I’ve been meaning to write a post in this for a while, but as usual, Barry Arrington has prompted me into action (I’m really very grateful to Barry sometimes :)) (Golly, just checked – it’s already half way down the UD page! Does Barry really want his posts buried quite so rapidly? We are going to see fossilisation at this rate!)
Anyhoo…. Neuroimaging is one of the things I do. Here is one of my favorite images (probably the most reproduced fMRI image of all time), by Fox et al, 2005:
Although it may not have the form that some readers might be more familiar with, as it’s plotted on a “flat[ish] map” of the cortical surface.
Here’s another nice one though:
Both of them are functional MRI images. One of the things MRI can do is to measure the oxygen in the blood, what’s known as the Blood Oxygen Level Dependency, or BOLD. Because neural activity is followed by a surge in blood supply to that neural region, peaking at about 5 seconds after the neural activity has occurred, and subsiding via a short underswing over about 30 seconds. This means that the BOLD signal is a nice proxy measurement for neural activity in the brain. There are two main problems with it, however: one is that not all brain regions have the same haemodynamic lag (some areas might peak earlier, some later), and the other is that the signal is heavily low pass filtered in time – we are seeing the trace of neural activity averaged over a substantial time window, moreover one that is really very slow compared with the rapid oscillatory dynamics of neural electrophysiological activity.
Nonetheless it is an amazing technique, but it’s probably worth saying a bit about the blobs. What are they? Well, essentially they are statistical values – usually t, sometimes f values. Still the most common way of generating an fMRI image is to get the participant to undertake an actual task in the scanner, or present them with some kind of sensory stimulus. Because the time course of the stimuli can be measured directly (indeed, it is usually pre-programmed) we can set up a simple linear model in which we model the expected haemodynamic response to a stimulus as the predictor variable, and the time course of the BOLD signal in each voxel (like a pixel, only a volume) of the brain, as the dependent variable (we also throw in various nuisance covariates, such as the log of the participants movements in the scanner). And then it’s just a simple multiple regression, just as you’d do if you wanted to know whether, say, smoking was a significant predictor of blood pressure after covarying for age and weight, say. Except that we do it for every voxel in our brain image. Then we look at the regression coefficients for our predictors of interest, and plot the t values in each voxel (generally thresholded at a p value that takes into account the very large number of voxels we have put into our analysis).
So when people say a part of the brain “lights up” when we do a certain task, what that actually means is that those task stimuli are a significant predictor of the BOLD response in that region, suggesting that those stimuli evoke a neural response in that region. However, we can also do other kinds of analysis, including measuring correlations between the BOLD response in different voxels. We can take a single “seed” voxel in some region of interest, and plot the correlations between that voxel and all other voxels in the brain. Or we can take several Regions of Interest (ROIs) and use graph theory to characterise the networks of correlations between them. Or we can use techniques like Independent Components Analysis (ICA) to pull out networks of regions that tend to coactivate, or to be negatively correlated. The first image I posted is from Fox et al, 2005, in which they used seed-region analysis of a BOLD time series acquired not during a task but during rest, and revealed three distinct networks, two positively correlated with each other, and one negatively correlated with the other two. The last one is usually known as the Default Mode Network (DMN), for reasons that are largely historical, and it is the one shown in orange/pink. The others are shown in blue/cyan. The time course of each is shown below, and you can see that the brain spontaneously “flips” between .one and the other at a frequency of about .1 hz (an example of highly functional “chaos”!)
Since Fox et al, many other networks have been identified, some of which are shown in the second image, and they are readily replicated. Some are evoked by specific tasks, and many tend to be “task positive” – the BOLD signal in these networks tends to be higher following a task-relevant stimulus, while some, including the DMN tends to be “task negative” (reduced following a task-relevant stimulus, probably because it is active when we are attending to our “internal world” rather than the external world.
And we also can identify these networks using Magnetoencephalography, which measure the magnetic field generated directly by neural elecrophysiological actiivity, and, while having less good spatial resolution than MRI, has extremely good temporal resolution (we usually sample at 600 hz).
Here are some classic MRI networks also elicited using MEG by one of my colleagues:
The upper images in each panel are MRI, the lower are MEG. This tells us that the BOLD really is a good proxy for neural activity.
