Frequently Asked Questions - Region-of-Interest (ROI) Analysis


1. What's the point of region-of-interest (ROI) analysis? Why not just use the basic voxelwise stats? Are you too good for them or something? Huh, Mr. Fancy Pants?

Whoa, now, no need to get touchy. A lot of people think ROI analysis is a really good idea, possibly where the real future of fMRI lies. Nieto-Castanon et. al (RoisPapers) make the argument kind of like this: brain imaging is concerned, among other things, with analyzing how mental functions are connected to brain anatomy. Note that the word "voxel" didn't enter into that statement. Voxels aren't, on their own, a particularly useful concept for us, but the standard statistical model does all of its preprocessing and analysis on that level. That analysis path tends to completely blur anatomical boundaries, often by a great deal, smearing our resolution to hell and preventing us from making good clean associations between structure and function. So ROI analysis offers us a way to get around individual anatomical variability and sharpen our inferences.

As well, even if we don't start our statistical analysis at the ROI level, ROIs offer us a reasoned way to extract measures that differ from the standard voxelwise t-statistic. Measures like percent signal change timecourses or fit coefficients are additional information that you can extract from your data only by looking within a particular region. These measures can shed light on otherwise obscured aspects of your study - temporal characteristics, particular sizes and directions of effects, or correlations with behavior. Seen from this perspective, ROI analysis is a valuable parallel tool to the standard voxelwise GLM analysis and can often provide new and interesting pieces of the puzzle of your data.

2. How should I generate ROIs? What are the pros and cons of each way?

Funny you should ask; I just happened to have this little grid lying around, which has been helpfully converted into a sort of tree for the Digi-Web. The methods of ROI definition can be split along two axes - the type of brain ROIs are defined on (individual, group, atlas) and the features used to define it (microanatomy/cytoarchitecture, macroanatomy, function). The breakdown goes like this:

  • Microanatomy / cytoarchitecture ROIs

    • Individually-defined: This would mean drawing (by hand or automated method) ROIs on the cytoarchitectonic map for an individual. The pros of this method: you're quite close to the neuronal level of organization, you've got very high resolution on your ROI, and the mapping between cytoarchitectonic regions has generally proven to be close (see Brett et. al on RoisPapers). Cons: Not all functional info is represented (columnar structure, for example, is sometimes represented in cytoarchitecture and sometimes not), and, of course, getting the cytoarchitectonic map for a living human non-invasively is practically impossible at this point. That may change in the future, though - some groups are attempting to mine this data from fMRI maps.
    • Group-defined: This is almost never done - it would mean drawing cytoarchitectonic maps on a template or something.
    • Atlas-defined: This means getting coordinates for particular cytoarchitectonic regions from an atlas, and generally this means using the Talairach Daemon or some similar tool to get Brodmann Areas for particular atlas coordinates. Pros of this: the Tal Daemon's easy to get, easy to use, fast, popular, and can be easily used for MNI-space coordinates with a simple transform. Cons: Lots of sources of error in labeling. The original Talairach BA labeling is very crude (see Brett et. al, RoisPapers) and based on eyeball. The Talairach - MNI transform isn't perfect. And perhaps biggest of all, there is enormouse individual variability of the shape and size of cytoarchitectonic regions - what's BA32 on the atlas may well be deep into BA10 on your subject. Some of those problems are solved by getting a better cyto. atlas - Zilles et. al are working on one - but the problem of individual variability remains.
  • Macroanatomy / anatomically-defined ROIs

