This is because, in my opinion, quite a lot of today's fMRI work suffers from methodological flaws. But that's not to say that all fMRI work is suspect, or, worse, that there's something inherently unscientific about fMRI as such. fMRI's a tool, an amazing one in a lot of ways, but like any tool it needs to be used well. Along with others, I've criticized various aspects of recent fMRI practice, but only because it's frustrating to see such a powerful tool not being used to its full potential.So I was very pleased by a recent paper by Sabatinelli et al, The Timing of Emotional Discrimination in Human Amygdala and Ventral Visual Cortex. The authors set out to test a hypothesis - that seeing an emotionally charged picture would activate the amygdala and the inferotemporal cortex (IT) before activating the extrastriate occipital cortex.
This is what should happen according to an influential model of how the brain processes emotionally meaningful information. The amygdala is part of a rapid "emotion detector" pathway, which responds faster than the standard visual perceptual system, so the theory goes. You see that it's scary before you see what it is, in other words.
To test the prediction, they scanned a single 5mm slice of the brain - see above - which cut through all of the regions of interest given the hypothesis. Most fMRI studies image the whole brain, but because scanning takes time, this produces one whole-brain image every 2 or 3 seconds.Sabatinelli et al's single slice approach gave them 10 scans/second (TR=100ms), which was crucial given that they were concerned with detecting which parts of the brain activated first. They scanned people while showing them a series of pictures. Some were boring images with no emotional impact, some were "positive" (i.e. porn), and others were "negative" (bloody pictures of mutilation).
The results are on the left. All images activated the visual system more than a blank screen did, unsurprisingly. Both kinds of "emotional" pictures activated the amygdala, IT, and more than the boring ones did (the green line), which is reassuring, since if they didn't, the basic assumptions of the experiment would be in question. And crucially, the emotional vs. non-emotional difference occurred about up to 1s earlier in the time course of the activation in both the amygdala and the IT than in the mOcc (extrastriate occipital cortex), in line with the original predictions.
In itself, this doesn't prove the "rapid emotion pathway" model, but it's an important piece of supporting evidence. It's also a great example of the flexibility of fMRI; while it's often thought of as a way to detect where neural activation happens, as opposed to when, with the right scanning parameters, it doesn't have to be that way. Although there's an unavoidable time lag in the BOLD response that fMRI measures - the response peaks about 5 seconds after the brain cells actually fire - this doesn't stop you from investigating the relative timing of activation in different areas, as in this study.
The key was that Sabatinelli et al had a specific hypothesis and designed their experiment to test it, as opposed to just scanning people under some conditions and looking to see which parts of the brain lit up - fishing for blobs, as it's known. fMRI is a very powerful tool for blob-fishing, unfortunately. But it's also a powerful tool for doing more informative science.
Sabatinelli D, Lang PJ, Bradley MM, Costa VD, & Keil A (2009). The timing of emotional discrimination in human amygdala and ventral visual cortex. The Journal of neuroscience : the official journal of the Society for Neuroscience, 29 (47), 14864-8 PMID: 19940182
12 comments:
Nice study. I'm surprised that they could get 10ms resolution, given the relative sluggishness of the hemodynamic response.
Totally unrelated dumb joke:
How many cognitive neuroscientists does it take to change a light bulb? Just one, but the light will have a 9-16s delay.
I study NEUTRAL faces. I believe my co-activation of pulvinar, amygdala and FFA to be due to subcortical unconscious pre-evaluation (unfortunately, FFA might also inform amygdala bottom-up. and pulvinar is just activated by attentional shift). maybe I should repeat my study this way...
Livia: Yeah, that's why I liked this study, it is quite surprising. As they say, although there's a HRF lag, assuming that this is predictable and constant across different areas, it doesn't stop you working out the relative timing of different areas.
MEG/EEG is always going to better in terms of temporal resolution but fMRI is a lot more flexible than is often assumed.
"assuming that this is predictable and constant across different areas"
Out of curiosity, is there an empirical basis for this assumption? I really don't know a whole lot about this subject.
"assuming that this is predictable and constant across different areas"
BrianW - may want to check out Logothetis' work, hes rather keen on do co-current electrophysiological invasive measures of neural activity (such as local field potentials) and fMRI (haemodynamic responses) in anaesthetised / awake monkeys.
He tends to do V1 related stuff, but invasive techniques on other brain areas would enable some off-the-cuff post-hoc analysis of Local field potential peak to BOLD response.
Theres also choosing which part of the haemodynamic response to image - analysing smaller regions with higher Tesla scanners (or unevenly staggering the acquisition of images) will enable you to see 'initial dips' in BOLD response which should be less variable (as its a effect of increased deoxygenation via neural activity increase rather than the HRF)than the subsequent big positive BOLD response function we're used to seeing in the majority of location-based studies (which typically takes 5 seconds to peak and likely to be far more variable).
