Dear Johan,
Well, the help text doesn't really state that pushforward is better, does it, it says it's a "newer way". And as stated in spm_cfg_preproc8.m :
---
Actually, in this version of SPM the warped data are not scaled by the Jacobian determinants when generating the "modulated" data. Instead, the original voxels are projected into their new location in the warped images. This exactly preserves the tissue count, but has the effect of introducing aliasing artifacts - especially if the original data are at a lower resolution than the warped images. Smoothing should reduce this artifact though.
Note also that the "unmodulated" data are generated slightly differently in this version of SPM. In this version, the projected data are corrected using a kind of smoothing procedure. This is not done exactly as it should be done (to save computational time), but it does a reasonable job. It also has the effect of extrapolating the warped tissue class images beyond the range of the original data. This extrapolation is not perfect, as it is only an estimate, but it may still be a good thing to do.
---
Which also leaves open whether it's better in practice or not, also e.g. when it comes to normalising fMRI data, for which one would usually want to have unmodulated data. Maybe John can provide some more information on that issue? It would also be interesting to know about recommend settings for when relying on Dartel (should functional data be resliced before as well).
Best
Helmut
|