Regional Homogeneity Analysis, Part II: Processing Pipelines

For regional homogeneity analysis (ReHo), many of the processing steps of FMRI data remain the same: slice-timing correction, coregistration, normalization, and many of the other steps are identical to traditional resting-state analyses. However, the order of the steps can be changed, depending on whatever suits your taste; as with many analysis pipelines, there is no single correct way of doing it, although some ways are more correct than others.

One of the most comprehensive overviews of ReHo processing was done by Maximo et al (2013), where datasets were processed using different types of signal normalization, density analysis, and global signal regression. I don't know what any of those terms mean, but I did understand when they tested how smoothness affected the ReHo results. As ReHo looks at local connectivity within small neighborhoods, spatial smoothing can potentially inflate these correlation statistics by averaging signal together over a large area and artificially increasing the local homogeneity of the signal. Thus, smoothing is typically done after ReHo is applied, although some researchers eschew it altogether.

So, should you smooth? Consider that each functional image already has some smoothness added to it when it comes directly off the scanner, and also that any transformation or movement of the images introduces spatial interpolation as well. Further, not all of this interpolations and inherent smoothness will be exactly the same for each subject. Given this, it makes the most sense to smooth after ReHo has been applied, but also to use a tool such as AFNI's 3dFWHMx to smooth each image to the same level of smoothness; however, if that still doesn't sit well with you, note that the researcher's from the abovementioned paper tried processing pipelines both with and without a smoothing step, and found almost identical results for each analysis stream.

Taken together, most of the steps we used for processing our earlier functional connectivity data are still valid when applied to a ReHo analysis; we still do the basic preprocessing steps, and still run 3dDeconvolve and 3dSynthesize commands to remove confounding motion effects from our data. However, we will only do smoothing after all of these commands have been run, and use 3dFWHMx to do it instead of 3dmerge.

One last consideration is the neighborhood you will use for ReHo. As a local connectivity measure, we can specify how much of the neighborhood we want to test for correlation with each voxel; and the most typical options are using immediate neighbors of 7, 19, or 27 voxels, which can be specified in the 3dReHo command with the -nneigh option. "7" will mean to consider only those neighboring voxels with one face touching; "19" will calculate the time-series correlation with any voxel touching with a face or an edge (e.g., a straight line bordering the current voxel); and "27" will do the analysis for any voxel with an abutting face, edge, or corner (think of it as the test voxel in the center of a Rubik's cube, and that voxel being correlated with every other voxel in the cube).

In the next part, we will go over some important rhetorical questions, along with a video showing how the command is done.

Head Motion and Functional Connectivity

Yesterday as I was listening to a talk about diffusion tensor imaging, a professor asked about the influences of head motion on DTI data, and whether it could lead to spurious effects. Another professor vigorously denied it, stating that it was much more of a problem for bread and butter FMRI analyses, and in particular resting state functional connectivity analyses. At one point he became so animated that his monocle fell off, his collar stud came undone, and eventually he had to be physically restrained by the people sitting next to him. It was then that I knew that I should pay heed, for it is unlikely that a scientist becomes highly excited and talkative over matters that are trivial; in short, I could sense that he was talking about something important.

I have done few functional connectivity analyses in my brief life, but what I understand about them is this: You take the timecourse of one voxel - or the average timecourse over a group of voxels, also known as a "seed" - and then compare that timecourse with the timecourse of every other voxel in the brain. (When I speak of timecourses, I mean the sampled signal over time.) If it is a good fit - in other words, if there is a significantly high correlation between the timecourses - then we say that the two regions are functionally connected. This is a bit of a misnomer, as we cannot make any direct inferences about any "real" connectivity from two correlated timecourses; but it can serve as a good starting point for more sophisticated analyses, such as psychophysiological interactions (PPI; also known as context-dependent correlations) which measure changes in functional connectivity as a function of task context. For example: Does the timecourse correlation between cognitive control regions and reward-related areas change depending on whether the subject is instructed to upregulate or downregulate their gut reactions to rewarding stimuli?

One of the most popular variations of functional connectivity is something called resting state functional connectivity (rsFC), where a subject is simply instructed to relax and listen to Barry Manilow* while undergoing scanning. Functional connectivity maps are then calculated, and usually a comparison is made between a control group and an experimental or patient group, such as schizophrenics. For us FMRI researchers, this is about as close as we can get to simulating a person's natural environment where they would be relaxing and thinking about nothing in particular; except that they're in an extremely tight-fitting, noisy tube, and unable to move in any direction more than a few millimeters. Other than that, though, it's pretty normal.

These types of experiments have become all the rage in recent years, with several studies claiming to have found meaningful resting-state differences between healthy controls and several different patients populations such as schizophrenics, depressives, Nickelback fans, and drug addicts. However, a few publications have called into question some of these results, stating that many of these differences could be due to head motion. As we've talked about before, head motion can be a particularly insidious confound in any experiment, but it is especially troublesome for functional connectivity analyses. This can arise due to systematic differences between control and patient populations that are possibly confounded with motion. Take, for example, an experiment contrasting young versus older populations. Older populations are known to move more, and any observed differences in functional connectivity may be due solely to this increased motion, not underlying neural hemodynamics.

A study by Van Dijk, Sabuncu, & Bruckner (2012) looked at this in detail by scanning over a thousand (!) subjects, and binning them into ten groups based on increasing amounts of motion (e.g., group 1 had the least amount of motion, while group 10 had the most motion). The authors found decreased functional connectivity in the "default network" of the brain - usually referring to the functional connectivity between the medial prefrontal cortex and retrosplenial cingulate cortex -, decreased connectivity in the frontal-parietal network, and slightly increased local connectivity among clustered voxels, simply based on motion alone. (Keep in mind that each bin of subjects were matched as closely as possible on all other demographic measures.) Furthermore, even when comparing bins of subjects closely matched for motion (e.g., bins 5 and 6), small but significant differences in functional connectivity were seen.

Figure 3 from Van Dijk et al (2012). Functional connectivity among different networks measured as a function of head motion. Both linear and nonlinear (quadratic) terms were modeled to fit the data.

Figure 4 from Van Dijk et al (2012). Note the comparison on the far right between groups 5 and 6; the mean motion difference between these two groups is a few thousandths of a millimeter, but noticeable functional connectivity differences are still seen between the two groups.

Lastly, a subset of subjects were rescanned in order to see whether motion was reliable; in other words, if a subject that moved a large amount on one day had the same tendency to move a large amount on the next day. A clear correlation was found between scanning subjects, suggesting that motion might need to be treated as a trait or individual difference, just like any other.

Figure 5 from Van Dijk et al (2012). There is a robust correlation between the movement of scanning sessions, even with the outliers removed (marked in diamonds).

So, what to do? A few recommendations are to match subjects for motion, correct motion prospectively (Ward et al, 2000), and regress out motion when performing a group-level analysis, as you would any other covariate. Apparently traditional methods of motion correction on a subject-by-subject basis are not enough, and increasing awareness of the pitfalls of between-subject motion is important for evaluating current functional connectivity analyses, and for conducting your own experiments.

This study hit me in the face like a wet mackerel since I am beginning to investigate a recent AFNI tool, 3dReHo, to do local functional connectivity analyses for publicly available datasets on the ABIDE website. However, as far as I can tell, motion limits were not used as exclusionary criteria, which may be a possible confound when examining, say, autistic children to controls. More to come soon. Or not.

*I Don't Want to Walk Without You