MRIprocessing (spatial normalization and segmentation)
To make statistical analyses across many MRIs of different individuals with different brain structure, preprocessing of the brain is necessary. The first process is known as spatial normalization or anatomical standardization. It is generally achieved by registering all images in a population into the same template image so that they are all in the same standard stereotactic space. After this process, an anatomical location in one subject’s MRI corresponds to the same location in another subject’s MRI. Different algorithms can be used to perform this registration [13, 64, 314]. The most commonly applied algorithm available in the SPM software (described later) includes a 12-parameter affine transformation followed by a nonlinear registration using a mean squared difference matching function [13,265]. The template image used for the spatial normalization could be one specific MRI scan, which is selected from a population based on defined criteria, or could be created by averaging across a number of different MRI scans such as the MNI152 template that was created by averaging 152 healthy brains at the Montreal Neurological Institute (MNI). Customized templates may be created using the given study group or a group that is matched to the study group in terms of age and disease status. Such templates may improve the normalization of each subject in a study group .
In SPM, in which a low-parameter shape transformation is performed for spatial normalization, a step called modulation is then often necessary to correct for volume changes during the spatial normalization step . Image intensities are scaled by the amount of contraction that has occurred during spatial normalization, so that the total amount of gray matter remains the same as in the original image. Then, statistical comparison of volumetric differences between scans is performed. If the spatial normalization was precise and all the segmented images appear identical, no significant differences would be detected in unmodulated data, and the analysis would reflect registration error rather than volume differences. Other techniques employ different normalization procedures that use high-dimensional elastic transformations , or ELAST . These procedures preserve the volume of different tissues and do not require a separate modulation step.
Images are segmented into different tissue compartments (gray matter, white matter, and cerebrospinal fluid (CSF)), and statistical analysis is performed separately on either gray or white matter, depending on the target tissue to be analyzed
S. Hanaoka et al.
Fig. 3.18 Image processing and analyses (Figure 1 of Ref. )
(Fig. 3.18). There are a number of ways to perform the segmentation, including classification using voxel signal intensity combined with prior probability maps, as in SPM. Such prior probability maps may be more unbiased when generated from the specific population under study. However, weighting balance between signal intensities and prior probability should be considered, when differences in the transformation vector itself in two different populations are a matter of concern. The accuracy of the segmentation will also depend on the quality of the normalization. Iterative versions of normalization and segmentation methods have been developed which enable optimization of both processes concurrently, to improve the final segmentation . In this method, the original MRI in native space is segmented, and then the segmented images are spatially normalized to gray matter and white matter templates to obtain optimized normalization parameters. The method is termed “optimized VBM” . Segmentation errors can occur because of displacement of tissue and partial volume effects between gray matter and CSF. Both are more likely to occur in atrophic brains in subjects of older age or with degenerative brain diseases. The use of customized templates can help to minimize some of these potential errors .