Multiscale Shape Priors for 3D Branching Structures in Vasculature Segmentation

In this paper we present a multiscale non-parametric shape model for the segmentation of 3D branching structures such as vasculature or biological neuronal networks. The proposed technique models these structures as surfaces formed by varying-sized interconnected cylinders using higher order active contours. A geometric shape model is built using this approach and incorporated into a surface estimation problem. The model is defined as a geometric cost function that favors the formation of 3D branching shapes when minimized. The resulting functional is incorporated into an active contour framework that is evolved using level set methods. The proposed method is applied to vasculature segmentation in images obtained from 3D digital subtraction angiography. This technique is novel in that it extends previous higher order active contours models for the modeling of branching structures in 3D; it uses the geodesic metric instead of the Euclidean to define the region of interaction between boundary subsets of points; and it defines a shape model that is multiscale, which was not possible with previous higher order active contours models.