Neurons, blood flow, and their intimate relationship

As a tangential follow-up the Bloom article and my post yesterday, Jonah Lehrer has a post today noting that Bloom failed to discuss the technical limitations of fMRI as another factor that should curb people’s enthusiasm for neuroimaging. While fMRI certainly has important technical limitations people should be aware of (low spatial and temporal resolution, high costs giving rise to underpowered studies, etc.), I think the issue Lehrer chooses to focus on–namely, the relationship between the BOLD signal (the signal measured by fMRI machines) and underlying neuronal activity–is actually one of the few areas that aren’t controversial. Here’s what he says:

For example, in 2001, Professor Nikos Logothetis, of the Plank Institute in Germany, published a paper in Nature in which he simultaneously recorded the electrical signals of neurons and measured blood flow using fMRI. No one had ever done this before. Logothetis found that the increases in blood flow measured by fMRI do not necessarily parallel increased neural firing rates. In fact, increased blood flow can also parallel a constant, or even a decreasing neural firing rate.

The 2001 paper was indeed seminal—it’s already been cited a thousand times (in just 5 years!)—but not for the reasons Lehrer suggests. What Logothetis actually showed conclusively was that the BOLD signal correlates very strongly with neural activity. There’s some nuance involved (see below), but the basic message is pretty clear. Here’s a figure from the paper that captures the findings nicely:

There’s a lot going on in this figure, but focus first on panels b and d. These panels illustrate the timecourse of both the BOLD (i.e., fMRI) signal and neural activity in response to sensory stimulation. Panel b illustrates the effect of a sustained stimulus (on for 12 seconds, off the rest of the time); panel d shows what happens when you rapidly turn the stimulus on or off (the stimuli here were all rotating checkerboard patterns with varying degrees of contrast). The dark grey blocks are periods when the stimulus is present, and the light grey shows the neural response. The red and blue solid lines reflect the BOLD response. You’ll notice that both lines are delayed a few seconds relative to the stimulus. This is expected behavior: blood takes time to make its way to a particular brain region following increased activation. Since this delay is always explicitly modeled and we know exactly what shape it has, it’s not a problem. In fact, you can see it’s not a problem by looking at the correspondence between the red and blue lines. The blue line is the ‘predicted’ activation—in other words, it’s what you would predict the BOLD timecourse should look like if you assume the correlation between neuronal activity and the BOLD signal is perfect (i.e., if they measure exactly the same thing). The red line is the actual data Logothetis et al. measured. Notice how strong the correspondence is.

If you look at panel c, you can see the association between the two measures quantified. The panel shows the distribution of correlations between the two measures: the x-axis indicates the strength of the correlation (in bins) and the y-axis shows the number of trials that showed a correlation in that range. What you see clearly is that, across all of the samples, including all of the different stimuli lengths (ranging from 4 – 24 seconds), there’s a tight coupling. How strong is it? Well, the modal amount of overlap between BOLD and LFP (local field potential) is 0.67. That number is given in r², which is an index of how much variance two measures share. The 0.67 indicates that the two measures, share, on average, 67% of their variance. If that doesn’t seem very high, consider that in typical human samples, weight and height only share about 35% of their variance. Yet it’s perfectly obvious to just about anyone that, on average, taller people tend to also be heavier. Moreover, the 67% value is the mode for individual trials. Because individual trials contain substantial measurement error, the correlation is likely to be much higher when averaged over many trials (as fMRI studies do)–likely near unity. In sum, there’s simply no question that the BOLD signal reflects neuronal activity. Here’s what the authors had to say about it in the Discussion:

Our results show unequivocally that a spatially localized increase in the BOLD contrast directly and monotonically reflects an increase in neural activity.

And here’s an even stronger sentiment from another Logothetis paper Lehrer links to:

The combination of this technique with electrophysiology has fully confirmed the longstanding assumption that the regional activations measured in MR neuroimaging do indeed reflect local increases in neural activity.

Now, I mentioned that there were nuances involved. Lehrer alludes to this when he observes that blood flow “can also parallel a constant or even a decreasing neural firing rate.” The basic issue here is that ‘neural activity’ is an ambiguous construct. You can measure activation levels in individual neurons, or in a few neurons, or you can measure the local field potential, which is essentially an average of a whole bunch of neurons in relative proximity carrying out a whole bunch of functions. From the perspective of fMRI researchers, it’s really only the last of these that matters. The reason is that the response to stimulation in individual neurons tends to habituate very rapidly. Meaning that if you present a stimulus for 20 seconds, most individual neurons don’t fire at the same increased rate for the entire 20 seconds; they’ll fire rapidly at first and then quickly tail off. Given that fMRI has a temporal resolution on the order of seconds, and that researchers often want to use relatively long-lasting trials, it would actually be disastrous if the BOLD signal was perfectly correlated with the firing of individual neurons, because then fMRI could only provide meaningful data for very brief presentation durations. Fortunately, the BOLD signal correlates most strongly with the LFP, which typically sustains throughout the duration of a stimulus presentation.

