Three-dimensional point distribution models for tubular objects

Landmark based statistical shape models are becoming increasingly popular in medical image analysis. These models require a set of training shapes in which corresponding landmarks are indicated. If a limited number of training shapes is available while a large number of landmarks is needed to describe a shape properly, a model based on observations in the training set alone may be too restricted. This will often be the case in three-dimensional shape description. This paper describes the construction of a three-dimensional shape model of a tubular object. Two approaches for generalizing this model are proposed. First, the model flexibility is increased by modeling the axis deformation independent of the cross-sectional deformation. Second, supplementary cylindrical deformation modes are added to the model. The methods are evaluated on model construction of an abdominal aortic aneurysm from segmented CTA data. In leave-one-out experiments on 23 datasets, shape approximation errors are successfully reduced using the two shape model extensions. The combination of statistical and synthetic deformation modes performs better than either of the two models alone.