In the analysis of fMRI data, many use small volume correction (SVC, as implemented in SPM) to restrict their search area to a given region of interest. To my understanding, one can look at both cluster-level and voxel-level statistics within SVC. However, the authors of several articles I've read do not specify if they used cluster or voxel-level statistics.
Do any of you know how people use SVC in this regard? Am I wrong in thinking that both voxel and cluster-level statistics are feasible statistics in SVC? By looking at the t/F values of several papers, I get the impression that most use cluster-level inferences in SVC, but I don't understand why they would not specifically state this, unless voxel-level statistics in some way is not suitable for SVC...
Any input appreciated!
Hi Lasse,
I have used SVC a few times, and reported it in a published paper at least once. My understanding is that the SVC is simply a correction you can apply when your a-priori hypothesis do not apply to the entire brain, but only to certain areas that can be included in a volume. The idea behind it is that standard family-wise error correction procedures assume you will be looking for effects (i.e., significant voxels) all around the brain. When this is not the case, you are effectively adopting a more sever threshold than you need to, thus reducing the sensitivity of your analysis. SVC simply applies the error correction to a predefined volume. When the volume is equal to the brain, SVC should give identical results to the canonical whole-brain analysis. For these reasons, the type of inference that you can do with SVC (e.g., cluster-, peak- or voxel-level) are the same as those that are valid for whole-brain analysis, and carry the same pros and cons.
It makes no sense to use a "small volume" that includes parts of the image where brain activation cannot possibly occur--i.e. in white matter, bone or CSF. Thus the spherical "small volume" normally used in SPM is fundamentally inappropriate, since it would need to be less than 3 mm diameter to not include WM, bone or CSF. It is greatly preferable to use a shaped small volume that includes only grey matter, from one specified and well-defined cortical area, or one deep nucleus. This can only be achieved when a suitable individual-brain cortical and deep brain parcellation has been performed. See my recent paper 'Comparing like with like: the power of knowing where you are' (Turner R, Geyer S. Brain Connect. 2014 Sep;4(7):547-57) to see further details of this approach, which takes advantage of the unique sensitivitiy of MRI to details of myeloarchitecture and iron distribution. To use such small volumes considerably increases the experimental power, and enables well-grounded averaging across subjects and comparison between subject groups. Once this approach has been widely adopted, we will be able to place far more confidence in fMRI findings, in any type of study
To answer your unanswered question, SVC corrections in the literature (as well as FWE corrections generally) refer to voxel-wise corrections unless otherwise specified. Some researchers are very adament that cluster-level statistics should not be used unless there is a convincing theoretical reason for it -- although I do not myself find their argument very convincing.
I am wondering if one can use multiple predefined ROIs to determine statistical significance of several clusters in functional map. For example, if there is a cluster in the right insula and a cluster in the left postcentral gyrus (PCG), is it appropriate to use insual mask and PCG mask separately to estimate cluster size threshold for each cluster in the statistical map?
Dear Mahlega,
Not sure I completely understand your question. But yes you can define multiple a priori ROIs, and use small volume corrections or Monte Carlo simulations to correct for multiple comparisons only in these predefined regions (the regions should be combined for this correction, so the p value adjusts for all voxels within all ROIs at the same time). But if you already analyze the data whole-brain (i.e. without ROIs) and see some significant clusters, it is not appropriate to then perform small volume corrections or in some other way calculate different p values. ROIs should be a priori defined, a common procedure is to first look at ROIs, and then look at whole-brain data. If you do it the other way around you will artifically reduce type-I error rates. This is known as p-hacking.
Hope at least some of that answers your question :)
-Lasse
It's worth pointing out that if the raw fMRI data are spatially smoothed before analysis, the 'spatial extent' of activation and thus cluster size are basically artifacts of the smoothing process, and lose their association with any presumed neural substrate.
My opinion is different. At the end of any SPM map calculation you have an 'activation map' for the whole brain, in any case is for the whole brain. After that you can ask to the map : "where are you significant?". In other words you begin the inference process and , in conseguence, the thresholding process......but , remember the 'activation map' is ever the same. So you must decide how high should be the threshold. Realy it is an very ambiguos question because depends by the number of voxel in wich you want put your attention but , remember the 'activation map' is ever the same....so you could decrease artificially the treshold ( with an artificial decrease of the volume of research) discovering more and more 'significant' clusters or voxel in the same 'activation map'. This is realy a trick or a dodge for to resolve in some manner the problem of a very low power of signal bold in some fMRI experiments as cognitive one's.....
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts