Discussion

Registration has been the subject of several important reviews and comparative studies, the first of which was made by Van den Elsen [163] in 1993.
Zuk [189] compared manual and automatic retrospective methods. He found the automatic surface-based methods to be the fastest and automatic volume-based methods the most accurate, provided the initial position of the scans to register is close enough to converge to the correct minimum.
West and most other researchers in the field performed a major comparative study [181,182], using fiducial markers to define the gold standard transformation. These results were also used in a later study comparing surface-based and volume based methods [179,180]. It concluded that both approaches can give satisfactory results, but that volume-based techniques tend to be more accurate, especially in the rotational component of the rigid transformation.
On the other hand, there are cases where only surface-based registration is possible, because volumetric data is not available in one of the information sources to register. For instance, Herring [75] registers CT images to the physical space for image-guided spine surgery. The physical surface is acquired using a 3DSL consisting of a probe and an Optotrak system. The image surface is generated by an iso-surfacing algorithm on a tetrahedral decomposition of the volume data. Grimson [68] proposes a surface-based registration for frameless stereotaxy, image-guided surgery. Maurer [104] uses a combination of a point-based and surface-based methods to register head CT images to the physical space.
In what follows, we present two such applications, where the choice of a surface-based matching criterion is driven by the lack of volumetric data in one of the modalities.