Study On Liver Segmtation Method From CT Images Based On Deformation Optimization And Sparse Statistics

Liver is the primary digestive organ that circulates and detoxifies blood though human body. It is also the organ where lesions occur frequently. Recent statistic has confirmed that the liver cancer is about to surpass cancers of other visceral organs and rank the second vital lesion. The vitality of the lesion attracts worldwide concerns about the preventive measures and efficacy of the treatment. Computational Tomography(CT) is a clinically recognized approach that provides superior resolution of anatomic structure to assist the detection of lesions. These lesions usually carry graphic features of inhomogeneous grayscale and coarse edge that can hardly be distinguished precisely without the usage of CT imaging. The outline and structure of liver lesions can be detected by segmenting the CT images and are of significant values to functional evaluation, lesion identification and surgical treatment. However, the location of liver is tightly close to its neighbor visceral organs that causes low divergence in grayscale. Besides, the structure and spatial position of liver tend to vary elastically. So far, accurate outline extraction of liver segmentation and detection from CT images still remains as a challenge in medical imaging.The focus of this research is to combine an elastic model, sparse encoding and CT image segmentation based on a statistical shape model and construct an adaptive surface expansion model. The combined method incorporates an A-priori sparse lookup dictionary and three segmentation methods on the statistic shape model. Experimental results revealed that the segmentation methods successfully partitioned the CT images of liver with high precision, and the sparse model based approach outperformed over the other two. The original contributions of this thesis include the followings:(1) An adaptive surface expansion and triangular regional optimization based liver CT image segmentation is proposed to improve the segmentation precision, in particular, of the depressed regions on the liver, where the conventional approaches that consider local optimization are incapable of handling this issue. The proposed approach reconstructed the initial boundary of the liver using a simple surface model. The internal force and constraint relations are formulated according to the geometrical relationship between the vertex of the analytical model and the vertices in the neighborhoods. The constraint for the external force is built upon the Gabor boundary characteristic of the liver CT image, with a Balloon force induced in the external force of the model. Moreover, an optimized self-tuning algorithm that adaptively inserts new vertices during the deformation was attempted in this study and fulfilled an increased accuracy of deriving the fine structure of the liver.(2) A-priori sparse dictionary and cavity filling based liver image segmentation was investigated to resolve the segmentation efficiency problem due to a huge number of training samples. In order to minimize the sample quantity, the proposed method first registers the undivided images with the images that satisfy with the golden criteria of segmentation to identify the boundary of liver which is adopted as the initial boundary of the undivided images. The choices of test sample sets are from the neighborhoods of the initial boundary. The Gaber feature and grayscale images in the training image set are used to compose the lookup dictionary for liver boundary characteristics. The calculation of sparse coefficients and reconstruction errors are based on a combined test sets and the lookup dictionary. The cavity filling using the liver surface data is brought about to ensure the smoothness of the segmentation.(3) A statistic sparse shape model is constructed according to an A-priori shape data of the liver boundary, aiming to solve the challenging issue of registration between the current stochastic shape model and the undivided model. Results found that the registration accuracy is increased by an order of magnitude. This is as a result of an A-priori shape model vertex based dictionary in combination with the undivided images and the corresponding sparse encoded reconstruction and stochastic shape model in the A-priori shape model. The specific energy functions are constructed using the grayscale and boundary data from the vertex of the stochastic model on the undivided images. Finally, the sparse matching constraint model that combines the lookup dictionary and the grayscale data of the undivided images are built and tested to drive the stochastic shape model converging to the clearly partitioned liver boundary.