However, as Barry rightly says, these images are models. In fact they are fitted models – models fitted to data. And I’m going to make the point here, briefly, that prompted me to post this thread now, in response to Barry’s post. Barry says (to poster Joealtle):
Joealtle, I am going to clue you and your Darwinist comrades in here, so please pay close attention now: Your interpretation of the data is not the same thing as the data. Write that down.
Way before you, Barry 🙂 Yes, a fitted model is not the same as data. However, I will go further. The word “data” means “what is given”. And what is “given” at one level, may be the model of what is “given” at the next level down. I am not a physicist, so my “data” is a series of BOLD signal measurements that come off the scanner. I fit my models to those BOLD data. But those data are themselves models – they are models made using certain assumptions about magnetic resonance, and the change in spin of protons when exposed to different radiofrequencies in a large magnet. So my data are the physicists models. And my models then become data at the next level of analysis which is when we often compare the regression coefficients from one set of participants with those of another, and get a new set of regression coefficients from that fitted model. Then someone comes along and takes regression coefficients (or other model parameters) from a whole range of studies and does a meta-analysis, where those models become the input data – “what is given”. And so on.
So yes, Barry. In science we learn at our supervisor’s breast that all we have are models – even our measurements are models. And our whole hypotheses are a giant layercake where the data at the bottom are the models for the next, and so on. But what we must never forget is that the models are always fitted to the data, not the other way round.
Write it down, Barry 🙂
I should say that some of the worst fMRI studies are the ones that hit the headlines – ones that correlates the regression coefficients from single subject data on some task with some kind of personality measure, and when they get a hot spot in the second level analysis say, say “hey! here’s the bit that makes people kind/gnarly/good at math/religious” or whatever.
They are almost all crap.
Raymond Tallis had an excellent article in the Guardian last week about that.
Think brain scans can reveal our innermost thoughts? Think again
Lizzie,
Thanks for the explanation of what the colored regions mean, what it means for them to “light up”. So a gray region doesn’t mean that the region is inactive, just that its activity is not significantly greater than normal.
This graphical summary is then somewhat misleading. Most of us, me included, would read the gray regions as “off”.
I don’t believe any neuron is ever off.
All this lighting up just means to show activity in a part of the brain. Yet its not showing the brain is doing the thinking. Our soul does do the thinking and it being connected to the brain/body would have exactly this same result.
There is no evidence to connect the brain to anything intellectual in man. Its just a line of reasoning from presumptions of man being only a machine.
There can be no breakdown in mans thinking because of breakdown in the brain. Its impossible from a Christian belief in the soul.
The only breakdown that can interfere with our thinking is the memory. Except for the issue of sleep etc which just turns off things.
I have been reading recently about autism and memory and it seems they are coming closer to a conclusion autism and etc are just the result of memory interference. So perhaps a better chance of healing then before it required fixing brain parts.
petrushka,
True, but the gray coloring gives you the impression of minimized activity.
What is the brain for?
How do you explain the case of Phineas Gage?
How do you explain drunkenness?
The grey just means that t value for the regression coefficient was either below whatever the threshold p value value was for that voxel, or negative. Because there are so many voxels in the brain, the multiple comparison’s penalty is severe, so Type II error is an issue.
And then there is the issue of significantly negative coefficients. In a simple fMRI task, where we just want to compare stimulus vs baseline, some voxels will have significantly positive regression coefficients for the stimulus predictor, and some negative, in other words, some brain areas will be actually suppressed below baseline. The DMN is usually significantly negative, and so will only show up on an image where both positive and negative values are shown.
Some fMRI images show F values rather than t values, and as F is unsigned, you get any region that is significantly altered from baseline by the stimulus, rather than just those regions that receive additional blood flow.
However, that begs the entire question as to what we mean by “activity” – BOLD is only a proxy measure for neural activity, and probably measures brain areas that are oscillating at a frequency that has higher metabolic demands, like gamma, or some of the lower frequencies, like theta. Beta oscillations (about 13 to 30 hz) seem to have the lowest metabolic demands, so fMRI “blobs” are correlated with reduced beta power. In fact in those MEG images, most of those networks were elicited by correlating beta-power between voxels (one was correlated alpha power).