    • Individually-defined: This means drawing ROIs on each individual subject's anatomy, based on anatomical markers like sulci, gyri, and other features viewable by the naked eye. This drawing can be by hand or by an automated or semi-automated program (like that at (See SegmentationFaq and SegmentationPapers for more info on those - they're getting pretty good.) Pros: Gets around problem of individual anatomical variability - by extracting each measure from an individually defined structure, you eliminate the issue of whether everyone's structures line up. Allows you to keep your data unsmoothed (smoothing is often done to avoid problems of anatomical variability), and so gives you higher resolution without sacrificing power and even while gaining some. Better mapping of function to structure. Cons: Can be hard to get - requires either good use of an automated algorithm or a lot of labor and experience in hand-drawing ROIs. Makes it more difficult to report activation coordinates (but note Swallow et. al suggest you can normalize your data before defining and be okay - RoisPapers). Most importantly, the relationship between function and macroanatomy is very unclear in many regions of the brain, particularly associational cortex and prefrontal areas. Just because you're extracting from everybody's superior frontal gyrus doesn't mean you're going to get at all the same function. So this path may not gain you any power at all and might lose you some.
    • Group-defined: Again, this is rarely done - it would mean drawing your ROI on some group average. Probably all of the cons of above, with fewer pros.
    • Atlas-defined: This means taking some atlas system's description of anatomical features, like using the Talairach-defined amygdala or Talairach superior frontal gyrus, or possibly drawing an ROI on the MNI template brain. Often the ROIs are then reverse-normalized to fit the subject's non-normalized functional data. Pros: Very easy to get - probably the most-used method for defining anatomical ROIs. Very standardized, and so easily comparable across studies. There's not as much an issue with labeling as above - the Talairach and Tournoux atlas is very accurate at picking out coordinates for macroanatomical features, so you get to leverage their skill at drawing ROIs for your own study. Cons: The overlap of the atlas-defined region with your subjects' is only as good as your normalization - and Nieto-Castanon et. al (RoisPapers) offer a scathing commentary on how good standard normalization algorithms are, showing that even post-normalization, they had huge variations in the shape of particular gyri. Because of that, atlas-defined regions will always tend to obscure differences in anatomical variability, even with reverse normalization. As well, this method is still subject to the problem of unclear structure-function relationships in many areas of the brain.
  • Functional ROIs

    • Individually-defined: This means choosing functionally-activated voxels from each individual's results - either from a localizer task or from the actual study task. This is so popular these days we've actually broken it out to its own page, FunctionalLocalization.
    • Group-defined: This means choosing functionally-activated voxels from your group results and using that voxel set as an ROI to extract from individuals. Pros: Gets to function, as above. Can be a better guarantee that you've got some effect in a given region, since presumably voxels activated in the group had some decent activations in most or all subjects. A good chunk less labor than individually-defining ROIs. Cons: Swallow et. al (RoisPapers) demonstrate these ROIs are not as reliable (i.e., are more variable) than the individually-created ones. This method doesn't account at all for individual functional variability, which can be considerable (i.e., different voxels can be included or exluded wrongly in every subject).
    • Atlas-defined: Probably the closest thing to this is using some other study's reported sites of functional activation - choosing ROIs that correspond to the Tal or MNI coords of other studies' activations. Pros: Avoids the problem of defining localizers, allows direct comparison between studies. Cons: Ignores differences in subject pool and anatomical and functional variability in your own subjects. Also introduces whatever error the other study might have had in localization into your study - if they got their spot wrong, you might be even wrong-er...

Lotta options. All of 'em have their pros and cons - which one you choose will depend largely on the type of questions you want to ask.

3. When can I just look at peak voxels vs. whole regions?

This is still an open question in the literature. The argument for averaging across an ROI is that it should enhance signal to noise; the timecourse from a single voxel can be quite noisy and could, indeed, be some kind of outlier in the ROI. Averaging might give you a better picture of what's happening over the whole ROI. The argument for using a peak voxel, though, is that we know the peak voxel - the voxel that shows the most correlation to the task relative to its variance - is guaranteed to show the best effect of any voxel in the ROI. Additionally, since we know our resolution is blurred by the vascular structure in the region, any spatial smoothing we may have done, and registration and normalization errors, it's entirely possible that some of our ROI's activation isn't reflecting "true" neuronal activity but simply an echo or blurring of activity elsewhere in the ROI. So to average those timecourse together may well wash out our effect, which is after all calculated in voxelwise fashion.

Nieto-Castanon et. al (RoisPapers) choose to look at whole ROIs, and that's arguably the prevailing sentiment in the literature. Particularly with false discovery rate p-threshold correction rising in prominence (see PthresholdFaq), the risk of any given voxel being a false positive might seem too high.

On the other hand, at least one (and possibly more than one) empirical study - Arthurs & Boniface, RoisPapers - has found that peak-voxel activity correlates better with evoked scalp electrical potentials than does activity averaged across an ROI. They cite a couple other studies that have examined similar issues in animal models, and suggest that in mammal cortex in general, the brain may "water the garden for the sake of one thirsty flower," i.e., ROI activity may only reflect true neuronal changes in a few voxels of the ROI. So the question remains open...

