We propose a new generic framework for segmentation of 3D digital data, based on knowledge contained in a segmentation example of similar data. The integration of prior knowledge is made by registering the image to segment on the segmentation example. Since the registration step relies on binary segmented data, segmentation and registration are performed jointly in a coarse-to-fine way using a multiscale parametric representation of a threshold map and of a deformation field. The threshold map is required by the segmentation procedure which is also devoted to recover topological details. The benefit of such an approach is illustrated in the context of head bone segmentation in 3-D computed tomography (CT) images.