The fine spatial scales of the structures in the human brain represent an enormous challenge to the successful integration of information from different images. Cortex has a typical thickness of between 1 and 5mm (Fischl and Dale, 2000
). Thalamic nuclei have spatial extents on the order of a few millimeters (Kandel, 2000
). Cortex is also highly folded, meaning that points that are 20mm from each other as measured along the cortical surface can be within a few millimeters in three-dimensional space. This places stringent requirements on the accuracy and precision of algorithms for aligning images. Even a few millimeters of alignment error can cause a voxel in a tissue type in one image to be assigned to the wrong tissue type in another image. This challenge extends to several domains, including within-subject integration of images from a wide range of imaging modalities such as structural MRI, functional MRI (fMRI), arterial spin labeling (ASL), diffusion weighted imaging (DWI), positron emission tomography (PET), within-subject single-mode longitudinal analysis, and surface-based group analysis (Fischl et al., 1999
Automatic within- and cross-modal registration has a long history of research in neuroimaging. The most prominent and widely used algorithms currently extant are Cross-Correlation (Collins et al., 1995
), Mutual Information (MI; Maes et al., 1997
; Maes et al., 1999
; Wells et al., 1996
), Normalized Mutual Information (NMI), and Correlation Ratio (CR; Roche et al., 1998
). Surface-to-surface and shape registration (Borgefors, 1988
) have also been used for multimodal registration; however, West (West et al., 1999
) found that these were not as accurate as intensity-based techniques. The basic model used in CR is that an intensity value in one mode will have one, and only one, matching intensity value in the other mode. MI and NMI are similar to CR but less restrictive in that they attempt to sharpen the intensity joint histogram. For both methods, intensity mismatches are evidence for misregistration. Correspondence can be achieved by adjusting the registration parameters (i.e., translations, rotations, scalings, shears) until the best match (i.e., minimum cost) is achieved. Unfortunately, intensity inconsistencies can exist for other reasons, which may be mode-specific. For example, in a Blood-Oxygen-Dependent (BOLD) weighted image, the tissue outside of the brain often appears quite dark whereas it will be bright on an anatomical. The brain can be extracted from the anatomical image (Segonne et al., 2004
), but the quality of the resulting registration will then be sensitive to the aggressiveness and quality of the extraction. Echo planar images (EPI) are also subject to B0 distortion in the form of non-linear metric distortion and intensity “drop out” (Jezzard and Balaban, 1995
). In addition, coil sensitivity profiles and B1 inhomogeneity can create spatial intensity fluctuations. Coil sensitivity fluctuations can become extreme when surface coils are used as is often done when studying retinotopy with fMRI (Sereno et al., 1995
). All these effects create inconsistencies in input intensity matching that are unrelated to the quality of alignment, but will be indistinguishable from alignment errors and thus can drive an alignment away from the true optimum.
In longitudinal analysis, in which the same subject is imaged over time in different scan sessions, cross-session alignment of the brain is well described by a rigid transformation, but there are significant non-rigid effects such as differences in jaw and tongue placement or head-neck angle. In theory, these can be accommodated by brain extraction, but, again, the resulting registration will be sensitive to differences in the extraction of the brain at different time points. Finally, we point out that some imaging modalities may not have full brain coverage. For example, the brain coverage in EPI must be reduced in order to reduce the slice thickness and maintain temporal resolution. Partial field-of-view (FoV) brain coverage creates enormous problems when attempting to register to a whole head
Recently, Saad (Saad et al., 2009
) performed an extensive analysis comparing CR and MI and found that they had errors that could easily exceed 3mm. This led them to propose a new cost function optimized for T2*-T1 registration using a local Pearson correlation (LPC). LPC is a local method, meaning that the cost at each voxel is only dependent on the nearby voxels (roughly the 36 nearest neighbors). As part of the computation of the Pearson correlation, the mean intensity over a neighborhood is subtracted from the intensity at each voxel in the neighborhood. This is effectively a spatial high-pass filter and so enhances edges in both the input and reference images; the correlation is then computed over the neighborhood. Areas away from tissue boundaries should contribute little to the overall cost since these areas tend to have locally homogeneous intensities in both images which will be suppressed by the local mean removal. This means that the areas near the tissue boundaries will drive the cost function; in this way it is similar to BBR.
Like Saad, et al, we have also observed inaccuracies and sensitivities in the registrations found by CR and MI. In this paper, we propose a new algorithm based on the principle that the most salient registration cue is the contrast across a tissue boundary, and so we refer to it as Boundary-based Registration, or BBR. Unlike the methods reviewed above, BBR does not treat the two images as equal. One of the images (the “reference image”) must be a high-quality anatomical volume sufficient for extracting surfaces that separate brain structures and tissue types. The second image (the “input image”) can be any modality as long as it has tissue contrast. Alignment is achieved by maximizing the gradient of the input image intensity across the surface boundary. Intensity values in the anatomical image are not part of the cost function. If two or more input images need to be aligned, they can each be separately aligned to the reference with BBR. We emphasize that this is not a surface-to-surface registration like those described above. A surface is only extracted from one image, the high-quality anatomical; intensities are used from the second image. It is similar to LPC in that a local cost is computed from tissue boundaries. However, there are some differences in that BBR operates over a much smaller neighborhood (a few millimeters) and uses percent contrast instead of correlation coefficient. At the time this manuscript was being prepared, LPC had only recently been published, and so we have not had the chance to compare it to BBR directly.
We show that BBR yields superior accuracy compared to CR and NMI using blinded human raters as well as improved fMRI results. We then show that BBR is extremely robust to variations in its parameters and initialization, major fluctuations in image intensity inhomogeneity, and to partial brain images, even to the extent of accurately registering single slices, something for which most current registration methods fail. The software that implements BBR is publically distributed as part of the FreeSurfer (surfer.nmr.mgh.harvard.edu) software package.