One interesting measure of “activity” in MEG we are starting to use at the moment is something called “rank vector entropy” which might interest our ID proponent members 🙂 The fMRI BOLD response may be correlated with oscillatory activity that is very “white” – lots of power at lots of different frequencies, while a low BOLD signal may indicate a “purer” set of oscillatory frequencies. In fact, being a musician and all, I have been playing around with shifting the MEG signal into the auditory range (just jacking it up by a few octaves) so that I can “hear” the brain. It’s fascinating! The “shhh” and crackling sounds may be what produces the BOLD signal, while the pitched sounds may indicate lower BOLD. But I have yet to test this hypothesis 🙂 I need a better mixer. (I’ve been mixing channels spatially, and you can hear signals from different brain areas moving across auditory space :)).
How would that work without violating the laws of thermodynamics?
Well, it probably means reduced oscillatory activity in the beta band, and increased activity in other bands. But we don’t know yet, only that we have very good correlations between BOLD and stimuli.
No, it doesn’t. But it does show that when we see things or do things, the blood flow to the brain changes. We can also compare the blood flow between conditions, so that we can see regions that get additional blood flow when we make an error, for instance, or have to do something harder.
Yes, indeed. In fact with fMRI we assume that the BOLD response is a response to neural activity. It could in theory be a response to soul activity. But in fact we are starting to have evidence that it is a response to neural activity, and in any case, electophysiological neuroimaging tells us that neural activity is also highly correlated with the stimuli we receive, and how we react to them, including the decisions we make.
So we know that oscillatory neural activity is intimately related to thinking. And we know a great deal about how neuronal populations oscillate. So it’s not easy to know what job a soul would do if it had one.
But, you never know 🙂
Oh, there’s a vast amount of evidence, and has been for centuries. If it were not the case, head injuries wouldn’t have an effect on intellectual function, and we know they have specific and often devastating effects
In that case the belief must be is incorrect! But it’s perfectly possible, in my view, to have a belief in the soul that is compatible with a neuronal description of brain function. But be that as it may, the fact is that brains are vital to intellectual function – the evidence is massive and unambiguous.
Memory is certainly important but does not capture all that neurons do. I’d say that both memory and imagination are forms of “modelling” the brain does, and both of these are vital for learning and thinking.
Fascinating OP. I am not sure I understand about the significance levels. What is the null hypothesis?
Depends on the model. For a simple model in which you are simply interested in the effects of a stimulus on the brain, the null hypothesis is: no difference from baseline, in other words you use a one sample t test that the regression coefficient is not different from zero.
You can also test the hypothesis that the difference between regression coefficients for two different regressors (for instance stimuli to which the participant gave a correct response, and stimuli to which the participant made an error) is not different from zero. That way you can find what brain regions receive greater blood flow, for example, following an error.
That’s at within-subjects level.
You can then do a random-effects analysis where you pull the betas through from the within-subject analysis, and compare betas of one group against another.
etc.
Let me know if that doesn’t make sense 🙂
Lizzie,
Thanks – in that case how do you determine the base line ? I am just curious – no need to respond if it is tedious to explain.
Barry quotes a UD poster called Turell as saying:
Turell is of course correct that fMRI measures the bloodflow that follows certain kinds of neural oscillatory activity, not the neural oscillatory activity in real time.
However, from there onwards, I’m afraid what Turell writes is completely unjustified. First of all, nearly all (and in some senses all) measurements in science are proxy measures of the the thing we are interested in. Particle physicist do not directly observe particles, but the effects of those particles on other things. Same with fMRI – we measure the effects of neural activity on blood flow. And we know, from many other kinds of neuroimaging and empirical neuroscientific findings that fMRI is indeed highly correlated with neural activity. One thing we can do, for instance, though it’s a technical nightmare, is to do simultaneous fMRI and EEG.
Secondly, Turell seems unaware that one of the reasons we know the brain is functionally interconnected is – wait for it – from fMRI functional connectivity studies! Yes, a lot of “blobology” as we like to call it is, indeed, garbage, but that’s not because the BOLD signal is not an excellent measure of the timecourse of neural activity, albeit averaged over a substantial time-window, but because there are some really crap statistical analyses out there.
And I don’t think the problem is “the buddy system of peer review”. I think it’s a publishing problem, but my take is that it has more to do with the “sexier” a finding (“Hey, we’ve found the God spot!”) the more likely it is to go to a journal with wide readership, and be sent to non-specialist reviewers. Post-publication peer-review is much more important, as well as meta-analyses, and that’s what eventually weeds out the chaff.
