This paper describes a technique for using prior knowledge of shape variations to help guide a volumetric deformation process. The volumetric transform maintains the topology of a template while matching the template to an image under study. A statistical model is used to describe inter- and intra-shape correlations in the template. The parameters for the shape model are learned by performing eigenshape analysis on a training set consisting of deformations of a single template to several typical segmentations. The shape model is used to guide the deformation by the inclusion of a term to the deformation cost functional that promotes the most likely deformations according to the shape priors. Some advantages of the proposed method are that it inherently conserves the topology between multiple shapes, and that prelabeling of corresponding points and point ordering of the training set is not needed. Results are presented for segmentation of magnetic resonance and cryosection images with varying contrasts. A qualitative analysis shows that the inclusion of shape priors can significantly improve the final deformation result.
|Original language||English (US)|
|Number of pages||6|
|Journal||Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition|
|State||Published - Jan 1 2000|
ASJC Scopus subject areas
- Computer Vision and Pattern Recognition