Knowledge-Based Active Contour And Its Applications In Prostate Segmentation

Image segmentation is a fundamental task in computer vision. It is hard to get thesatisfactory segmentation results when the images are corrupted by noises, clutters andocclusions. Model-based segmentation is proposed to address this kind of problem, whichinvolve both of the low level visual information and prior knowledge of the object tosegment the object. It is very similar to the human perception that the low level visualinformation is reconstructed in human brain using prior knowledge to get a meaningfulresult. This dissertation studies the prior model-based segmentation and its applications inprostate segmentation. The primary works of this dissertation are as follows:(1) A new nonparametric region-based active contour driven by local histogram fittingenergy is presented. The distribution of each pixel in the image is characterized by localhistogram computed within a patch centered on the pixel. A local histogram fitting energyis defined in terms of an evolving curve and two fitting histograms that approximate thedistribution on the inside and outside of the evolving curve locally by a truncated Gaussiankernel. The evolution direction of a pixel on the evolving curve is decided by thecompetition of the local histogram and local fitting histograms of the pixel. The kernelwidth of the truncated Gaussian kernel decides the computation range of the local fittinghistograms. The kernel width for computing the fitting histograms should be different ondifferent pixels, since the same kernel width applied may cause local minima of the localfitting energy. Three inequalities are introduced to determine whether larger kernel widthshould be considered. Experimental results show desirable performances of this method.(2) In a shape-based active contour model, it is an important procedure to balance theinfluence of shape prior and conventional segmentation energy terms. Usually, a weightfactor is utilized and manually selected according to different image content and shapeprior. However, the range of the weight factor is from zero to infinity, and there is noguideline to decide how to tune it. An automatic and adaptive method is present to find theweight factor iteratively in this paper. An equality constraint for shape prior is presented tooptimize the energy function of conventional active contour model which is the objective function. Thus, the shape-based active contour is formulated by a constrained variationalmodel. Euler Equation and Lagrange Multiplier Rule are applied to solve this kind ofproblem. Experimental results show the ideal weight factor is the Lagrange Multiplier thatcan balance the impact of the two energy terms on the evolving curve automatically andadaptively when we combine the conventional active contour model with shape model.(3) A nonparametric statistical shape model using kernel density estimation is presentedto model the shape deformation. In a statistical shape model, Principle Component Analysisis applied to extract the shape feature, and then a Gaussian distribution is applied to modelthe shape deformation. If the shapes in the training shape set are different from each othertoo much, the Gaussian distribution cannot model the shape correctly, since it is notpossible to use a parametric method in such kind of circumstance. Kernel densityestimation is applied to model the shape deformation, and the influence of the kernel widthis analyzed when using a Gaussian distribution in kernel density estimation. Experimentalresults show the kernel density estimation is much better than the Gaussian one whenmodel a complex shape.(4) The deformation of a shape model is driven by appropriate feature, so if a bad featureis applied to deform the shape model, the deformation result of the shape model will beincorrect. An appropriate feature to drive the deformation of a nonparametric shape modelis presented to segment prostate in Ultrasound image. A principle is defined to select thefeatures, and the appropriate feature is applied to deform the nonparametric statisticalshape model. Experimental data consist of12patient image sets, each with6slices of theprostate ultrasound image. We leave one slice each patient for test the algorithm and use theothers to model the shape. Experimental results show the efficiency by using theappropriate feature to deform the shape model.(5) A nonparametric statistical method to model the shape or appearance of an object ispresented. Firstly, Principle Component Analysis is applied to extract the features, and thenkernel density estimation is applied to model the features. The presented method is utilizedto build the shape and appearance model of three dimensional prostate in CT imagerespectively. Then, the constrained variational method joins the appearance model andshape model together. Experimental results show a fine segmentation result will beachieved by place the shape model to the area of the real prostate.