Vessel Segmentation Using Implicit Model-Guided Level Sets

This paper proposes an automatic segmentation method of vasculature that combines level-sets with an implicit 3D model of the vessels. First, a 3D vessel model from a set of initial centerlines is generated. This model is incorporated in the level set propagation to regulate the growth of the vessel contour. After evolving the level set, new centerlines are extracted and the diameter of vessels is re-estimated in order to generate a new vessel model. The propagation and re-modeling steps are repeated until convergence. The organizers of the workshop report the following results for the 24 testing datasets. The sensitivity and PPV are 0.26, 0.40 for QCA and 0.05 and 0.22 for CTA. As for quantitation, the absolute and RMS differences for QCA are 29.7% and 34.1% and the weighted kappa for CTA are -0.37. As for lumen segmentation, the dice are 0.68 and 0.69 for healthy and diseased vessel segments respectively. Performance for QCA and lumen segmentation are close to the reported by the organizers for three human observers.