Theory Of Level Set Method And Its Application To Medical Image Segmentation

With the wide application of medical imaging technology, stable and reliable medicalimage segmentation and positioning method is of great significance in many aspects. Forexample, analyze abnormal anatomy, pathological tissue modeling, intelligent auxiliarydiagnosis, treatment planning, surgical implementation, and treatment effect tracking, etc.Compared with natural scenes，the tissues of the medical image vary greatly in size,shape, position and appearance property. During the development of the disease, theabnormal tissues may be associated with edema, which may be deformed, permeated, andoverlapped nearby abnormal tissues. Meanwhile, the inhomogeneous intensity in the sameregion and the weak edge may be caused by some factors, such as noise, bias field effect,motion, and so on. A robust，accurate and effective segmentation method for medicalimages is still a challenge task.Recently, level set method has been widely researched and developed in the field ofimage processing. During the level set segmentation, the evolving curve implicitlyrepresented by the zero level set of high dimension level set function. The level set methodcan naturally adopt its topological structure changes. Meanwhile, it is theoreticallysupported by some strong mathematical theories and can be easily extended to highdimensional data sets. Thus, level set method has become a research hotspot in the field ofthe medical image segmentation. Considering of the characteristic of different medicalimages, level set method should be enhanced and improved for medical imagesegmentation.The research focus on level set theory and its application in medical images. Someexpanding methods are used for medical image segmentation. The main worksaccomplished in this thesis include:1. A geodesic region-based variational level set segmentation method withoutre-initialization is presented in the energy respect. Curvature anisotropic diffusion is firstlyused for edge preserving smoothing. A new skull-stripped approach, which incorporates amorphological method, is implemented in order to reduce the affection of the skull. The energy function consists of the region term, local geodesic term and the distance-penalizingterms. The region term attracts the evolution curves to segment the image into differentregions, which have the minimal intensity difference among the regions. A geodesic lengthterm employing the edge gradient function is used to attract the curve to the boundary andreduce the dependence of intensity homogeneity. Moreover, an optimality penalizing termhas been used to keep the level set function remaining a signal distance function.Thus, thismethod can drive the curve converge to a desired boundary for the images with complexobjects and avoid the time-consuming re-initialization step of the level set.2. A multiphase variational level set method for image segmentation based on featurevector space is developed. This model presents a multiphase level set method using afeature vector space for automatic segmentation of normal brain tissues. The vector imageconsists of several main properties including intensity, clustering spatial consistency, andlocal texture properties. Clustering spatial consistency describes a spatial informationchannel, which can decrease the affection of the intensity inhomogeneity. Meanwhile, atolerance local binary pattern model and its histogram filter are used to produce a texturechannel. The feature components of the vector image are utilized to optimize the movementof the pixels during the numerical solutions of level sets. Results demonstrate that theproposed method is very effective for providing the feature properties of the medicalimage.3. A discrete optimization level set method is proposed. A Fast level set based ondiscrete optimization method is presented through the analysis of the Fast marchingnumerical methods. The algorithm relies on the assumption and equation of fast marchingmethod. Assume the front always moves “outside”, a weight-grid are constructed forobtaining the arrival time of the front as it crosses each point. The arrival time is computedby a fast shortest path algorithm, which is an optimal sorting technique for sparse graph.The position of the expanding front is determined by the arrival time. The edge detector ischaracterized as the local edge properties. Meanwhile, the global properties are describedby the intensity difference between the point and the foreground. The desired segmentationresults can be obtained by user interaction.4. A variational level set segmentation formulation based on a signal model is presented for images in the presence of intensity inhomogeneity. The bias field productsunsuccessful segmentation results by using the region-based level set method, which isassumed that the images are piecewise constant models. To overcome the difficultiescaused by the intensity inhomogeneity, the image model and local clustering properties areused to reduce the segmentation field from the entire domain to a local neighborhood. Thebias field, the clustering centers and the corrected image are estimated by the iterativeprocessing. Two data fitting terms which incorporate the local clustering properties into theglobal region information are defined to construct the energy function together. Experimentresults on biological images show desirable performance and demonstrate the effectivenessof the proposed algorithm.