blind learning strategy which introduces the information in the image to be segmented into the training process of machine learning model to carriy out targeted model training,and thus establishes the bidirectional relationship between the training and application of the model.The proposed method first detects the artifact region in the image to be segmented,then feedbacks the location information of the artifact region back to the model training process,and then in model training,the intensity feature components from artifact region are forbidden to participate in decision and only shape feature components have the right to make decision.The experimental results show that the proposed method can effectively overcome the influence of metal artifacts on segmentation results.