Monte carlo sampling for the segmentation of tubular structures

Abstract

In this paper, we present a multiple hypotheses testing for the segmentation of tubular structures in medical imaging that addresses appearance (scanner artifacts, pathologie)and geometric (bifurcations) non-linearities. Our method represents vessels/tubular structures as sequences of state vectors(vessel cuts/cross-sections), which are described by the position of the corresponding plane, the center of the vessel in this plane and its radius. Thus, 3D segmentation consists in finding the optimal sequence of 2D planes normal to the vessel’s centerline. This sequence of planes is modeled by a probability density function (pdf for short) which is maximized with respect to the parameters of the state vector. Such a pdf is approximated in a non-parametric way, the Particle Filter approach, that is able to express multiple hypotheses (branches). Validation using ground truth from clinical experts and very promising experimental results for the segmentation of the coronaries demonstrates the potential of the proposed approach.