Using Dartel

Dartel¹ is a suite of tools for achieving more accurate inter-subject registration of brain images. It consists of several thousand lines of code. Because it would be a shame if this effort was wasted, this guide was written to help encourage its widespread use. Experience at the FIL would suggest that it offers very definite improvements for VBM studies – both in terms of localisation² and increased sensitivity³.

Using Dartel for VBM

The following procedures could be specified one at a time, but it is easier to use the batching system. The sequence of jobs (use the Batch button to start the batch manager) would be:

• SPMSpatialSegment: To generate the roughly (via a rigid-body) aligned grey and white matter images of the subjects.

• SPMToolsDartel ToolsRun Dartel (create Template): Determine the nonlinear deformations for warping all the grey and white matter images so that they match each other.

• SPMToolsDartel ToolsNormalise to MNI Space: Actually generate the smoothed “modulated” warped grey and white matter images.

Further details of the steps are described next.

Using SpatialSegment

Note: This subsection will be elaborated on later.

The first step is to classify T1-weighted scans⁴ of a number of subjects into different tissue types via the Segmentation routine in SPM (Ashburner and Friston 2005), which can be found under SPMSpatialSegment. With this option, the “imported” tissue class images (usually rc1.nii and rc2.nii) would be generated directly. It is also suggested that Native Space versions of the tissues in which you are interested are also generated. For VBM, these are usually the c1*.nii files, as it is these images that will eventually be warped to MNI space. Both the imported and native tissue class image sets can be specified via the Native Space options of the user interface.

Segmentation can require quite a lot of memory, so if you have large images (typically greater than about \(256\times256\times150\)) and trying to run it on a 32 bit computer or have relatively little memory installed, then it may throw up an out of memory error.

Using Dartel ToolsRun Dartel (create Template)

The output of the previous step(s) are a series of rigidly aligned tissue class images (grey matter is typically encoded by rc1*.nii and white matter by rc2*.nii – see Fig 1.2). The headers of these files encode two affine transform matrices, so the Dartel tools are still able to relate their orientations to those of the original T1-weighted images. The next step is to estimate the nonlinear deformations that best align them all together (Ashburner 2007). This is achieved by alternating between building a template, and registering the tissue class images with the template, and the whole procedure is very time consuming. Specify SPMToolsDartel ToolsRun Dartel (create Template).

• Run Dartel (create Template)

• Images

  • Images: Select all the rc1*.nii files generated by the import step.

  • Images: Select all the rc2*.nii files, in the same subject order as the rc1*.nii files. The first rc1*.nii is assumed to correspond with the first rc2*.nii, the second with the second, and so on.

• Settings: Default settings generally work well, although you could try changing them to see what happens. A series of templates are generated called Template_basename_0.nii, Template_basename_1.nii, etc. If you run multiple Dartel sessions, then it may be a good idea to have a unique template basename for each.

The procedure begins by computing an initial template from all the imported data. If u_rc1*.nii files exist for the images, then these are treated as starting estimates and used during the creation of the initial template. If any u_rc1*.nii files exist from previous attempts, then it is usually recommended that they are removed first (this sets all the starting estimates to zero). Template generation incorporates a smoothing procedure (Ashburner and Friston 2008), which may take a while (several minutes). Once the original template has been generated, the algorithm will perform the first iteration of the registration on each of the subjects in turn. After the first round of registration, a new template is generated (incorporating the smoothing step), and the second round of registration begins. Note that the earlier iterations usually run faster than the later ones, because fewer “time-steps” are used to generate the deformations. The whole procedure takes (in the order of) about a week of processing time for 400 subjects.

Different stages of template generation. Top row: an intermediate version of the template. Bottom row: the final template data.

The end result is a series of templates (see Fig 1.1), and a series of u_rc1*.nii files. The first template is based on the average⁵ of the original imported data, where as the last is the average of the Dartel registered data. The u_rc1*.nii files are flow fields that parameterise the deformations. Note that all the output usually contains multiple volumes per file. For the u_rc1*.nii files, only the first volume is visible using the Display or Check Reg tools in SPM. All volumes within the template images can be seen, but this requires the file selection to be changed to give the option of selecting more than just the first volume (in the file selector, the widget that says “1” should be changed to “1:2”).

Using Dartel ToolsNormalise to MNI Space

The next step is to create the Jacobian scaled (“modulated”) warped tissue class images, by selecting SPMToolsDartel ToolsNormalise to MNI Space. The option for spatially normalising to MNI space automatically incorporates an affine transform that maps from the population average (Dartel Template space) to MNI space, as well as incorporating a spatial smoothing step.

