SKELETONIZATION OF TUBULAR OBJECTS USING SUPERQUADRIC ELLIPSOIDS

Open Access
Author:
Amjad, Syed Majeed
Graduate Program:
Computer Science
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
None
Committee Members:
  • Sukmoon Chang, Thesis Advisor
Keywords:
  • gradient vector flow
  • skeletonization
  • super quadrics
  • tubular objects
  • deformable objets
Abstract:
Anatomical structures contain various types of curvilinear or tube-like structures such as blood vessels and bronchial trees. Such tubular structures can be best represented by their skeletons. A skeleton of an object is the loci of center points of all maximally inscribed discs (or spheres in 3D). Skeletons reduce the complex 3D pixel-based representation of an object into a simpler 1D spatial line while pre- serving its boundary and region information. A skeleton should be thin, centered, and correctly connected and should preserve the topology of the object. Many algorithms for skeletonization rely on the boundary information of an object to extract its skeletons. These methods are not feasible for medical images, where the boundary is not known a priori or the boundary is not clearly de ned. For medical images, therefore, such methods require extensive pre-processing steps to identify the boundary of the objects of interest. Furthermore, these methods are sensitive to boundary noise. More recent approaches use ridge point detection for nding skeletons. Although these methods do not need boundary information they require complex grouping schemes to ensure the ridge points are correctly connected. These methods also require intensive user interaction, such as manual selection of seed points, which hinders the method from automation. In this thesis, we propose a novel skeletonization algorithm using deformable models. The proposed algorithm requires minimal user interaction and no pre- processing steps to obtain boundary information. The algorithm generates thin, connected skeletons that preserve the topology of the object. Experiments on data show that the proposed algorithm extracts skeletons in sub-voxel accuracy, and results on clinical data also show accurately centered skeletons.