4. What sort of measures can I get out of ROIs?

The two big ones that are usually looked at are:

  • beta weights (also called parameter weights or fit coefficients), which are voxel-by-voxel slope values from the multiple regression of your statistical model and correspond with effect sizes of particular conditions.
  • percent signal change or timecourse information - literally looking at the TR-by-TR image intensity at a particular voxel or ROI to get a timecourse of intensities in a particular area. That timecourse is often trial-averaged to come up with time-locked average timecourse, which correspond to the average intensity change following the onset of a particular trial type - essentially an empirical look at the shape of the hemodynamic response to different trial types in a given region.

Other measures are occasionally taken - for voxel-based morphometry, for example, where you might look at percentage of gray matter in a given voxel - but those two are the biggies. Beta weights are used in block-related and event-related experiments; percent signal change is usually more important for event-related experiments, although it's occasionally used for blocks as well.

5. What’s the point of looking at percent signal? When is it helpful to do that? How do I find it?

For everything you could want to know about percent signal change, check out PercentSignalChangeFaq.

6. What are beta weights / parameter weights / fit coefficients? When is it helpful to look at them? What types of analyses can I do with them?

When you run a general linear model to estimate your effect sizes (see BasicStatisticalModelingFaq for info on this), you're essentially running a giant multiple regression on your data, with the columns of your design matrix as the regressors. Each of those columns corresponds to a particular effect, and each of them is assigned by the GLM a particular parameter value: the B in the equation Y = XB + E, where Y is the signal, X is the design matrix, and E is error. That parameter value corresponds to how large an effect the particular condition had in influencing brain activity.

Importantly, beta weights are not an index of how well your condition's design matrix fit the brain activity - it is not the r or r-squared value for the regression. It's the slope of the regression. This means you could conceivably have a very small effect that fit the model incredibly well, or a very large effect with a great deal of noise in the response. This slope corresponds better to the idea of 'level of activation for a particular condition' that we want to find. As an example, a design matrix column that was all zeros might predict the brain activity perfectly - there might be essentially no change in a given voxel down the whole timecourse. In that case, our r-squared would be very high for that column of the regression - but we wouldn't want to say that voxel was active, because it was totally insensitive to any experimental manipulation. It makes more sense to look at how big the effect size was - whether a given voxel seemed to respond very highly to a given trial type - and then, if we're concerned about noise, we can normalize the effect size by some measure of the effect variance, to get a t-statistic. That's generally what's done in most neuroimaging programs these days.

The big reason to extract beta weights at all is that they give you a numerical estimate of the effect size of a particular condition. If the beta for A at a given point for one subject is three times that at the same point for another subject, you know that A had three times a bigger effect in the first subject. This can be an ideal measure to use in regressions against some behavioral measure. For example, you might want to know if a given's subject's self-reported difficulty with a task correlated with the size of the effect of that condition in a particular region. You can extract the beta weights from that regions for each subject, run a simple regression, and find how significant it comes out.

You can also use beta weights to correlate with each other as a crude way of indexing connections between regions. If subjects with anterior cingulates that responded more to condition A also had cerebellums that responded less to condition A, and that correlation is significant across your subject pool, that may tell you something about how the cingulate and cerebellum relate in your task. Check out ConnectivityFaq for more info that direction.

Essentially, beta weights can be used in a myriad of ways - any time you'd like to have some numerical estimate of a given effect or contrast size, rather than simply a statistical measure of the activation.

7. How do I find beta weights / etc.?

Depends on the program, but every major neuroimaging program can create, in the process of running a GLM, an image of the voxel-by-voxel beta weights for each condition. In SPM, these are the beta_00*.img files produced by estimating a model; in AFNI, they're the fit-coefficient or parameter images that can be produced as part of the bucket dataset output. Other programs generally have similar names for these images. Getting the beta weights is simply a matter of using some image extraction utility - something like roi_extract for SPM - to get the voxel-by-voxel intensity values in your desired ROI. These can then be averaged across the ROI, or you can look only at the peak beta value, etc. Check out RoisFaq for more.

8. How do I combine information from an ROI across the whole thing?

The most common strategy in dealing with multiple voxels in an ROI is simply to average your measure across all voxels in the ROI. This has the advantage of being simple to do and simple to explain in a paper. Friston et al. (2006) point out this method may be too conservative; if the ROI has any heterogeneity (say, half of it activates and half of it de-activates), you'll tend to miss things. More complicated methods can be used to identify different subsets of voxels within the ROI with separable responses. Taking the first eigenvariate of the response across voxels from a principal components analysis of the ROI is a simple version of this (and supported in SPM).