Ncut And Medical Image Segmentation

This thesis builds on previous work on image segmentation by using spectral graph theory,and focuses on the usage of segmentation in brain medical images.First,image Segmentation is the important and initial step in the processing of images.For instance,given a digitized image containing several objects,the pattern recognition process consists of three major phases,and the first phase is image segmentation.Image segmentation process can be defined as one that partitions a digital image into disjoint(nonoverlapping) regions.Image segmentation is becoming an important step in medical image processing,which give great support to the latter anlysis of medical images.The research area of this thesis focuses on brain images,which is the most difficult part of human body for image segmentation.The main purpose of the thesis is to find a way for automatic and efficient segmentation for these brain structures.We utilized Normalized Cut as our segmentation criteria,and broaden the affinity weight W to 3-dimensional space,which makes the Ncut method can consider all the points in the image space.The main contributes of this thesis are as follows:1) The definition and the implementation of Normalized cut are introduced,and the improvement of computation eigenvector and eigenvalue is summarized.Finally,the application of Ncut in different fields is emphasized,such as medical image segmentation, vector field segmentation,and video segmentation.2) The thesis presents new approach for how to recognize focus in a series of brain MR images.These images are relevant with each other and the approach proposed in this paper can consider all the features in the 3D space through reconstruction weight matrix while using Normalized Cut.The eigenvectors shows segmentation of the focus in different images,even in some case the focus is not very clear.We also analyze two different feature linking functions for construction weight matrix and the result given in the thesis.3) However,the original Ncut method can not cover large scale images,especially some medical image,like MRI.A new method proposed by Timoth(?)e Cour successfully overcome the shortcoming by compressing different images in a scaled matrix. This thesis continues the research and application of their algorithm by developing and reconstructure the affinity matrix form 2D to 3D. 4) In the end,a conclusion and some direction for further research are discussed.The thesis tributes to recognize the rough object in one's brain MRI images by collecting the pixel information in the 3D space,and propose a new way to reconstructure the affinity matrix,as well as get the eigenvalues and eigenvectors.Our method has reasonablely solved the problem in brain images,but still need more improvement to apply in other medical images.