All this is basically to say that, based on these results, you can think of the BOLD signal as reflecting the aggregate activity of a whole bunch of neurons located in a particular part of the brain. And that’s exactly what researchers have been assuming all along. (Actually this is still an oversimplification, since LFP reflects not just local processing but also inputs and outputs from other regions. But in subsequent work, Logothetis and colleagues have shown that all of these are highly correlated with the BOLD signal, with inputs and outputs being marginally more influential.)

What about the other problems? Lehrer notes that:

In 2004, Logothetis’ lab found something even stranger . Neurons that had been chemically silenced – they could no longer become active – could still generate an fMRI signal that appeared active.

The irony here is that this finding was actually used by Logothetis et al. to support their argument that the BOLD signal reflects the LFP. What they showed was that applying serotonin to the target brain region (monkey visual cortex) completely abolished the firing of individual neurons, but barely affected either the LFP or the BOLD signal, which remained closely coupled. They concluded:

The response to the stimuli was unaltered, indicating once again the possibility of a total dissociation between spiking activity and hemodynamic responses. On the basis of all of these dissociations, we conclude that the LFP signal is the key variable for the BOLD response.

Again, this isn’t at all problematic—it’s great. It’s exactly what fMRI researchers want out of the BOLD signal. We want to know that greater blood flow means greater general involvement of a region in a cognitive task; we don’t care what individual neurons X, Y, and Z are doing. Now it’s certainly pretty interesting that you can get a dissociation between individual neuron spikes and the local field potential at all; but that’s not an issue that concerns fMRI researchers, since it’s pretty clear that it’s a lower-level phenomenon. Put differently, if the dissociation between individual spikes and the LFP is a reason to question imaging results, then a lot of other areas of neuroscience are in trouble, because local field potentials are used all over the place.

The last three of Lehrer’s points I think I can mostly agree with, but they’re hardly damning criticisms:

It gets worse. A 2002 study by Robert Harrison at the University of Toronto showed that fMRI signals “emanated only from areas endowed with a rich vascular network, and [that] no signals were obtained from adjacent regions in which the vasculature was less dense.”

I wasn’t familiar with this article, so I’m grateful for the pointer. Not having read it thoroughly, I’ll restrict my comments to the following. First, the authors note that capillary density likely develops as a result of a regions demands:

Our working hypothesis is that the capillary density of any brain area develops in direct relationship to the metabolic demand of local neurons.

If this is true, it’s not a huge concern, in that it just means brain regions that do more work than others are going to be easier to detect with fMRI. Second,the study used chinchillas, so it’s unclear to what extent the findings generalize to humans (though the basic organizing principle certainly seems plausible). And third, supposing it’s true that there are some regions in which signal is impossible to detect, such regions are clearly in the minority. Virtually every part of the brain has been demonstrated to show reliable activation in one literature or another. I’ve personally had the (frequent) experience of being frustrated at the fact that too much of the brain is activated by a task; I suspect other researchers have too. So empirically, this doesn’t seem like a major concern, though it’s probably worth keeping in mind.

Furthermore, blood also moves slower than the electricity in our neurons, so it’s always difficult for fMRI to decipher what thought process the blood flow actually correlates with.

This is true, but isn’t really a ‘problem’ per se. It’s an acknowledged limitation of the technology, and every neuroimaging researcher is well aware of the lag between neural activity and the hemodynamic response. So long as it’s modeled correctly (and it isn’t always, but that’s a literature unto itself), there’s no problem in attributing activation to the appropriate timeframe (subject to the general temporal resolution limitation, of course).

Finally, whole parts of our brain remain invisible to fMRI machines. The base of our frontal lobe – a brain area crucial for consciousness – is too close to our nasal ducts to be visualized. The magnetism of air interferes.

They’re not really invisible; they’re just shy. You can get at the orbitofrontal cortex and surrounding areas, but it requires extra catering to if you want to get a really good signal. This isn’t a dirty secret, though; there’s a small literature discussing methods for improving signal-to-noise ratio in different parts of the brain, with people proposing using different pulse sequences, brain orientations, field strengths, and so on. So if you want to do a study you think will specifically activate the OFC (e.g., decision-making or emotion studies often do), you can go the extra step. But the OFC is certainly not invisible, and you do frequently see activation in deeper brain regions even without doing anything different procedurally.

The bottom line is that while fMRI has limitations, just like every other method, it didn’t become the workhorse of cognitive neuroscience just by taking pretty pictures (though it certainly does that). Other methods can in principle take equally pretty pictures of the brain: PET and EEG, for example. But those methods are limited in critical ways (PET is invasive, and EEG has terrible spatial resolution). The reason most people like fMRI is because it optimizes a bunch of trade-offs in a way that previous methods haven’t been able to do. And while Lehrer’s right inasmuch as the question of the mapping from BOLD signal onto neural activity was once a major issue, that concern has mostly gone away now. Which isn’t to say that the BOLD signal always measures exactly what we think it does; there are always going to be circumstantial factors we don’t know about and can’t anticipate. But that’s not a problem with fMRI—it’s a problem with doing research!