Many techniques of knowledge-based segmentation consist of building statistical models that describe the deformations of the structure of interest, and then fit these models to the image data. In this paper, we introduce a novel family of shape prior models that aim to capture such varying support. To this end, 3D segmentation is considered progressively with 2D slices segmented in a qualitative fashion, starting from the ones with strong data support toward the ones of limited support.

more »

... ve segmentation maps are linked through a locally adaptive autoregressive prediction mechanism -that is learned through training -where confidence of the data from prior slices constrains the results. Such prediction is integrated with a contour minimization technique, leading to a Bayesian sequential procedure that iteratively predicts and corrects 2D contours leading to complete reconstruction of 3D anatomical structures. A quantitative comparative study with 3D Active Shape Models demonstrate the potential of the method.