The initialisation of segmentation methods aiming at the localisation of biological structures in medical imagery is frequently regarded as a given precondition. In practice, however, initialisation is usually performed manually or by some heuristic preprocessing steps. Moreover, the same framework is often employed to recover from imperfect results of the subsequent segmentation. Therefore, it is of crucial importance for everyday application to have a simple and effective initialisation method at one’s disposal. This paper proposes a new model-based framework to synthesise sound initialisations by calculating the most probable shape given a minimal set of statistical landmarks and the applied shape model. Shape information coded by particular points is first iteratively removed from a statistical shape description that is based on the principal component analysis of a collection of shape instances. By using the inverse of the resulting operation, it is subsequently possible to construct initial outlines with minimal effort. The whole framework is demonstrated by means of a shape database consisting of a set of corpus callosum instances. Furthermore, both manual and fully automatic initialisation with the proposed approach is evaluated. The obtained results validate its suitability as a preprocessing step for semi-automatic as well as fully automatic segmentation. And last but not least, the iterative construction of increasingly point-invariant shape statistics provides a deeper insight into the nature of the shape under investigation.