Knowledge-guided segmentation of 3D imagery

This paper presents a computationally efficient and robust approach to locate, label, and isolate three-dimensional (3D) structures from discrete 3D imagery. The emphasis is placed on extracting a convex, singly connected 3D structure of interest imbedded in discrete, volumetric data sets that are sparse, noisy, and possibly misleading. In particular, we focus on the segmentation of man-made objects (phantoms) imaged tomographically and on the segmentation of the myocardial mass obtained in 3D nuclear cardiac imagery. The salient characteristics of this method are that the segmentation process is accomplished in a fully automated fashion and that the volume of interest can be isolated and labeled even in cases where ambiguities or structural incompleteness are inherent in the original imagery. The method presented here can be viewed as a knowledge-guided approach that is iterative and self-correcting, and that is shown to consistently evolve toward incrementally refined segmentation solutions that are quantitatively and qualitatively accurate. The approach innovatively combines image analysis techniques with morphological, knowledge-based, and model-based grouping operations in a highly integrated fashion. In the subsequent discussions, we fully describe the underlying mathematical and algorithmic details of the approach, and discuss the results obtained from its application in numerous experimental studies. (C) 1996 Academic Press, Inc.