|
One of the changes made in SPM2 is that in order to customize your analysis procedure you are now required to change the default settings related to the steps you are required to take. Here is a list of the options that you are allowed to customize with some brief notes on which ones are prefered.
Defaults for...
Realignment
Quality: This is pretty self-explanatory; it determines the quality of your registration. If you are worried you are not achieving the appropriate motion correction you should up the quality to time tradeoff.
Weighting: This will allow you to select a mask for areas that should not be weighted as much in the registration procedure. If you are dealing with brains that have known artifacts in certain areas that you wish not to effect the normalization you should enable this feature and create the appropriate mask (i.e. values from 0, not used at all, to 1, fully used).
Reslice interpolation method:
Main observable differences have to do with the histogram of the data. The tri-linear interpolation seems to begin at zero whereas the spline interpolations begin before zero with the peak beginning about where the tri-linear peak begins (a little after zero). The second observable slope in the data, presumably from white to gray matter, begins later in the spectrum (110 vs 130?). The tri-linear method produces a smoother result (looks like gaussian photoshop smoothing ~1px), loosing what I interpret as noise that was acquired during scanning. It also minimally changes the values of some voxels (i.e. from ~114 to ~116) depending on the interpolation method.
% The method by which the images are sampled when being written in a
% different space.
% 'Nearest Neighbour'
% - Fastest, but not normally recommended.
% 'Bilinear Interpolation'
% - OK for PET, or realigned fMRI.
%
% 'B-spline Interpolation'
% - Better quality (but slower) interpolation, especially
% with higher degree splines. Don't use B-splines when
% there is any region of NaN or Inf in the images.
% 'Fourier space Interpolation'
% - Rigid body rotations are executed as a series of shears,
% which are performed in Fourier space (Eddy et. al. 1996).
% Unfortunately, this method can only be applied to images
% with cubic voxels (since zooms can not be done by
% convolution in Fourier space).
% [defaults.realign.write.interp]
Way to wrap images: This option allows you to choose to wrap images around if they are coregistered beyond the voxel space. Ideally it should not be necessary, however, it is suggested that you wrap in Y if you are dealing with un-resliced MRI images where phase encoding is in the Y direction.
Mask Images:
% Because of subject motion, different images are likely to have different
% patterns of zeros from where it was not possible to sample data.
% With masking enabled, the program searches through the whole time series
% looking for voxels which need to be sampled from outside the original
% images. Where this occurs, that voxel is set to zero for the whole set
% of images (unless the image format can represent NaN, in which case
% NaNs are used where possible). This is in order to avoid artifactual
% movement-related variance the realigned images.
% [defaults.realign.write.mask]
Back to top
Spatial Normalization
Parameter Estimation
Weight template when registering: A mask is a series of values, one for each voxel, with a value between 0 and 1, which determines the amount that the voxel with be utilized when computing the normalization. The goal of this option is to weight the template image as so not to use parts of the image that you do not wish to influence the normalization. The default mask is meant to mask the skull out of the T1 template provided with spm2b. If you are running a study that only wishes to normalize the cerebellum, you may wish to create a mask that reflects this decision.
Weight source images when registering: This option allows you to mask out individual subject abnormalities when normalizing by creating a mask as described above.
Cutoff: Currently, I read this to document the smallest warp that will be applied to the data, i.e. if you set the cutoff to 25mm SPM2 will not attempt to fit any changes smaller than 25mm.
% Cutoff of DCT bases. Only DCT bases of periods longer than the
% cutoff are used to describe the warps. The number used will
% depend on the cutoff and the field of view of the template image(s).
% [defaults.normalise.estimate.cutoff]
Nonlinear Regularization: This option modifies the weight of smooth deformations in the trade-off between smooth deformation fields and the residual sum of squares difference between the subject and template images. If your normalization tends to warp images in unnatural ways you may wish to increase the amount of regularization, but otherwise it is acceptable to use medium or light regularization.
Nonlinear Iterations: This is the number of iterations SPM2 will perform when searching for the best set of non-linear basis functions. In practice it does not take 16 iterations but it is generally preferred, unless time is of the essence.
Back to top
Writing Normalized
Preserve What: SPM99 always preserved the concentrations, and this would be considered the standard way to normalize, however, “preserve total” will preserve the total amount of signal in a given area regardless of whether it expands or contracts. For example, if you have 6 voxels that all have a value of 1 that are combined into 1 voxel: with preserve concentration the value of that voxel will be 1, and with preserve total the value of that voxel will be 6.
Bounding Box: This determines the number of voxels relative to the origin (which should be set at the Anterior Commisure) that will be included in the final image. If you are normalizing to MNI space, using the default bounding box will generally include all of the brain, but get rid of surrounding space (i.e. where the skull is).
Voxel Sizes: This is the size of the voxels that the normalization procedure will output. It makes sense to preserve the voxel-sizes of your data, but since SPM2b converts your data into cubic voxels when normalizing (CHECK THIS) you should be sure to choose the most common or smallest original voxel size to minimize the warping of your data.
Interpolation method:
Main observable differences have to do with the histogram of the data. The tri-linear interpolation seems to begin at zero whereas the spline interpolations begin before zero with the peak beginning about where the tri-linear peak begins (a little after zero). The second observable slope in the data, presumably from white to gray matter, begins later in the spectrum (110 vs 130?). The tri-linear method produces a smoother result (looks like gaussian photoshop smoothing ~1px), loosing what I interpret as noise that was acquired during scanning. It also minimally changes the values of some voxels (i.e. from ~114 to ~116) depending on the interpolation method.
% The method by which the images are sampled when being written in a
% different space.
% 'Nearest Neighbour'
% - Fastest, but not normally recommended.
% 'Bilinear Interpolation'
% - OK for PET, or realigned fMRI.
%
% 'B-spline Interpolation'
% - Better quality (but slower) interpolation, especially
% with higher degree splines. Don't use B-splines when
% there is any region of NaN or Inf in the images.
% 'Fourier space Interpolation'
% - Rigid body rotations are executed as a series of shears,
% which are performed in Fourier space (Eddy et. al. 1996).
% Unfortunately, this method can only be applied to images
% with cubic voxels (since zooms can not be done by
% convolution in Fourier space).
% [defaults.realign.write.interp]
Way to wrap images: This option allows you to choose to wrap images around if they are coregistered beyond the voxel space. Ideally it should not be necessary, however, it is suggested that you wrap in Y if you are dealing with un-resliced MRI images where phase encoding is in the Y direction.
|
