Modeling of bifurcated tubular structures for vessel segmentation

Abstract

Segmentation and geometric modeling of blood vessels from medical imaging is important for diagnosis of ischemia and atherosclerosis. Since conventional voxel-based methods are slow and unable to incorporate expert knowledge, machine learning methods are proposed for segmentation of cardiac structures by not only increasing the speed but also learning from the manual annotations. However, to our knowledge, all previous learning-based methods assume a loose combination of tubular structures and do not account for the bifurcation geometry. In this chapter, we propose a novel method for construction of complex lumen vasculature with a focus on explicit modeling of bifurcations for learning-based vessel segmentation. A bifurcation is modeled by using convex hulls to join tubular structures guided by centerlines. Subdivision and boosting-based segmentation are performed to adapt the bifurcation model to the target vessel boundaries. Our experiments show that constructed coronary artery geometry from CT imaging is not only water-tight but also accurate by comparing to the manual annotated ground-truths.