Adaptive energy functions and efficient graph-based optimization for deformable image registration

Deformable registration, the task of bringing two images into spatial correspondence, is a prerequisite step in many medical image analysis problems, including multi-modal image fusion, multi-temporal image analysis, atlas construction, and atlas-based segmentation. An increasingly popular approach is to cast the registration task as a graph-labelling problem and solve the problem via discrete optimization, which allows for efficient and robust registration. In this thesis, we advance registration methods in the graph-based paradigm by developing strategies to increase the model fidelity of various registration objectives and to perform their optimizations more efficiently. Specifically, we propose different ways to adapt image similarity and regularization by leveraging prior knowledge, information learned from previous examples, and information derived solely from the given input data. To reduce the computational complexity of the optimization without compromising accuracy, we also present various schemes that use contextual information or prior knowledge to conduct the solution search efficiently. To this end, we have applied our proposed methods to various deformable registration problems, as well as shape matching and contour tracking tasks with demonstrated success.