Globally Optimal Active Contours, Sequential Monte Carlo and On-Line Learning for Vessel Segmentation

摘要

In this paper we propose a Particle Filter-based propagation approach for the segmentation of vascular structures in 3D volumes. Because of pathologies and inhomogeneities, many deterministic methods fail to segment certain types of vessel. Statistical methods represent the solution using a probability density function (pdf). This pdf does not only indicate the best possible solution, but also valuable information about the solution's variance. Particle Filters are used to learn the variations of direction and appearance of the vessel as the segmentation goes. These variations are used in turn in the particle filters framework to control the perturbations introduced in the Sampling Importance Resampling step (SIR). For the segmentation itself, successive planes of the vessel are modeled as states of a Particle Filter. Such states consist of the orientation, position and appearance (in statistical terms) of the vessel. The shape of the vessel and subsequently the particles pdf are recovered using globally active contours, implemented using circular shortest paths by branch and bound that guarantees the global optimal solution. Promising results on the segmentation of coronary arteries demonstrate the potential of the proposed approach.