A Novel Shape Prior Segmentation Method Based On Levelset Model And Its Application In Liver Segmentation

Medical image segmentation is a fundamental problem in medical image processing and analysis, and is the basis of computer aided diagnosis and treatment as well. The general segmentation problem is the process of partitioning an image or data-set into a number of uniformity or homogeneous segments. Image segmentation is important for medical image analysis, for example, 3D Visualization, Computer Aided Operation, and Radiology Treatment all assume that Region of Interesting (ROI) are well segmented. In the project of liver perfusion in our lab, the first step is liver segmentation. Perfusion analysis is based on the segmentation result.Level Set Method is a popular and widely used algorithm in present Image Segmentation Area. Its basic idea is to construct an energy function for the model and let the curve evolve under the model's inner control force and outside image force. When the energy function reaches its minimum value, the evolving curve will describe the target region. Therefore, image data, initial contour and target contour are included in one uniform mathematical mode. Level Set Method is independent of detail parameters during the evolving process. The evolving curve or surface can be represented as the zero level set of a higher dimensional function, which can deal with the topological change of ROI automatically.However, there are still some shortcomings in Level Set Methods, such as under-segmentation, over-segmentation and leakage problems. Here shape prior knowledge is employed. The Level Set Methods based on shape prior model improve the segmentation precision greatly through introducing the prior shape information of the area to be segmented. Nevertheless, some new problems emerges that in traditional shape prior models, the prior shape is required to have the same scale, position and pose as the target object. Otherwise the segmentation result will not be satisfying. Under this background, we propose a novel model, where we get an initial prior shape through statistical methods, then train the initial prior shape with gradient descend method, regulating it to an appropriate scale, pose and position. Thereby it can guide the segmentation process better. In real applications, our model is more practical than traditional ones. Besides that, we also make some algorithm optimization, which improve the segmentation velocity, precision and robustness.The main innovation points are described as below:1. Introduce shape prior knowledge into CV model (a popular model based on Level Set), and provide the numerical deduction and program implementation.2. Improve the Cubic Spline Function, thus we can get smooth curve when sample points are given. Afterwards we use some statistical methods to get the initial prior shape.3. With the help of Chamfering method, Feature Image and the Fast Solution, we improve the speed remarkably.4. Construct the measurement of Gradient Descend Method, thereby we implement the rigid-body registration training between initial prior shape and initial segmentation result. Therefore our model becomes more practical.