Dear Martijn,
Sorry for the late reply.
> I guess it can be computed by summing the jacobians
The expansions/contractions reflected by the jd file (or alternatively, the dv file) are not specific for particular tissue types. Note that during the longitudinal registration, there's no segmentation or normalisation into template space (e.g. MNI), just warping one image onto another in "subject space" (plus some other operations like bias corrections). Accordingly and assuming the FOV is identical between the two anatomies, the sum of the voxels within the jd file should be (close to) 0 except there's not just brain atrophy (with intracranial volume remaining the same) but really head shrinkage. Thus, for GM changes over time you would have to combine the c1 file and the jd file (or alternatively, the dv file), see below.
> I want to calculate the percentage brain (or gray matter) volume change per year for each subject
During longitudinal registration write out the individual average images, which you would then forward into Segment, resulting in c1, c2, c3 files, which reflect probabilities for GM, WM, CSF on a voxel-by-voxel basis in native space for the arbitrary average time point. The probabilities are also commonly interpreted as percentages, thus e.g. c1 value 0.3 would mean 30% of the voxel are GM. To convert into absolute volume units you would multiply by the volume of one voxel and the number of voxels for the region under consideration, e.g. 0.3 * 1x1x1 mm^3/voxel * 10 voxel = 30 mm^3. Note that there's some confusion, as others go with an initial arbitrary threshold and any voxel with c1 values > e.g. 0.1 would be classified as being GM to 100% and any voxel with c1 values =< 0.1 would be classified as being not GM. This approach could be understood as mimicking manual ROI drawings/volume estimates, in which you rely on the intensities for defining what is GM and what is not, without further considering intensity differences.
For a global value of the relative change you just have to multiply the c1 file and the jd (dv) file and sum up then. For proper interpretation make sure about the settings during the longitudinal registration (which time difference did you enter), if you want to convert into volume units multiply with the voxel resolution and the no. of voxels.
> within the ICV mask
You don't necessarily have to rely on the mask as long as the segmentation is reasonably good. I had the feeling that it cropped of some (very) minor parts of the temporal cortices, but this might well depend on preprocessing/data. However, masking might be good if e.g. brainstem was covered to a different extent for different subjects, as the mask would exclude voxels of the most ventral slices. As the ICV mask is in MNI space you would have to apply the inverse deformation field iy_avg obtained during segmentation to transform the ICV mask into native space. Make sure about the settings, the default bounding box corresponds to settings for MNI space, possibly the values have to be adjusted.
> to register the average image to dartel
This depends on your preprocessing strategy. If you want to use Dartel for normalisation, you would go with the rc1, rc2 files and possibly create your own study-specific template or rely on an already existing version. Accordingly, you would have to go with a different inverse deformation field, the one obtained during Dartel preprocessing. But this doesn't affect volume estimates as obtained from above (which are those in native space and thus, always appropriate). You could also calculate the volume based on the normalised files, but for that purpose you would go with modulated versions taking into account warping from native space (or that space of the avg image) into MNI/template space (in which the total amount of the intensity values would be identical to that of the raw file, plus/minus some interpolation inaccuracies), the same would hold when normalising the jd files into MNI/template space.
Hope this helps
Helmut
|