Description |
-
Abstract: In this dissertation, we investigate the identification and application of geometrical graph-based models (GGMs) of tubular structures with a focus on the vascular networks. Our initial contribution is in the development of a method that directly extracts microvasculature from highly artifacted raw 3-D fluorescence microscopy images. This method comprises two novel initialization and con... read morestrained recovery and enhancement stages. The approach is fully automated using features derived from bi-scale statistical measures and produces results robust to non-uniform illumination, low SNR, and local structural variations. We next introduce a GGM-based method that identifies a piece-wise linear skeletal approximation of a microvascular network that merely requires a rough segmentation of the structures. The nodes of the graph represent the critical points (CPs), defined as locations of large structural deformation and detected with template and convex hull filterings that are independent of any a priori geometric and probabilistic information such as direction, degree, or intensity distribution. The anatomical connectivity of the CPs is derived by solving a binary integer program whose utility function reflects both intensity profile and structural information of the vasculature along the edges. In a "divide and conquer" manner, we have designed a graph interpolation technique that extends applicability of the GGM identification method to larger data sizes. Finally, the GGMs are employed to non-rigidly register cranial artery networks which is formulated as a homologous landmarks guided point correspondence problem. We have developed a novel collection of features, which we call a "signature," that captures geometrical attributes of nodes (location of junctions) and edges (length and curvature of vessels) in a topologically encoded form. Using this signature, we formulate registration as a Linear Assignment Problem (LAP) rather than the more commonly employed (and NP-hard) quadratic assignment problem. Using signatures allows us to relax the combinatorial problem to a convex form that results in a profound computational complexity reduction. By solving the LAP via a graduated assignment technique, nodes are first matched, and then the edge correspondences are determined using a heuristic approach. The performance of this method is tested using clinical angiography images and synthetic data sets. Quantitative results suggest that this method is highly reliable under the influence of different perturbing factors that turns it into a potential technique for inter-subject and multi-modal registration.
Thesis (Ph.D.)--Tufts University, 2015.
Submitted to the Dept. of Electrical Engineering.
Advisor: Eric Miller.
Committee: Xiaoyin Xu, Brian Tracey, and Misha Kilmer.
Keyword: Electrical engineering.read less
|
This object is in collection