Research On Liver CT Image Segmentation Algorithms Based On Sparse Representation And Low-rank Recovery

Computer-aided diagnosis and surgical planning systems are playing an increasingly important role in the diagnosis and treatment of liver cancer.CT imaging is widely used in the clinical diagnosis of liver cancer,because of its high resolution and high signal-to-noise ratio.Therefore,the study of accurate and robust liver segmentation from CT images is of considerable significance in both medical research and clinical diagnosis and treatment.However,liver segmentation from CT images is still faced with many difficulties: On the one hand,the boundaries between liver and adjacent organs are fuzzy,and the liver shape varies much between different individuals;On the other hand,the CT images are very susceptible to partial volume effects and noise.In order to solve these problems,the segmentation methods based on global prior knowledge are studied in the CT images,on the purpose of accurate liver segmentation.Meanwhile,we introduce the latest sparse representation and low-rank recovery theories in the signal processing and computer vision communities,and integrate them into the global shape prior knowledge based active shape model and atlas methods to achieve accurate and robust liver segmentation in CT images.Semi-automatic liver segmentation method has the characteristics of high segmentation accuracy and low time cost.In clinical application where there is a specific requirement for the segmentation time,this kind of approach is more suitable than fully automatic method.Since the traditional methods which are based on image or local shape priori information tend to yield large segmentation errors,an active shape model based semi-automatic liver segmentation method that incorporates global shape prior knowledge is proposed.In active shape model method,the existing shape prior knowledge modeling methods have three main issues: high sensitivity to the non-Gaussian errors of input shapes;inability to effectively model complex shapes with non-Gaussian distribution;weak recovery of the local details of input shapes.To this end,a multilevel local region-based sparse shape composition is proposed.When building the shape model,we firstly decompose the liver shapes into multiple regions with homogeneous shape variation in a multilevel fashion,then we build a corresponding local shape repository for each region and refine the input shape as a sparse linear combination of training shapes in the shape repository in a region-by-region manner.To improve the accuracy of liver initialization,we also proposed a blood vessel-based liver shape initialization method,and a hierarchical shape model optimization method.Experimental results based on clinical data and international competition proved that the proposed semi-automatic segmentation method can be deployed for accurate and efficient liver tissue segmentation.Semi-automatic segmentation method can achieve high accuracy in limited time through human interaction;However,due to the need of human participation,the results of semi-automatic segmentation method are affected by human factors,which then affects the reproducibility of segmentation results.In view of the disadvantages of semi-automatic segmentation methods on this respect,an atlas based fully automatic liver segmentation method that incorporates global shape prior knowledge is proposed.The atlas-based segmentation methods can easily lead to large registration errors when the atlas intensity image is aligned to the target image.To this end,a sparse representation theory based deformation model is proposed to correct the registration errors generated during the image registration procedure,that is refining the derived nonrigid transformation as a sparse linear combination of existing training transformation.In order to let the final generated liver atlas reduce bias towards the particular anatomy of the selected initial template image,we also proposed an iterative method to construct the liver atlas.Experimental results based on clinical data show that,the proposed automatic segmentation method can obtain the segmentation accuracy close to the semi-automatic segmentation method without human intervention,and the experimental results are reproducible,thus verifying the effectiveness and advancement of the proposed method.The atlas-based liver segmentation approach can achieve high segmentation accuracy when dealing with liver tissue under healthy or minimal pathological conditions.However,the segmentation accuracy is significantly reduced when faced with liver tissue with severe pathology.Nevertheless,fully automatic methods that incorporate shape prior model can effectively improve the segmentation accuracy of highly deformed liver tissue,in particular,the sparse representation theory based sparse shape composition model successfully solves the main issues of existing shape prior knowledge modeling methods.However,the ?1-norm based sparse representation method lacks grouping effect.This affects the generalization ability of the model to a certain extent.In view of the model defect,a low-rank and sparse decomposition based shape prior model is proposed.The model overcomes the shortcomings of the traditional shape prior models,and can accurately recover the true subspace of the liver shape matrix containing sparse gross errors.The model decomposes the liver shape matrix into the following three parts: a low-rank part of the global liver anatomical structure,a sparse part of the sparse gross errors,and a part of the small dense Gaussian noise.Experimental results demonstrate that the proposed method can be deployed for effective liver shape prior modeling,especially for the shape of liver tissue with severe pathology,thus it improves the segmentation accuracy and segmentation stability of the automatic segmentation method to a certain extent.The construction of the low-rank and sparse decomposition based shape prior model provides a model basis for the highly accurate and fully automatic segmentation.In order to achieve accurate segmentation of the liver tissue with severe pathology,an active shape model based automatic liver segmentation method is proposed.When segmenting the liver tissue with severe pathology,we face the following three additional challenges: the regions with large lesions exhibit completely different intensity values from the normal liver tissue;the existence of low contrast between peripheral liver lesions and adjacent organs;the CT images are very susceptible to imaging artifacts.To this end,a patient-specific low-rank and sparse decomposition based probabilistic atlas for shape initialization is proposed,to largely eliminate the negative effects of pathology on the final constructed probabilistic atlas and liver likelihood image.In order to accurately recover the local details of input shapes,we also used the low-rank and sparse decomposition based shape prior model to construct a population-specific shape prior model,and finally we proposed a hierarchical active shape model search method for effective model optimization.Experimental results based on clinical data and international competition show that the proposed method can be deployed for accurate and robust pathological liver tissue segmentation.