Examination Committee

Prof Chiew Lan TAI, CSE/HKUST (Chairperson)
Prof Albert Chi Shing CHUNG, ECE/HKUST (Thesis Supervisor)
Prof Weichuan YU, ECE/HKUST

Abstract

Curvilinear structure analysis is the foundation of a wide range of applications, for instance image enhancement, vascular segmentation and surface reconstruction. Specifically in medical imaging, the appearance, morphology and topology of curvilinear objects can be important indicators of many vascular diseases and systemic disorders. Thus, the analysis of curvilinear structure can be of great diagnostic value. 
 
In this thesis, we first develop the curvilinear structure descriptor - Optimally Oriented Flux (OOF) for three dimensional curvilinear structure analysis, namely inter-scale OOF analysis. OOF attempts to find an optimal axis along which the image gradients are projected prior to quantifying the amount of gradient flows. The performance of conventional OOF suffers from overshooting problem and circular cross-sectional assumption. Therefore, in the proposed method, responses obtained in different radii are competed to eliminate the disturbances introduced from the nearby non-curvilinear structures. The residual responses are acquired separately along the object normal and tangent spaces in order to benefit the analysis of eccentric structures. Further, local orientation coherence is enforced to generate a more consistent OOF vector field. With the inter-scale OOF analysis, three important characteristics of curvilinear structure can be estimated simultaneously: curvature, dimension and orientation. We experimentally demonstrate that the new descriptor delivers more accurate and stable detection under the interference of adjacent high-contrast structures and exhibits extraordinary noise robustness.
 
Vascular segmentation is one important application of curvilinear structure analysis. In order to evaluate the capability of inter-scale OOF and to reversely facilitate the understanding of vasculatures, in the second part of this thesis we propose to feed the descriptor into two existing general frameworks for vascular segmentation, which are Random Walks and Continuous Max-Flow. The proposed methods are evaluated on public available databases and achieve comparable segmenting results.