Tubular Objects in 3D Medical Images: Automated Extraction and Sample Applications

In this paper we present and evaluate a novel technique for generating representations of tubular objects in 3D medical data. Tubular objects are abundant in medical images, e.g., vessels, bones, ducts, spinal cords, and bowels. Tubes can be characterized as smoothly varying, yet possibly branching, structures in 3D that have nearly circular cross sections. While other techniques have been suggested for segmenting tubular objects, our method rapidly generates accurate and consistent tubular representations with minimal user interaction by exploiting the geometry of tubes. Specifically, tubular objects defined via contrast are special in that blurring produces a central intensity ridge that well approximates the objects’ central skeleton. Our method operates by traversing those central skeletons. Once extracted, those central skeletons also serve to stabilize a width estimation process. Our method is also advantageous because the form of the representations it generates (i.e., central skeleton and widths). These representations enable abstract reasoning and hence a variety of clinical applications. These representations can be easily joined, split, quantitatively evaluated, and viewed in isolation. As examples, we illustrate the formation and manipulation of vascular and bronchial trees for surgical planning, and we highlight our work regarding the use of our representations for multi-modal, multi-dimensional registration.