3 Cross-sectional registration module


This module uses SPM12's DARTEL implementation to spatially variably non-linearly register pGM and pWM maps between subjects. DARTEL iteratively creates a group mean pGM and pWM (from T1 segmentations in T1 module) and non-linearly registers individual probability maps to the mean maps, with decreasing smoothness 5. Note that these pGM and pWM maps (rc1T1.nii and rc2T1.nii) are already in a very similar space between subjects, because of the use of the inverse transformation from the SPM12 segmentation. This results in \\analysis\dartel/u_rc1T1_SubjectName_T1_template.nii files, which contain flow fields that describe for each subject the transformation to the common space. For each voxel, these flow fields gives the relative coordinates (i.e. how much voxels should the voxel move up/down/left/right/forward/backward).

In case of longitudinal registration, only SubjectName_1 are included in the DARTEL list. Otherwise, all subjects are included. An error will occur if not all subjects can be included. In that case, check whether all T1 modules ran correctly. Then the number of subjects included in ExploreASL should be the same as the number of *.jpeg check files in e.g. //analysis/dartel/T1_CHECKDIR.

Separate DARTEL runs

DARTEL has by default 6 "outer iterations", which are repetitions of DARTEL with increasing spatial precision (i.e. less smoothing). ExploreASL runs these 6 iterations separate instead of a single DARTEL run. This way, if ExploreASL crashes for whatever reason - e.g. you forgot to recharge your laptop - it will simply continue at the iteration where it was, rather than restarting DARTEL completely at the first iteration. Each iteration rerun will re-use the same flow field files, and will store its group mean templates as template_0.nii to template_1.nii (for before and after the DARTEL transformation estimation). So the final template will be named \\analysis\dartel\template_1.nii rather than template_6.nii.

These Figures show the original T1 image in its original native space, and in the row below from left to right the T1 image in MNI space after the T1 module (using the inverse deformation from segmentation), the DARTEL flow fields for this subject and the T1 image in MNI space after applying the DARTEL flow fields. pGM and pWM are projected in red and blue on top of the T1 image. Note how before DARTEL the images are already very nicely positioned, but the brain still looks like an individual brain, there are unique spatial variations that make it look like an individual brain. These are taken out by the DARTEL flow fields, and the resulting images after DARTEL look very symmetrical, and very similar for all subjects.

This shows the DARTEL template creation procedure, with the enhanced priors and longitudinal registration (n~300 subjects). Improvements are small, which can be explained by the large population (and hence averaging out of variability and only showing consistency), and by the good initial starting point of the affine transformation into MNI space by the SPM12 segmentation.

2) Rerun DARTEL if additional T1 images are found

If there are new T1 images (i.e. subjects) added to the population, after DARTEL was previously ran and the DARTEL flowfields exists for all but the new T1 images, then ExploreASL checks whether the number of new subjects is smaller than 5% of total population, otherwise we are better off creating new templates altogether. If we add less than 5% subjects, we simply use the existing templates. If not all the template exists (but only the latest, most sharp one), then we recreate the smoother, earlier, DARTEL templates by assuming that the 7 templates [0 1 2 3 4 5 6] should follow a Gaussian [8 6.25 4.5 2.75 1 0 0] FWHM smoothness additional to the latest, sharpest template.