This research is a feasibility study of developing a semi-automatic segmentation method for CT images of fresh cadavers. The segmentation method should be efficient in computation and require minimum human intervention. One popular strategy is to detect 3-D surface directly from the dataset. We have developed a facet-model-based surface detector which is capable to extract 3-D surfaces from a series of CT images. This surface detection  method can achieve subpixel accuracy. But further efforts requiring the registration of extracted surfaces of various anatomical entities and processing are not part of this effort.

     Another popular strategy is to extract two dimensional edge points of closed contours from each image first, and then stack these contours together to form three dimensional surface. Using this strategy, we can minimize the computation and algorithm complexity. We have developed an active contour model program to extract surfaces of a particular anatomical entity from a series of CT images. Once an initial contour in one CT image is specified, the contour in the subsequent CT image can be extracted automatically if there is no abrupt change between two contours, as to say, the slice thickness has to be reduced so that such abrupt change will not occur. One by one, all contours of the anatomical entity can be extracted and stacked together to form the surface. By this way, surface extraction and registration can be done in one step, although separate entities need separate initial contours. At the same time, we are also investigating level set method. Comparing with active contour model method, the level set method does not require the initial contour to be close to the desired contour and it has the capability extracting contours of multi-entities in the same image. After contour extraction, it requires set registration, which is much easier than point registration.

     In the next section, the tools we have investigated in our study are introduced and their principles and functions are summarized. In the third section, our processing algorithm are presented in step-by-step details. In the fourth section, the segmentation results are presented and discussed. In the last section, we conclude our study.