There is also individual variability in haemodynamic timecourses for the HRF, which I believe we're just scratching the surface of. So my prediction is initial dip studies would be better for time-based studies using fMRI.
Im not sure if thius posted.
What are the best most recent papers (in your opinion) on Autism, and brain.imaging?
Whats with the fusiform gyrus and amygdala.?
Thanks M
First of all -- disclaimer -- I've not read paper yet but just your post, so I could be utterly wrong in details ;-)
Livia: "Nice study. I'm surprised that they could get 10ms resolution, given the relative sluggishness of the hemodynamic response."
Indeed it is a nice study; I just would prefer to use "Xms sampling" as opposed to saying "Xms resolution" since BOLD simply lacks such (even if you look at those gross-averages with error-bars, it takes seconds to detect significant change between two "neighboring" timepoints) ;) Therefore I get a bit alarmed when people draw conclusions about difference in temporal characteristics of the processes which happen in under 150ms duration from the signal which simply does not have that resolution. I would probably believe more results from mcMRI (AKA DND, AKA msMRI) which are yet to be proved to "exist" ;)
BrianW: ""assuming that this is predictable and constant across different areas"
Out of curiosity, is there an empirical basis for this assumption? I really don't know a whole lot about this subject."
So far I've seen only empirical basis for the opposite: characteristics of BOLD response vary across subjects/areas/experimental design. Especially interesting I find recent studies
doi:10.1038/nature07664
Anticipatory haemodynamic signals in sensory cortex not predicted by local neuronal activity
To briefly summarize: BOLD is not neural activation but some after-product of it, it is highly complex phenomenon, details of which we do not know (yet).
And in this paper, once again, comparison is done across different areas... now it becomes even more interesting to look at amygdala activation, actual plots are PST against 0-onset of the event... not even some mean in preceeding x00ms before the onset (common practice in EEG ERP plots) which could question either there was ANY significant difference between response in amygdala (also pay attention that amygdala PST <= 0.3% whenever in the other areas <= 1.5%). Why does it never return to baseline condition (ie 0) for the green line but rather seems to dive further and further into negatives?
Sorry for seems being even more skeptic than Neuroskeptic ;-)
Neuroskeptic is actually a bit incorrect in describing this paper's methods. The timing of the BOLD response definitely varies across brain regions making it impossible to directly compare timing differences between regions.
Fortunately, this is not what the authors did in this manuscript. They looked for relative timing differences between neutral and emotional stimuli WITHIN each region. Their results claim to show a shorter lag difference between neural and emotional stimuli. At that level, this study is very good and gets around the issues of hemodynamic variations across brain regions.
Unfortunately, I think their method of calculating these lags is a bit flawed. They look at the time series, such as those in figure 2, and identify the time point where the emotion signal magnitude is significantly different from the neural stimuli magnitude. Since this is also a function of the signal-to-noise ratio, it can be affected by the total response magnitude. For example, take two time series with identical temporal pattern. If one is 10 times larger than the other, the time point where they significantly diverge will be earlier than if it is only 5 times larger. Assuming these trials are representative, this looks like it could be the case when comparing the middle occipital cortex & IT cortex. Their task design is good and their concept is good, but I'm not sure the specific anlaysis method actually proves their point.
Also, as far temporal sampling, a 100ms sampling rate is great, but we're measuring a signal that is much slower. This means you can actually use a slower sampling rate and still model the signal as if it were continuous to get very find relative temporal differences (look of Nyquist rate to learn more). That said, the signal is noise so reconstruction isn't perfect and more samples means less noise.
Interestingly, they take their 100ms data and smooth it to 1s. This means they are only using their higher sampling rate for the signal-to-noise boost and to remove some noise related to the cardiac cycle. It doesn't give them any higher temporal resolution of the signals that interest them.
bsci: You're right about their methods. That's what I was trying to say with this statement:
"crucially, the emotional vs. non-emotional difference occurred about up to 1s earlier in the amygdala and the IT than in the mOcc (extrastriate occipital cortex), in line with the original predictions."
Although as you say, this is in the sense that the difference occurred one second sooner in the time course for that region; not necessarily one second sooner in absolute terms.
You're right that signal-to-noise ratio could skew these results, but I don't think this is what's happened here; by inspection the amygdala time course is noisier than the mOcc, but the lines separate sooner; the IT vs mOcc comparison is more iffy but again, by inspection, it looks kosher - the mean lines clearly separate very early.
I've edited the post to make the methods clearer.
Hi Neuroskeptic, can you give me an example of 'blob-fishing'? Perhaps it's because I'm pretty new to the field, but it seems like most studies specify some ROI and look for its response to a particular stimulus, similar to what you've described.
Of course, I might unwittingly only be reading ones that do.
Post a Comment