Normalization of fMRI data

Hi Neuroimagers.

I’ve been doing some experiments comparing pre-processing pipelines where I do consistent methods for realignment and the like, but change the method and template I use for my normalization (e.g. kid template, adult template, study-specific templates, SPM12/8 vs ANTs and the like). I’ve gotten some VERY different results depending on whether the “final” step prior to analysis is non-linear vs affine registration of templates to MNI space (on the order of ICCs of 0.4-0.95 for the same data).

I’m curious, what do the most popular pipelines do for their normalizations? I read in the SPM12 Manual that DARTEL does an affine registration of the study-specific template to the MNI template for it’s pipeline, and I assumed the @toMNI_Qwarpar or similar AFNI workflow did the same.

Can anyone provide any thoughts (or pipeline clarifications) as to whether an affine is sufficient or if a second-stage non-linear warp is necessary when moving from a study or unique prior template to a standard template?

quick thought: an ICC is a measure of intra- vs inter- subject variance. One way of boosting ICC is to increase inter-subject variance, e.g. through poor anatomical alignment. I would not use ICC as a measure of the quality of spatial normalization.

People instead typically consider some gold standard (manual) segmentation of regions, and check whether a given normalization tool align these across subjects. With this type of metric, non-linear alignment provides a clear advantage.

Now, can linear alignment be sufficient? It is going to really depend on the tool you use after preprocessing. If you apply a crazy deep neural networks with tons of data, it can likely learn how to cope with alignment on its own, maybe better than a generic anatomical alignment tool does. But for standard mass univariate stats I believe the clear consensus is that non-linear is preferable.

1 Like

Apologies I should have clarified:
The ICCs are not on the images themselves, but rather the FC values from the same regions based off one normalization method vs the others.

Howdy-

From an AFNI perspective, the main workhorse tool for nonlinear (NL) alignment is 3dQwarp:
Cox RW, Glen DR (2013). Nonlinear warping in AFNI.
OHBM: https://afni.nimh.nih.gov/pub/dist/HBM2013/Cox_Poster_HBM2013.pdf
In modern times, we recommend using a particular wrapper for this program, depending on whether you are doing human or animal studies; each includes both skullstripping (SS) and nonlinear warping, because these processes can help each other: for humans, @SSwarper, and for animals, @animal_warper. In each case, the outputs can be directly input into afni_proc.py for your full, FMRI single subject processing.

It is hard to imagine any case where you have an analysis that ends with having multiple subjects in a single space that you would not do nonlinear alignment. Having poor alignment basically means that parts of a particular gyrus whose results should be piled together across subjects are not really well compiled, and hence you don’t get a strong build up of the similarities-- it is kind of like having very blurry, noisy data. I can’t imagine a scenario where just affine alignment would be preferred-- even in cases of having non-standard structures in a brain (e.g., a tumor, or changes in shape due to surgery), one can use masks to help reduce the effects of those structures (at least with 3dQwarp et al).

You can see more thoughts on alignment in general (as well as in specific cases) here:
AFNI Academy alignment series

Note that NL alignment is not perfect. And the brain is a tricky subject to align nonlinearly: the variability of structure even within a healthy, homogeneous group is such that some people will have different numbers of sulci and gyri in a particular location-- how can a diffeomorphic, NL warp overcome that kind of difference?? So, there will still be imperfect alignment of structures across any group. And you still need to visually check all your results-- that’s why we try to include automatic images of alignment to help that important step of running a program. (And NL alignment does take longer-- that’s just life.)

Note that you also have to consider cost functions carefully with your alignment-- see here for more details:
Saad ZS, Glen DR, Chen G, Beauchamp MS, Desai R, Cox RW (2009). A new method for improving functional-to-structural MRI alignment using local Pearson correlation. Neuroimage 44 839–848. doi: 10.1016/j.neuroimage.2008.09.037.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2649831/

Our general rule: when you are aligning brains of different subjects (e.g., sub-001 to template, or sub-001 to sub-002), nonlinear alignment is always recommended.

–pt