Graph and Point Cloud Matching for Image Registration

Abstract

This thesis focuses on the topic of image or volume registration of data containing tree and graph shaped structures, with a special focus on medical imaging. The geometrical information is first extracted from the volumes or images and then used for registration. We propose a method for the segmentation of trees in images acquired at different time instances, by enforcing time consistency. This results in an overall improvement of the extraction accuracy. The method was tested on medical, biological and road images. The focus of this thesis is finding the alignment between segmented graphs and trees. We first propose a method called Active Testing Search (ATS) that explores partial correspondences of branching points of the structures. The method estimates the probability of partial match correctness based on training data and incrementally grows these partial matches. The ATS approach was able to align real data from several different medical imaging modalities, and is robust to initial position, rotation, deformation, missing data and noise. The second proposed method is called Graph Matching using Monte Carlo tree search (GMMC). The approach uses a stochastic state-space search algorithm inspired by the Monte Carlo tree search method to build a large set of compatible curves. Further acceleration is achieved by pruning using novel curve descriptors. The method can handle partial matches, topological differences, geometrical distortion, does not use appearance information and foes not require an initial alignment. Moreover, our method is very efficient – it can match graphs with thousands of nodes, which is an order of magnitude better than the best competing method.