A Pseudo-distance for Shape Priors in Level Set Segmentation

We study the question of integrating prior shape knowledge into level set based segmentation methods. In particular, we investigate dissimilarity measures for shapes encoded by the signed distance function. We consider extensions and improvements of existing measures. As a result, we propose a novel dissimilarity measure which constitutes a pseudo-distance. Compared to alternative approaches, this measure is symmetric and not biased toward small area. Based on this pseudo-distance, we propose a shape prior for level set segmentation methods which is pose invariant. In numerical experiments, we demonstrate that the resulting prior permits the segmentation of corrupted versions of a familiar object which is independent of the pose of the object. Moreover, we demonstrate the advantage of the symmetric formulation of the dissimilarity measure when segmenting corrupted images of known objects which consist of multiple components.