• Normalise to MNI Space

• Dartel Template: Specify the last of the series of templates that was created by Run Dartel (create Template). This is usually called Template_6.nii. Note that the order of the N volumes in this template should match the order of the first N volumes of the toolbox/Dartel/TPM.nii file.

• Select according to either Few Subjects or Many Subjects. For VBM, the Many Subjects option would be selected.

  • Flow Fields: Specify the flow fields (u_rc1*.nii) generated by the nonlinear registration.

  • Images: You may add several different sets of images.

  • Images: Select the c1*.nii files for each subject, in the same order as the flow fields are selected.

  • Images: This is optional, but warped white matter images can also be generated by selecting the c2*.nii files.

• Voxel sizes: Specify the desired voxel sizes for the spatially normalised images ([NaN, NaN, NaN] gives the same voxel sizes as the Dartel template).

• Bounding box: Specify the desired bounding box for the spatially normalised images ([NaN, NaN, NaN; NaN NaN NaN] gives the same bounding box as the Dartel template).

• Preserve: Here you have a choice of Preserve Concentrations (ie not Jacobian scaled) or Preserve Amount (Jacobian scaled). The Preserve Amount would be used for VBM, as it does something similar to Jacobian scaling (modulation).

• Gaussian FWHM: Enter how much to blur the spatially normalised images, where the values denote the full width at half maximum of a Gaussian convolution kernel, in units of mm. Because the inter-subject registration should be more accurate than when done using other SPM tools, the FWHM can be smaller than would be otherwise used. A value of around 8mm (ie [8, 8, 8]) should be about right for VBM studies, although some empirical exploration may be needed. If there are fewer subjects in a study, then it may be advisable to smooth more.

The end result should be a bunch of smwc1*.nii files⁶ (possibly with smwc2*.nii if white matter is also to be studied).

Pre-processing for VBM. Top row: Imported grey matter (`rc1A.nii` and `rc1B.nii`). Centre row: Warped with Preserve Amount option and zero smoothing ("modulated"). Bottom row: Warped with Preserve Amount option smoothing of 8mm (`smwc1A.nii` and `smwc1B.nii`).

The final step is to perform the statistical analysis on the preprocessed data (smwc1*.nii files), which should be in MNI space. The next section says a little about how data from a small number of subjects could be warped to MNI space.

Spatially normalising functional data to MNI space

Providing it is possible to achieve good alignment between functional data from a particular subject and an anatomical image of the same subject (distortions in the fMRI may prevent accurate alignment), then it may be possible to achieve more accurate spatial normalisation of the fMRI data using Dartel. There are several advantages of having more accurate spatial normalisation, especially in terms of achieving more significant activations and better localisation.

The objectives of spatial normalisation are:

• To transform scans of subjects into alignment with each other. Dartel was developed to achieve better inter-subject alignment of data.

• To transform them to a standard anatomical space, so that activations can be reported within a standardised coordinate system. Extra steps are needed to achieve this aim.

The option for spatially normalising to MNI space automatically incorporates an affine transform that maps from the population average (Dartel Template space) to MNI space. This transform is estimated by minimising the KL divergence between the final template image generated by Dartel and tissue probability maps that are released as part of SPM’s segmentation (Ashburner and Friston 2005). MNI space is defined according to affine matched images, so an affine transform of the Dartel template to MNI space would appear to be a reasonable strategy.

For GLM analyses, we usually do not wish to work with Jacobian scaled data. For this reason, warping is now combined with smoothing, in a way that may be a bit more sensible than simply warping, followed by smoothing. The end result is essentially the same as that obtained by doing the following with the old way of warping

• Create spatially normalised and “modulated” (Jacobian scaled) functional data, and smooth.

• Create spatially normalised maps of Jacobian determinants, and smooth by the same amount.

• Divide one by the other, adding a small constant term to the denominator to prevent divisions by zero.

This should mean that signal is averaged in such a way that as little as possible is lost. It also assumes that the procedure does not have any nasty side effects for the GRF assumptions used for FWE corrections.

Prior to spatially normalising using Dartel, the data should be processed as following:

• If possible, for each subject, use SPMToolsFieldMap to derive a distortion field that can be used for correcting the fMRI data. More accurate within-subject alignment between functional and anatomical scans should allow more of the benefits of Dartel for inter-subject registration to be achieved.

