With resting state fMRI (rs-fMRI) and functional connectivity booming, and an increasing number of fMRIers adding a resting state scan to their otherwise task-based protocols (even if they don't know what they'll do with the data), the question of whether there is an optimal protocol, perhaps even a standard that could be established across multiple centers, seems timely. From my limited investigation it appears that many fMRIers are doing the logical thing: they are using a version of their task-based fMRI experiment for their resting state acquisition. Is that a good, bad or indifferent thing to do?
A recent review from van Dijk et al. (J. Neurophysiol. 103, 297-321, 2010) set out to determine whether parameters such as run duration, temporal resolution (i.e. TR), spatial resolution (voxel size) and a series of processing steps made any appreciable difference to the detection of default mode and attention networks, against a reference network (a set of nodes not expected to have functional connectivity). Their findings can be summarized as follows:
- Spatial smoothing with a 6 mm Gaussian kernel produced default and attention networks robustly, as detected by ICA and correlation analysis. The addition of subsequent processing steps (temporal filtering 0.01-0.08 Hz, including motion regressors) had minimal effect on the strength of the attention and default networks, but did reduce slightly the "erroneous" connectivity in the reference network. The erroneous connectivity was abolished by regressing with whole brain signal (also reducing the connectivity in the attention and default networks slightly), however, this step also introduced the now well-known negative correlation between the default and attention network. (See, for example, Murphy et al., Neuroimage 44: 893–905, 2009).
- Spatial resolution, either 2x2x2 or 3x3x3 mm voxels, made little difference to the strength of connectivity in attention or default networks, nor to the anticorrelation between them, following all the processing steps mentioned above (i.e. including whole brain signal regression).
- Temporal resolution, either 2.5 or 5 second TR, also made very little difference.
- A duration of about 6 minutes of resting state data acquisition was found to provide a good compromise between total experiment time and robust functional connectivity. Runs shorter than 4 minutes had reduced sensitivity, whereas runs much longer than 6 minutes added little to the data.
- The use of eyes open rest without a fixation target produced similar results to the use of a target, however eyes closed rest and a task (word classification) reduced functional connectivity considerably.
- In a final test, the authors evaluated the use of a breathing waveform-based correction to account for what is possibly the largest physiologically (and non-neural) driven signal change in the frequency range of interest. They used Birn's RVT method (Birn et al., Neuroimage 31: 1536 –1548, 2006) although they discussed the Chang & Glover alternative (Neuroimage 47: 1381–1393, 2009, also 1448 –1459 in the same issue) which I think is conceptually simpler and perhaps easier to implement when using a canonical. In any event, the RVT correction was applied following whole brain signal regression and was found to make no difference to the connectivity results, including the anticorrelation of the attention and default networks. Unfortunately, they didn't try the RVT step instead of whole brain signal regression, only following it, so it isn't possible to tell whether the anticorrelations would have remained when not using an alternative to whole brain regression.
Towards an optimal protocol?
The van Dijk results provide a nice starting point for considering the question of an optimal protocol for rs-fMRI; the positive correlations of the default and attention networks seem robust to just about every parameter choice and processing option (apart from smoothing)! It would appear that if we are interested in getting functional connectivity of attention or default networks, we can select a TR and voxel size that permits the brain coverage we are interested in. And if we are chasing an optimal, potentially standard, protocol then whole brain coverage is a good aim.
What's left? The authors didn't explicitly test slice angle but we can probably assume that we should be using whatever provides the least dropout, as for task-based fMRI. In descriptive terms, then, we should be looking to get the most thin slices we can get in a TR of about 2 to 3 seconds, with in-plane resolution as high as we can achieve, and the slices should be set with a prescription which covers as much brain (including cerebellum - somebody once told me that the cerebellum is also considered part of the brain...) as we can get in the time available (i.e. TR). Finally, we want to acquire a total of 6 minutes' of data if we can, but not less than 4 minutes' worth.
Throwing all of the aforementioned criteria into the mix produces the following approximate parameters for a Siemens Trio: TR=2000 ms, TE=25, 3x3x3 mm voxels, 30 slices. Tilted in an axial-oblique fashion (approximately AC-PC, say) will cover a good fraction of most brains. If we run out of coverage - say we want to acquire the entire cerebellum - then the TR could be extended to 2500 or 3000 ms to increase the number of slices proportionally. For generality, one would have to select the modest setting of 3000 ms and approx. 45 slices, of course.
The anticorrelations introduced by whole brain signal regression are unsatisfying. Presumably there could be valid, interesting anticorrelated networks that must be discarded if they cannot be separated from those introduced as a result of statistics. Thus, it seems to me prudent to have a subject's breathing recorded as another part of our general acquisition strategy. One might then use the RVT or Chang & Glover corrections and omit the whole brain signal regression entirely.
Fixation would seem to be prudent, too. It's hard enough to stay awake in an MRI as it is; allowing a subject's eyes to wander without a defined target is likely not good. But since most people already use a fixation target, I think we can move on.
Some issues that come to mind for me, as someone who doesn't actually do fMRI experiments but helps people set up the acquisition side, is whether some sort of arousal monitoring, or eye tracking, would be useful. It might be sufficient to assure compliance with the "task" - rest, stay alert, don't think of anything, don't fall asleep - though perhaps some clever person will think of a way to use a time course of pupil dilation as a regressor somehow!
I also wonder about the potential of refluxed CO2. Unless you have a well ventilated magnet bore, e.g. you run the fan, there is every chance your subject's hemodynamics are being modulated by this effect. Perhaps this implies monitoring ET-CO2 instead of just the chest motion for breathing rate, I don't know. At the very least, the potential of this issue to introduce another physiologic variable suggests ventilating the magnet bore, just in case.
If spatial coverage per unit time becomes an issue it is tempting to opt for a "go faster" method, such as GRAPPA or partial Fourier acquisition. I would stay away from both for now, for the simple expedient that both methods deliver considerable SNR hits (40% drop best case) and in the case of GRAPPA can considerably enhance motion sensitivity (leading to far worse SNR hits!). Given that the van Dijk results suggest there's little benefit to having 2x2x2 mm spatial resolution, why take an SNR hit for little probable benefit?
The corollary to this situation is MUCH faster temporal resolution. One might expect that if the whole brain could be sampled in under 500 ms then new, interesting dynamics between the networks might emerge. This is wild speculation and is in any event not possible with the present EPI variants available. To get to TR below 500 ms would necessitate reducing the brain coverage to a small fraction, and that's probably not a good basis for an optimal protocol with aspirations of generality. (The cerebellar folks get annoyed enough as it is when their part of the brain fails to make the 'cut.')
I am going to acquire some test data sets using the guidelines above, and will post some analysis here soon. I'll also probably make the raw data available so you can play along at home. (It'll be my brain so there are no HIPAA violations.) I'm thinking of axial full brain, axial-oblique (AC-PC) full brain, an aggressive axial-oblique to reduce OFC dropout, and a sagittal. The latter is just a hunch, and a nod in the direction of the cerebellar folks. Sag is the only sure fire way to get true whole brain (and brainstem) in a reasonable TR. The question is whether the motion sensitivity, distortion and dropout make that option a no-go for some folk.