Non-linear registration for brain images by maximising feature and intensity similarities with a Bayesian framework

The objective of this work was to provide a new, precise registration of the cortical mantle with a non-linear transformation. Image registration is broadly classified into two types, using intensity similarity and feature similarity. Whereas the former approach has merit in global brain matching, the latter provides a fast registration centred on a region of interest. The hybrid registration proposed in this paper was achieved using a Bayesian framework, which consisted of a likelihood model including intensity similarity and a prior model including feature information and a smoothing constraint. In this approach, each voxel was spatially transformed, so that the distance between corresponding features was shortened and also so that the intensity correlation was maximised. The result of the hybrid method clearly showed a good match of global brain (r = 0.930) by including intensity similarity. Moreover, this method compensated for the approximated sulcus of the feature-based method with intensity information, so that the geometric shape and thickness of the sulcus at the feature-defined region was likely to be registered. The accuracy in the feature-defined area was improved by 33.4% and 7.5% compared with feature-based and intensity-based methods, respectively.