Authors Boyacioglu and Barth claim remarkable things for the technique:
We find that the spatial localization of activation for GIN is comparable to an EPI protocol and that maximum z-scores increase significantly... with a high temporal resolution of 50 milliseconds.EPI, the current standard fMRI sequence, would have a temporal resolution of 2000 or 3000 milliseconds, so it's about 50 times faster.
Other super-fast fMRI methods already exist (e.g. this one I blogged about), but they've generally achieved speed only at a cost: they've had to either sacrifice spatial resolution to achieve that, or limited themselves to scanning only a small fraction of the brain, or have been more subject to random noise and hence less sensitive.
GIN, however, is said to cover the whole brain, with decent spatial resolution and signal-to-noise ratio. The data can be analyzed in exactly the same way as any other kind. So that's up to fifty times faster with no real drawbacks.
That would be truly revolutionary - as the major limitation of fMRI at the moment is that it's much slower than other methods of recording brain activity.
Check it out: this shows brain activation in response to simple visual stimuli, imaged with bog-standard EPI and GIN:
So this is a big deal... if it does work, I'm sure neuroscientists the world over will be lining up to buy Boyacioglu and Barth a GIN and tonic.
How does it work, and is it all it's cracked up to be? Well, I can't really say: the math is beyond me.
In essence, rather than scanning the brain in 3D, slice by slice (like this), GIN only scans one 2D slice, but then manages to reconstruct the rest of the brain in 3D from just that slice, using dark, forbidden magicks... I mean mathematics. The principle is called parallel imaging and it's been around for several years, but with image quality limitations that GIN claims to have overcome.
Perhaps my more technically-inclined readers will have more insightful comments.
Boyacioglu R, and Barth M (2012). Generalized iNverse imaging (GIN): Ultrafast fMRI with physiological noise correction. Magnetic Resonance in Medicine PMID: 23097342
19 comments:
Fifty times less accurate as well i imagine? Please stop with this machine, it's a waste of time. It's going nowhere, better cut the losses and try and invent a direct method for measuring activity. The indirect method will always have the same error possibilities factored in.
All the money wasted on this thing is better spent on developing something that actually works
"In essence, rather than scanning the brain in 3D, slice by slice (like this), GIN only scans one 2D slice, but then manages to reconstruct the rest of the brain in 3D from just that slice, using dark, forbidden magicks... I mean mathematics. The principle is called parallel imaging and it's been around for several years, but with image quality limitations that GIN claims to have overcome."
I'm sure I read about this technique in the Harry Potter series, when Hermione had to quickly reconstruct 3D image of Harry's brain after a horrific quidditch injury.
People say that Isamov was ahead of his time for predicting future technology, even going so far as suggesting that it would be so advanced as to be viewed as magic. Rowling took the opposite approach, and just used magic to create sufficiently advanced technology.
This uses MEG-like reconstruction so there is a fundamental issue in dealing with the inverse problem. Also, the large array of small coils enhances the motion sensitivity; see MathematiCal Neuroimaging's latest blog post for similar issues as pertains to EPI.
Right now the best hope for faster-than-EPI performance seems to be multiband EPI (or EVI), although this also suffers from increased motion sensitivity. These tools will have to be used with caution, probably with well-trained subjects and perhaps bite bars and other constraints. Definitely not ready for routine neuroscience at this juncture!
Similar to what petrossa said, making fMRI 50 times faster is like using an atomic clock to time the cooking of a chicken. The underlying error isn't in the technology, it's in the biology, and at the end of the day you're still measuring the hemodynamic response, which is just painfully slow.
Even if you got fMRI down to the sub-millisecond temporal resolution of electrophysiology, so what?
It's a neat technology, but I don't see why people try to force it to do something it fundamentally cannot. It's like source localization in EEG: sure it might kinda sorta work, but at the end of the day it's not what the device is best at.
Use the scientific question to guide your tool use instead of trying to force the tool to do something it can't.
@Voytek stole my comment. The effects we are looking for happen on a time scale of 5-10 seconds, so speeding up data-collection won't help fMRI compete with EEG or MEG on timing. It might make it easier to model the (excruciatingly slow) hemodynamic response a little more accurately, though. Which is something.
An extended response. Consider yourself pinged.
Despite the inherent slow hemodynamic response and indirect nature of fMRI measurements (pointed out by other commenters), I still think improving the sampling rate could be considered as a significant advance for particular applications.
For example, in the field of "resting state" fMRI, researchers are not generally concerned with linking brain activity with specific events. The hemodynamic delay is not really a concern if the goal is to examine correlated activity between remote brain regions (assuming the delay is roughly the same in all brain regions). Improved sampling rate would allow us to look at higher frequencies of correlated activity, approaching the type of synchrony that is studied with EEG/MEG.
This is just another "neuroimaging" method hiding the nature of its hideously ill-posed inverse problem.
It's nonsense. How the hell does this stuff get published. Are the editors really this damned ignorant? Apparently so.
Thanks for the comments!
I'm not sure I buy the atomic-clock-chicken analogy though. BOLD lags neural activity by several seconds, yes, but if that lag is consistent (at least within a given subject, and brain area), then it doesn't affect the temporal resolution, it just adds a constant delay factor.
Only if the BOLD lag varies unpredictably would it be a big problem.
Whether BOLD is consistent... I'm not sure we have enough data to say yet... but maybe we need ultrafast fMRI to find out?
