Semi-automatic segmentation approaches tend to ignore the problems caused by missing or incomplete image information. In such situations, powerful control mechanisms and intuitive modelling metaphors should be provided in order to make them practically applicable. Taking this problem into account the usage of subdivision curves in combination with the simulation of edge attracted mass points is proposed as a novel way towards a more robust interactive segmentation methodology. Subdivision curves provide a hierarchical and smooth representation of a shape which can be modified on coarse and on fine scales as well. Additionally, local adaptive subdivision gives the required flexibility when dealing with a discrete curve representation. In order to incorporate image information, the control vertices of a curve on a certain subdivision level are considered as mass points which are attracted by edges in the local neighbourhood of the image. The usage of this framework for the segmentation of medical data sets shows that the required manual interaction depends highly on the coarsest level defining the shape that serves as a starting point for subsequent segmentation. In order to reduce both processing time and the number of manual interventions, the consideration of model-based information for supplying a good initial shape promises to be very helpful. Hence, the first subdivision model should be determined by a shape statistics based guess and not by the simple application of a subdivision rule. This can be achieved by progressively removing shape variation coded by the position of single points from a statistical shape description. If the correct positions of the corresponding points are known for an object to be segmented, the removed part of the statistic can be used for the generation of a reasonable initial subdivision model. The resulting combination of these two frameworks can be considered as the first steps towards a more robust and practically applicable, semi-automatic segmentation tool.