Necklaces: Inhomogeneous and Point-Enhanced Deformable Models

Abstract In many advanced segmentation problems objects have inhomogeneous boundaries, hindering segmentation under uniform boundary assumption. We present a multifeature image segmentation method, called necklaces, that exploits local inhomogeneities to reduce the complexity of the segmentation problem. Multiple continuous boundary features, deduced from a set of training objects, are statistically analyzed and encoded into a deformable model. On the deformable model salient features are identified on the basis of the local differential geometric characteristics of the features, yielding a classification into point landmarks, curve landmarks, and sheet points. Salient features are exploited within a priority segmentation scheme that tries to find complete boundaries in an unknown image, first by landmarks and then by sheet points. The application of our method to segment vertebrae from CT data shows promising results despite their articulated morphology and despite the presence of interfering structures.