Research On Nonlinear Statistical Shape Analysis And Complex Medical Image Segmentation

Medical image segmentation, with its aim to identify the object of interest, is notonly one of the difficult and hot spots in medical image processing, but also animportant part of computer aided diagnosis and is significant in the clinical diagnosis,pathologic analysis and treatment of diseases.Retinal optic nerve diseases (such as glaucoma) are caused by a variety of retinaland neural tissue lesions, and ultimately lead to blindness. These diseases have highincidence rate in all over the world, and most of them have no clinical signs at earlystage. Therefore, diagnosis and treatment in early stage are very important. Studiesshow that the segmentation and measurement of the colour fundus image of opticnerve heads played a key role in clinical diagnosis and screen for glaucoma.However, due to the poor quality (such as non uniform illumination, low contrast,obscur due to blood vessels), distinct inter differences of individuals and other reasons,image information may not sufficient to segment the optic disk from the colour imageof the optic nerve heads. Therefore it is indeed necessary to integrate the statisticalshape prior into the image segmentaion medol. As a result, the key to the problem liesin finding an effective shape prior representation and integrating it into thesegmentaion medol.In classical linear statistical analysis of an object shape,2order statisticalinformation was considered only, while the high order statistical dependencies such asthe relationship among three or more pixels were ignored, which marrs itsperformance in practical opration.In this paper, a vector valued Mumford Shah model with nonlinear statisticalshape prior for complex medical image segmentation was proposed, and then wasapplied to the segmentation of the optic disk in the colour image of the fundus ofglaucoma patients at different stage. Firstly, we represented the training shapes usingthe narrow band level set method and mapped the narrow band level set of shape priorinto its kernel space by a nonlinear kernel function. Then, the Principal ComponentAnalysis (PCA) was performed in the kernel space so as to acquire its base vectors (i.e.deformation model). Moreover, the nonlinear statistical shape prior of the object ofinterest can be built and was integrated into the Mumford Shah model forvector valued image segmentation. The experimental results on segmentation of the optic disk in optic nerve headimage showed that the proposed model is effective and practicable for thesegmentation of the low contrast optic disk obscured partly by blood vessels in colouroptic nerve head images of different stage glaucoma patients.