Dense deformation estimation for pairwise and multi-subjects registration

Medical imaging is nowadays a vital component of a large number of clinical applications. For comparing images of the same patient (sometimes acquired using different modalities) or for comparing different patients, the images need to be aligned. When images from the same patient or from a collection of patients are acquired using different modalities, their alignment is necessary. Registration is the process of finding the best alignment between a pair or a collection of images. The main topic of this thesis is the conception and the application of intensity-based techniques, both for pairwise and multi-subjects registration. In the context of pairwise registration, we have investigated the use of a stochastic gradient descent technique (SPSA) for optimizing the mutual information metric. For non-rigid registration, the use of volumetric tetrahedral meshes has been implemented as a deformation model in collaboration with A. du Bois d'Aische. The targeted application for our algorithm is the tracking of anatomical changes between pre-operative and intra-operative images in brain, prostate and liver surgery. A second method, equivalent to optical flow but developed for multi-modal images is also described and applied to the problem of atlas to pathological brain registration. In the context of multi-subjects registration, we developed an unbiased atlas generation technique in the Expectation-Maximization framework. At each iteration, the method estimates a reference for the registration problem by performing an average giving more weight to consistent experts (E step). The consistency parameters and the transformation parameters are estimated in the M step. Our atlas generation technique has been applied for aligning 80 brain images segmented in 4 labels (background, gray matter, white matter and ventricles).