Segmentation of Medical Image Objects Using Deformable Shape Loci

Robust localization and segmentation of normal anatomical objects in medical images require (1) methods for creating descriptive object models that adequately capture object shape and expected shape variation across a population, (2) methods for combining such shape models with unclassified image data, and (3) means for localizing and extracting corresponding objects from the image data using the model. A Bayesian approach is well suited as a general analytic framework for such a process; object shape models are associated with the prior, image information is associated with the likelihood function, and the posterior provides a means for combining models with images in a way that makes it possible to localize and segment normal figural shapes. Prior models that include multiscale medial and boundary analysis are well matched with a Bayesian approach since such models directly capture i) figural shape, ii) inter-figural shape relationships, and iii) boundary shape and location relative to figural shape. Described in this paper is such an approach to model-based segmentation, called deformable shape loci (DSLs), that has been successfully applied to 2D MR slices of the brain ventricle and CT slices of abdominal organs. The method combines the model and image data by warping the model to optimize an objective function measuring both the conformation of the warped model to the image data and the preservation of local neighbor relationships in the model. Large scale medial information stabilizes the localization of the object boundary region in the presence of image disturbances such as noise, blurring, and poor contrast resolution. Methods for forming the model on which the prior is based and for optimizing the posterior are described.