Automatic retrospective methods

Manual (human-based) methods rely on a small set of highly semantic information such as the exact location of specific anatomical features in both images. Automatic (computer-based) methods compensate the lesser semantic ability of computers by increasing the amount of data used in the matching criterion.
Surface-based method, such as those proposed by Pellizari [119], Jiang [80,81], Mangin [101], Lemoine [92] or Hemler [73], are divided in two steps. First, similar features are segmented from both images. Then, the best transformation is defined as minimizing the distance between those features.
Volume-based methods use the complete dataset into account in the matching criterion. Such methods were proposed by Woods [184], van den Elsen [162,161], Maintz [99], Studholme [150], Collignon, Maes [21,98] and Wells [47]. The central point of these methods is how they define the similarity between pixels from modalities where similar gray-levels do not necessary correspond to similar tissue types. For instance, Van den Elsen [161] correlates ``ridgeness'' images to register MRI and CT scans. Wells [47] and Maes [98] maximize the mutual information between images from any two modalities.