Abstract :
In many cases, the accuracy of statistical pixel classification can be improved by applying a spatially varying prior that can be derived from a shape model. We propose to represent the prior knowledge on the spatial distribution of tissue classes by a distribution of shape particles, each representing one possible distribution of tissue classes. Classification and shape can then be optimized jointly by alternating a particle filtering step, in which the shape particle distribution is evolved under the influence of the current classification, with an update of the classification estimate using the shape distribution. Since a large number of shape hypotheses is used this method does not easily get trapped in local maxima. By applying shape models that are conditional on other, more easily discernible, objects in the image one can perform shape guided classification even if the shapes themselves are hardly visible. The method is demonstrated on the task of detecting aortic calcifications in X-ray images, in which calcifications can only be present inside the aorta - mainly on the aortic wall - but the aorta itself is not visible.
