From pairwise medical image registration to populational computational atlases

Medical image registration is a widely used strategy for intrasubject and intersubject matching. Over the last twenty years, numerous strategies have been designed to address the challenges raised by medical applications including the capability to handle images acquired by different imaging sensors and the estimation of flexible but realistic transformations for modeling intersubject and pathology-induced deformations. This chapter presents a survey of pairwise intensity-based automatic registration algorithms by classifying them according to the similarity cost function and extracted features used for quantifying the matching, the representation of the space of allowed transformations and the regularization strategy used to ensure continuous and smooth transformations. Joint alignment of a population of subjects is an interesting problem that can be seen as a natural extension of pairwise registration. It brings forward the open question of generalizing pairwise similarity cost functions and the definition of a common reference space where the population under study can be projected. An overview of existing techniques for the construction of computational atlases based on a collection of subjects is presented. The two main representations of such atlases — probabilistic and statistical — are described. Finally, we present a review of state of the art techniques for the joint alignment of a population of images.