IF that lag is consistent. How on earth could it be consistent? Bloodflow in same subject varies across time depending on mood, stressfactors, focus of testsubject, and then a whole list of other factors.
As far as my knowledge goes, admittedly not far but still, the least constant factor in a human being is the bloodflow and therefore you rather need a slower machine then a faster to average out the differences.
At such high sampling rate you just get a lot of extra noise.
How perhaps an fMRI could work somewhat representing what goes on is by doing the same test over and over and then averaging out the results.
The more often the same test is done in the same setting the more likely the result you get is consistent in theory.
However that introduces training in the neural network which in turn influences the results.
A lose/lose situation.
Just build a proper machine. Ask some engineers who know nothing about brains but everything about measuring electrical potential or something like that. That excludes preconceptions.
Neuroskeptic:
Like petrossa said the HRF is far from consistent. We know full well that its shape varies by brain region and age at least.
E.g., http://www.cumc.columbia.edu/dept/sergievsky/pdfs/investigatinghemodynamic.pdf
The fact that so many researchers use the "canonical" HRF as their parameter fit regardless of brain region, subject age, etc. is downright ludicrous to me.
Yeah. Like Voytek said.
(Except for the hit on the "canonical" HRF. For the majority of fMRI studies—which have power at low temporal frequencies—the individual and regional variability in the HRF is insufficient to make much difference in the model.)
Huh, I didn't know that Geoffrey. Interesting.
@Neuroskeptic: EEG or MEG measure effects that happen on time scales of tens of ms. So for you to measure effects of that size using fMRI, you'd need the BOLD response to be like *clockwork*. That is, it would need to peak at exactly (e.g.) 4.75 +/- 0.05 ms. That's really unlikely. Generally, the later the reaction time, the larger the variability (in absolute measurements). I expect that applies to fMRI as well.
More exact measurement of the BOLD response would probably help, but only so much.
There are two questions here. One is whether this specific method really works as well as claimed & the second is whether a fast sampling rate for fMRI is useful.
Whether this method works as claimed is questionable. On a basic level, if you collect more information from the same number of data points, it should be noisier. This isn't a perfect analogy, but it's like taking a picture 10 megapixel picture focused on a person standing on a hill or of the person & surrounding landspace. You get more information, but the information at any one location isn't of as high quality.
In addition, for fMRI, the spacing between excitation of protons in a region affects the steady state signal. If you're exciting the same tissue every 50 ms, it will have significantly less signal than if you excite it every 2000ms. Having more time points balances this out a bit, but I'd be surprised if it makes up for the hit of using such a short TR. Testing a sequence for signal quality on a very robust visual stimulation paradigm doesn't show much.
Regarding faster sampling rate. One big fMRI benefit is that it could really filter out cardiac pulsation to clean up the data. This is useful regardless of the hemodynamic response resolution. Also, the theory behind hemodynamics responses is that for a given piece of tissue, the HRF is part of a linear time invariant system. That means, once you estimate voxel-specific hemodynamic response, you can distinguish events even 100ms apart. This has been shown several times even with much larger TRs. The short TR might not help resolve causality between brain regions, but the fine temporal resolution within a region is definitely interpretable.
bsci: Thanks, that's what I meant about how the theory of a consistent HRF would make this method a lot more useful... if that theory is true.
bsci is correct to separate the discussion into two parts. The bigger debate is whether faster sampling helps with BOLD. (There is a related question concerning spatial resolution, since T2* BOLD has an inherent downstream bias, which in the limits of today's technology equates to a bias towards the pial surface.) Such discussion emphasizes the intrinsic limits of (T2*) BOLD as a contrast mechanism, regardless of how it is sampled/acquired.
The specific discussion related to the inverse imaging approach is entirely technical and mathematical. Is this approach to spatial encoding better than other approaches, e.g. 2D EPI? If so, how? And what are the penalties for the speed? At this juncture, what we find is that the farther away from a 2D FT - which MathematiCal Neuroimaging will tell you is a nice, well defined mathematical approach - the less stable the mathematics of the reconstruction, and the more motion sensitivity is introduced. This has been shown for parallel imaging such as GRAPPA. As a field, we are defining a continuum with 2D EPI at one end and the latest whizz-bang inverse encoding method at the other. (GRAPPA and multiband imaging sit somewhere in the middle at the moment.)
The cost of acceleration from that possible with 2D EPI seems to be some sort of reincarnation of Heisenberg's uncertainty principle: You can have speed or you can have spatial precision, but not both to arbitrary accuracy. Until there are methods that can tackle the enhanced motion sensitivity, then, I see all of these "go faster" methods as very specific, expert-only tools. Use only with extreme caution. (I could go on a rant now about how people only adopt these unproven methods because there are no lives at stake, but I'll quit and save the sermon for another day.)
A timely paper on the usefulness of speed for BOLD-based fMRI:
Measuring Relative Timings of Brain Activities Using FMRI
S.B. Katwal, J.C. Gore, J.C. Gatenby, B.P. Rogers.
NeuroImage in press:
http://www.sciencedirect.com/science/article/pii/S1053811912010555
Thanks, well spotted! It's worth quoting the conclusion:
"High resolution imaging at ultrahigh field, signal extraction via self-organizing map, and appropriate use of Granger causality permit the detection of small timing differences in fMRI data, despite the intrinsically slow hemodynamic response."
However, this was in the primary visual cortex V1, where BOLD responses are very predictable. It might not work elsewhere...
Post a Comment