• Use either SPMSpatialRealignRealign: Estimate Reslice or SPMSpatialRealign Unwarp. If a field map is available, then use the Realign Unwarp option. The images need to have been realigned and resliced (or field-map distortion corrected) beforehand - otherwise things are not handled so well. The first reason for this is that there are no options to use different methods of interpolation, so rigid-body transforms (as estimated by Realign but without having resliced the images) may not be well modelled. Similarly, the spatial transforms do not incorporate any masking to reduce artifacts at the edge of the field of view.

• For each subject, register the anatomical scan with the functional data (using SPM Spatial Coreg Coreg: Estimate). No reslicing of the anatomical image is needed. Use SPMUtilCheck Registration to assess the accuracy of the alignment. If this step is unsuccessful, then some pre-processing of the anatomical scan may be needed in order to skull-strip and bias correct it. Skull stripping can be achieved by segmenting the anatomical scan, and masking a bias corrected version (which can be generated by the segmentation option) by the estimated GM, WM and CSF. This masking can be done using SPMUtilImage Calculator (ImCalc button), by selecting the bias corrected scan (m*.nii), and the tissue class images (c1*.nii, c2*.nii and c3*.nii) and evaluating i1.((i2+i3+i4)>0.5). If segmentation is done before coregistration, then the functional data should be moved so that they align with the anatomical data.

• Segment the anatomical data and generate “imported” grey and white matter images.

• To actually estimate the warps, use SPMToolsDartel ToolsRun Dartel (create Templates) in order to generate a series of templates and a flow field for each subject. Previous steps processed each subject’s data individually, but this stage requires imported data for all subjects in a study to be processed together. The idea is to align all subject’s data to the average anatomy of the study population.

In principle (for a random effects model), you could run the first level analysis using the native space data of each subject. All you need are the contrast images, which can be warped and smoothed. Alternatively, you could warp and smooth the reslices fMRI, and do the statistical analysis on the spatially normalised images. Either way, you would select SPMToolsDartel ToolsNormalise to MNI Space:

• Normalise to MNI Space

• Dartel Template: Template_6.nii,1 is usually the grey matter component of the final template of the series. An affine transform is determined using this image.

• Select according to either Few Subjects or Many Subjects. For fMRI analyses, the Few Subjects option would be selected, which gives the option of selecting a flow field and list of images for each subject.

  • Subject

  • Flow Field: Specify the flow field (u_rc1*.nii) for this subject.

  • Images: Select the images for this subject that are to be transformed to MNI space.

• Voxel sizes: Specify the desired voxel sizes for the spatially normalised images ([NaN, NaN, NaN] gives the same voxel sizes as the Dartel template).

• Bounding box: Specify the desired bounding box for the spatially normalised images ([NaN, NaN, NaN; NaN NaN NaN] gives the same bounding box as the Dartel template).

• Preserve: Here you have a choice of Preserve Concentrations (ie not Jacobian scaled) or Preserve Amount (Jacobian scaled). The Preserve Concentrations option would normally be used for fMRI data, whereas Preserve Amount would be used for VBM.

• Gaussian FWHM: Enter how much to blur the spatially normalised images, where the values denote the full width at half maximum of a Gaussian convolution kernel, in units of mm.

Warping Images to Existing Templates

If templates have already been created using Dartel, then it is possible to align other images with such templates. The images would first be imported in order to generate rc1*.nii and rc2*.nii files. The procedure is relatively straight-forward, and requires the SPMToolsDartel ToolsRun Dartel (existing Template) option to be specified. Generally, the procedure would begin by registering with a smoother template, and end with a sharper one, with various intermediate templates between.

• Run Dartel (existing Templates)

• Images

  • Images: Select the rc1*.nii files.

  • Images: Select the corresponding rc2*.nii files.

• Settings: Most settings would be kept at the default values, except for the specification of the templates. These are specified in within each of the SettingsOuter IterationsOuter IterationTemplate fields. If the templates are Template_*.nii, then enter them in the order of Template_1.nii, Template_2.nii, …, Template_6.nii.

Running this option is rather faster than Run Dartel (create Template), as templates are not created. The output is in the form of a series of flow fields (u_rc1*.nii).

Warping one individual to match another

Sometimes the aim is to deform an image of one subject to match the shape of another. This can be achieved by running Dartel so that both images are matched with a common template, and composing the resulting spatial transformations. This can be achieved by aligning them both with a pre-existing template, but it is also possible to use the Run Dartel (create Template) option with the imported data of only two subjects. Once the flow fields (u_rc1*.nii files) have been estimated, then the resulting deformations can be composed using SPMUtilsDeformations. If the objective is to warp A.nii to align with B.nii, then the procedure is set up by:

• Deformations

• Composition

  • Dartel flow

  • Flow field: Specify the u_rc1A_Template.nii flow field.