There are far more good fMRI studies than bad, but the bad ones make the headlines because spectacular findings are far more likely to be wrong than less spectacular findings.
However, in my experience, it’s not spectacular findings that get you grants – it’s really solid work with clear translational applications. Which is generally much more time-consuming and expensive. So I suspect it’s tempting to get a load of cheap students in to do a personality questionaire, bung them in the scanner to do a task, and hope for a sexy finding, which you can write up quickly and get published in a fancy journal. But the real expensive work is, I’d say, much more solid.
Not at all 🙂 I can happily do this stuff or hours. In fact I just finished giving a tutorial on preprocessing to a new imager 🙂
It’s important when setting up an fMRI experiment to include time not spent doing the task – either periods of rest, or spacing out the stimuli enough that the predicted haemodynamic response has had a chance to drop.
For a “block design” it’s easy – you just have rest periods between blocks of doing the task . For an event-related design, you get a time stamp for each event (say the onset of a task-relevant stimulus), and then convolve either that as a stick function (if you are interested essentially in an intaneous event) or a short box car, with the predicted haemodynamic response function which usually looks like this:
You usually include a temporal derivative as well, in the model, to give a bit of play in the timing of the peak.
And what you should end up with is a kind of wave function that rises when you predict the BOLD to be high following an event, and low when you predict it to be low.
But you need to ensure that your design has enough gaps for the predicted response to drop.
That’s the standard way of doing it, using the GLM – what’s called Statistical Parametric Mapping. There are also other techniques, including non-parametric models. But SPM has been hugely successful, and I have a soft spot for it as it was born in our kitchen 🙂
My husband hired Karl Friston to figure out a voxel-based parametric mapping system for PET imaging (which is why googling my surname and imaging gets you a lot of hits – it’s mostly by marriage :)). And he did.
And it’s conceptually very simple – just what he calls a “massively univariate” regression model, carried out on every voxel, the clever bit being the technique for correcting for multiple models, given the huge number of voxels.
Well, all neurons are in some state of polarization/depolarisation or other at all times.
Lizzie,
Just to push this one step further:
“not doing the task” could in theory cover a vast range of mental activities from sleeping to intense mental arithmetic. Is there some kind of generally accepted neutral state for mental activity which acts as the universal base line?
No, not at all. In fact we know (for example from the Fox et al paper from which the first image is taken) that we activate and deactivate exactly the same networks during rest as during a task. The point is that we are comparing predicted event-related BOLD against the average unmodelled data.
As an analogy: let’s say that we see a machine outputting lots of coloured beads. there at first appears to be no systematic way of predicting which colour comes out next. But we observe that there are a number of workers ladling dye into vats, and that the plastic in the vats takes 2 minutes to reach the bead-stamper. And we say: let’s see what colour bead comes out 2 minutes after Worker A lifts her ladle. And we observe that whenever Worker A lifts her ladle, two minutes later out comes a blue bead. And when Worker lifts her ladle, two minutes later out comes a red bead. Quite often blue and red beads come out at other times, from which we can deduce that there are lots of other workers with blue and red dye as well, but because Worker A and B are consistently associated with red and blue beads, we can deduce that there is a temporal relationship between their actions and the bead colour. Moreoever, if we actually experimentally manipulate the actions of Worker A and B, we can establish a causal relationship.
Probably a really stupid example! A better one would be a study that show that in the UK, death rates are higher following surgery on a Friday on other days. The baseline isn’t zero deaths, but the mean rate across other days of the week.
Your analogy works fine. I am just surprised because I would have thought that independently of the worker’s activities the different vats would be producing a range of different colour patterns and those patterns would vary over time (depending whether they were doing the vat equivalent of watching TV, doing maths, playing squash etc) so it would be impossible to pick out a significant signal from the noise. Obviously it is possible (I am not querying the science!) but surprising (to me).
Yes indeed, which is why picking out the signal from the background is challenging and the statistical power is low. Nonetheless, we do get repeatable results for certain kinds of hypotheses. Where I emphatically agree with Raymond Tallis (and I agree with him on almost everything!) is that we are miles from (and will never get to) the point where we can look at a brain image and tell what someone is thinking. We are looking at the effects of many millions of neurons, and are much better at imaging what different tasks have in common than in what differentiates them, unless they are radically different.
