Active contour based segmentation in medical imaging

Date of Completion

January 2005


Mathematics|Engineering, Biomedical|Engineering, Electronics and Electrical|Artificial Intelligence|Computer Science




Active contours have been extensively used in image processing and computer vision. The existing active contour models can be broadly classified as either parametric active contour or snake models or geometric active contour. In this research, we investigate some fundamental and important issues in these two types of active contours and apply our methods to different modalities of images, with emphasis on medical images. ^ For parametric active contours, we propose an improved gradient vector flow (GVF) as the external force, which has a desirable edge-preserving property. We call this vector field edge preserving gradient vector flow (EPGVF). In snake models, automatic initialization and topological changes are difficult problems. We solve these problems by segmentation of external force field. The segmented force field is then used for automatic initialization and splitting of snakes. To segment the external force field, we represent it with a graph, and a graph-theory approach can be taken to determine the membership of each pixel. ^ In traditional geometric active contour models, the level set function has to be periodically re-initialized during the evolution, in order to maintain stable evolution and usable results. The re-initialization procedure has serious problems, such as when and how to re-initialize, and there is no answer that generally applies to date. Moreover, the re-initialization is computationally expensive and can cause errors in computation. We present a new level set method that completely eliminates the need of re-initialization. Our level set evolution is derived from an energy functional minimization problem. The energy functional consists of an internal energy term that penalizes the deviation of the level set function from a signed distance function, and an external energy term that drives the motion of the zero level set. Compared with traditional level set methods, our method has several advantages, such as faster convergence, more accurate computation, and more efficient and flexible initialization. Moreover, our level set formulation can be easily implemented with an efficient and stable narrow band level set evolution algorithm. Our level set method can be extended to 3D and higher dimension. ^