Traditional segmentation systems apply frequently a labeling methodology using a voxel representation as geometric model. This descriptor is of extraordinary simplicity and does not differ from the original volumetric data representation. For visualization purposes, the labeled voxels can be transformed into a polygonal surface using the well-known Marching Cubes algorithm ([8]). Since the resulting meshes consist usually of a huge number of triangles, an additional mesh decimation step ([9]) is essential for further usage.

     Another equally important segmentation approach employs a boundary description which can be considered as a dual representation of the aforementioned region based model. The set of methods based on boundary representations can be subdivided into two-dimensional and three-dimensional approaches varying only in the underlying geometric description of the model. While the 3-D approach uses directly a spatial model, the 2-D algorithms based on arbitrary curve representations work in a slice-by-slice manner, making a reconstruction step absolutely essential. This postprocessing operation, connecting the stack of contours to a single three-dimensional model, has been studied extensively ([10], [11], [12], [13]) and proved to be an awkward problem not being solvable in a unique way. One possible workaround is to convert the boundary representation into a voxel data set and to proceed as mentioned before. Since the reconstruction problems do not occur while working with a spatial model, present day research focuses the longer the more on the employment of surface representations.

     Considering the interaction metaphor of state of the art segmentation methodologies, one can divide up existing algorithms into three classes:

• fully automatic
• semi-automatic
• manual

     Alternatively, manual segmentation is highly insensitive to noise, tolerant of missing information and sufficiently precise. The drawbacks, however, are those of being horribly time consuming and tedious, and not to mention the absence of any reproducibility. Nevertheless, in many ambiguous situations it is an advantage, if comfortable manual editing of the geometric models is possible.

     In between these two extremes, the computer vision community has developed a whole family of semi-automatic tools ([4], [14], [3], [15]), combining the advantages of computational support by precise border detection with the benefits from manual manipulation possibilities. This man-machine co-operation implies some kind of collaboration between the human operator and the computing engine, wherefore members of this class of algorithms are also called interactive.

     Considering all these possibilities, we have chosen the following strategy to build a highly detailed and consistent anatomical model of the abdominal cavity:

     In spite of all the advantages of a three-dimensional approach, we relied on the traditional slice-by-slice technique using the familiar two-dimensional outlining methodology for the following reasons:

• It is much simpler for the user to control, edit and interact with a two-dimensional model, since the degrees of freedom grow exponentially by switching to a three-dimensional model, potentially overwhelming the sculpturing capabilities and the powers of imagination of a human operator.
• It is much simpler to implement a two-dimensional user interface using state of the art desktop metaphors than grafting on a three-dimensional interaction methodology to existing mouse-based user interfaces.
• Human perception is not designed for discerning volumetric data. Volume rendering techniques are one promising approach to alleviating this problem, but the computational complexity of these algorithms is still very high and interactivity is at present hard to achieve.

     Therefore, we decided to rely completely on manual outlining based on cubic interpolating B-splines. Figure 1 illustrates the user interface of the developed segmentation package. To establish a powerful segmentation environment, we implemented a complete multi-user segmentation system with an underlying anatomical database. Special attention was paid to an intuitive user interface and very fast data access. To reduce the requirements on our computational infrastructure, we cropped the original images and reduced the spatial resolution by a factor of two.