  • Forward/Backwards: Backward.

  • Time Steps: Usually 64.

  • Dartel template: leave this field empty.

  • Dartel flow

  • Flow Field: Specify the u_rc1B_Template.nii flow field.

  • Forward/Backwards: Forward.

  • Time Steps: Usually 64.

  • Dartel template: leave this field empty.

  • Identity

  • Image to base Id on: Specify B.nii in order to have the deformed image(s) written out at this resolution, and with the same orientations etc (ie so there is a voxel-for-voxel alignment, rather than having the images only aligned according to their “voxel-to-world” mappings).

• Output

  • Pullback

  • Apply to: Specify A.nii, and any other images for that subject that you would like warped to match B.nii. Note that these other images must be in alignment according to Check Reg.

  • Output destination: Specify where you want to write the images.

  • Interpolation: Specify the form of interpolation.

  • Mask images: Say whether you want to mask the images.

  • Gaussian FWHM: The images can be smoothed when they are written. If you do not want this, then enter 0 0 0.

Suppose the image of one subject has been manually labelled, then this option is useful for transferring the labels on to images of other subjects.

Composition of deformations to warp one individual to match another. Top-left: Original `A.nii`. Top-right: `A.nii` warped to match `B.nii`. Bottom-left: Original `B.nii`. Bottom-right: `B.nii` warped to match `A.nii`.

Warping the Dartel template to match an individual

Sometimes the aim is to warp one of the templates created by Dartel, or some other image that is in alignment with this template, to match one of the scans of an individual. This can be achieved using SPMUtilsDeformations by:

• Deformations

• Composition

  • Dartel flow

  • Flow field: Specify the u_rc1_Template.nii flow field.

  • Forward/Backwards: Forward.

  • Time Steps: Usually 64.

  • Dartel template: leave this field empty.

  • Identity

  • Image to base Id on: Specify a NIfTI image in order to have the deformed image(s) written out at this resolution, and with the same orientations etc (ie so there is a voxel-for-voxel alignment, rather than having the images only aligned according to their “voxel-to-world” mappings).

• Output

  • Pullback

  • Apply to: Specify the Dartel template or any other images that are in alignment with the Dartel template, such as label maps etc, that you would like warped to match the individual. You can use Check Reg to assess this alignment.

  • Output destination: Specify where you want to write the images.

  • Interpolation: Specify the form of interpolation.

  • Mask images: Say whether you want to mask the images.

  • Gaussian FWHM: The images can be smoothed when they are written. If you do not want this, then enter 0 0 0.

Ashburner, J. 2007. "A Fast Diffeomorphic Image Registration Algorithm." *NeuroImage* 38 (1): 95--113. .

Ashburner, J., and K. J. Friston. 2005. "Unified Segmentation." *NeuroImage* 26: 839--51. .

Ashburner, J., and K. J. Friston. 2008. "Computing Average Shaped Tissue Probability Templates." *NeuroImage* 45 (2): 333--41. .

---

1. Dartel stands for “Diffeomorphic Anatomical Registration Through Exponentiated Lie algebra”. It may not use a true Lie Algebra, but the acronym is a nice one. ↩

2. Less smoothing is needed, and there are fewer problems relating to how to interpret the differences. ↩

3. More sensitivity could mean that fewer subjects are needed, which should save shed-loads of time and money. ↩

4. Other types of scan may also work, but this would need some empirical exploration. ↩

5. They are actually more similar to weighted averages, where the weights are derived from the Jacobian determinants of the deformations. There is a further complication in that a smoothing procedure is built into the averaging. ↩

6. The actual warping of the images is done slightly differently, with the aim that as much of the original signal is preserved as possible. This essentially involves pushing each voxel from its position in the original image, into the appropriate location in the new image - keeping a count of the number of voxels pushed into each new position. The procedure is to scan through the original image, and push each voxel in turn. The alternative (older way) was to scan through the spatially normalised image, filling in values from the original image (pulling the values from the original). The results of the pushing procedure are analogous to Jacobian scaled (“modulated”) data. A minor disadvantage of this approach is that it can introduce aliasing artifacts (think stripy shirt on TV screen) if the original image is at a similar - or lower - resolution to the warped version. Usually, these effects are masked by the smoothing. ↩