The really interesting thing about the last decade or so has been the discovery of a relatively small number of networks of coactivated regions that seem to be involved in a large number of tasks, in a coordinated manner, on a slow time scale.
But the specific thought content is going lie at a level that we will never have access to (except for very crude differences like telling the difference between someone thinking or viewing a house, as opposed to a face).
I think the troubling scifi screnario is that a skilled interrogator could cold read your thoughts by observing the reactions to questions.
To some extent I think that is a real worry. Or rather, what is an even greater worry is over-confidence by investigators in their answers. Just as in polygraphs.
You can be damn sure that if creationists had models that had any actual value in explaining the world they would do everything possible to divert attention away from them being ‘just models’. When you haven’t got the goods dismiss your opponents goods as ‘just models’.
davehooke,
The brain is to connect the inanimate soul with our physical body. To make the thing work. I don’t know PG and drunkenness is simply a interference with the triggering mechanism for the memory as i see it. Mental retardation is a form of drunkenness .
Lizzie,
All these neurones etc should not deceive that they show a brain thinking machine at work.
All they show is activity in the brain. Yet the soul would do this also if it was using the brain for this or that or just for memory operations.
Head injuries simply need to be seen as injuring the triggering mechanism for the memory. Its possible they might disconnect some things but only basic operations. Not ones thinking ability. Yes moving ones hand. We are in a machine but not oif the machine. No reason to see it that way I say.
I don’t think modelling ideas are needed. We just are thinking beings that must use the memory machine in our bodies as all creatures must.
All problems with peoples “minds” are simply interference with the triggering mechanism for memory and simply there is a great spectrum.
Forgetfullness for most comes from distraction/memory on some other thing to the far end of the spectrum of savants who memorize books in a hour but can’t dress themselves.
Its all about memory and its triggering and nothing to do with a soul made in Gods image.
Its impossible for a Christian to believe his soul, which travels to the afterlife fully thinking, can be affected by his brain. Only the machine of the body gets in the way with info.
Oh that’s going up on FSTDT.
I see that Denyse O’Leary is touting what looks like an excellent book, Brainwashed: The Seductive Appeal of Mindless Neuroscience by Sally Satel and Scott O. Lilienfeld.
But I am getting a bit fed up with all the neurobashing. What hits the headlines is simply not what the vast majority of neuroscientists use the tools for, and, as Satel and Lilienfeld point out in their introduction (on Amazon preview), make most neuroscientists “cringe”.
You don’t have to be a dualist to see where the problem is, and this straw man is really starting to bug me 🙁
Lizzie,
Its really taking advantage of accepted problems in this field and saying, rightly, the whole field is never been done well.
Its presumptions are its failings. its too evolutionary concerning the mind/brain of humanity.
There is so much wrong some authors see a chance to get a book out of it.
The creationist sees a bigger presumption problem and sees, as predicted, the whole field going nuts about hypothesis.
The field goes wrong in denying the soul as a option behind all human thinking and wiring.
I think the “field” as you put it, simply ignores the concept of a “soul”, and will continue to do so until there’s a means of investigating it.
In the meantime, the science barrels on, investigating that which can be investigated, and improving the instruments and methods as it goes. When they see the flicker of the soul’s fin in the murky waters of the mind, it’ll be brought into the picture, no doubt.
The soul is a famous conclusion of religion and makind.
To ignore it is a statement they are ignoring the intellectual heritage of a belief in the soul.
The soul is a legitimate hypothesis for explaining human thought and intelligence.
Dismissing it out of hand and then going nuts about ideas of what human intelligence is and is made up of is a invading aggression on history on this subject.
Its more likely we think in our soul and the body/brain is jusrt a middleman .
There is nobody home in the parts or accumulated parts called the brain.
Thinking there is leads to strange hypothesis people can write books about.
That’s fine, Byers. You’re perfectly entitled to your opinion, and I to mine. I find the “conclusions of religion” of limited utility at best, and useless, or cruel, at worst.
I hope you and your soul ( or you and your body) will be very happy together.
This is a great post! Very clear. Thanks Lizzie.
For the record, I saw Raymond Tallis at a talk in London. The whole talk was a waste of time. Lots of straw-men; smoke and mirrors. Maybe his books/articles are better.
Once again, thanks for this clear post.