Functional MRI, Round 3: Six Items to Keep in Mind

This is the third and final installment in a series on biophysical mechanisms of functional magnetic resonance imaging (fMRI) technologies. My overarching goal has been to explain why great care must be exercised when interpreting data derived from these magnets. The inspiration for the series came as I was reading a magazine article while waiting for a plane to take off—my reaction to what I read may have resulted in a bit of trauma to the seat pocket in front of me.

In part 1 , I talked about how innocent little protons spinning inside biologically important molecules end up giving us insights into brain function. The protons get caught in powerful magnetic fields; are blasted with radio waves; then allowed, gasping, to return to their former, lower-energy state. This gasping (energy release, actually) is what an fMRI machine actually detects.

In part 2, we discussed 2 facts:

• fMRI can only measure changes in blood flow (something called BOLD signals).

• There exists a somewhat ambiguous relationship between hemodynamic changes in the brain and neural activity. Although neural activity is supposed to be associated with an increase in blood flow, that’s not always true. Sometimes increases in neural activity result in a decrease in blood flow.

We ended by observing that there is an array of switching mechanisms from which the brain has to choose when assessing energy/oxygen needs; I then gave a detailed example of 1 such mechanism. A full understanding of all these mechanisms would have to occur before a completely accurate interpretation of imaging data emerges.

Here I formalize this need for caution by describing 6 items to keep in mind when examining fMRI images. The reason to end our series this way is altruistic (I hope): the view these scans provide about brain function is quite spectacular—and rendered all the more remarkable by conservative, thoughtful interpretations about what that revelation is.

Here then, are 6 things to keep in mind.

Item to Keep in Mind #1

You are always looking at machine-selected populations.

Not everyone who signs up for fMRI brain experiments can actually carry them out. Researchers estimate that up to 20% of the subject pool becomes claustrophobic as they climb into the machine. This makes it impossible for them to start (or in some cases, continue) the imaging process. Even those who stay put often report feelings of anxiety—especially as the machine groans into action. (This is not all that surprising, given the almost coffin-like tube into which the subject must crawl). Subjects also have to keep their heads perfectly still, sometimes from a few minutes to a few hours, while locked inside the tube (to keep the image as clear as possible). This stationary requirement is facilitated by packing the subject’s head into tight foam wedges before starting the scans.

Such conditions necessarily keep the subject pools from being completely randomized; their selection is biased by the needs of the machine. This may sound like a trivial matter, except when one beholds the sheer volume of fMRI papers that have been (and are currently being) published. Taken as a whole, we are not examining a randomized representation of a human family, but rather of a human family member who can stand to be in small spaces with his or her head in traction for a long time. There is no question that much information can be derived by imaging stress-tolerant individuals—but that is hardly the only population important out there.

Item to Keep in Mind #2

Resolution issues can be a problem.

Another issue has to do with the resolving power of the “typical” scanner, which affects the researcher’s ability to thoroughly characterize all relevant neural tissue recruited during an activity. The smallest block of tissue your garden-variety fMRI machine can image is a little cube a few millimeters on a side (these cubes are called voxels, a collision of the word volume and pixel). A few millimeters of brain tissue is a ridiculously large amount of cellular real estate, representing thousands upon thousands of neurons. But the machine can only examine a relatively large macroscopic block of tissue. That means there is a resolution problem—especially if you’d really like to see individual cells. It’s the equivalent of taking a picture of a 21st century battlefield using spy satellites from the early 1960s. As you know, many important neural activities do not happen in conspicuous, large blotches in the brain, but rather occur in subtle, more refined, electrically weaker networks distributed throughout the organ. fMRI technologies cannot currently capture these more subtle patterns. We are at constant risk for seeing only an incomplete picture of the brain’s response to the stimulus being examined.

The very idea of blotches themselves can be misleading, which leads to Item #3.

Item to Keep in Mind #3

Watch out for the edges.

Standard brain imaging scans often look like Doppler satellite images in a weather report: there are conspicuous regions of inactivity and sharply defined regions of activity. That sharpness can be misleading, however, because the activity levels between blotch and non-blotch regions are often quite small (in some cases, so small that the boundaries may be arbitrarily determined). One can easily be misled into believing that these sharply defined boundaries indicate just as sharply defined regions of activity, which may or may not be true.

This difficulty in determining the “actual” signal is confounded with the fact that the brain is hardly silent, even when no measured stimulus is occurring. The auditory cortex lights up after all in response to sound, which can be quite abundant in these machines, even when no experiment is in place. (There is also something called “dark energy” in the brain—electrical activity occurring throughout the organ in the complete absence of any stimulation.) Its a familiar signal-to-noise issue common to most engineering problems. Determining an acceptable threshold level, one capable of detecting the stimulus the researcher is after and not anything else, can surprisingly be difficult to achieve.

Item to Keep in Mind #4

Machine time and brain time may not be the same thing.

Temporal limitations of the technology must be taken into account when one is interpreting fMRI scans. The image most machines create develops quite slowly, usually over several seconds (that’s why it is important to keep the subject’s head so still). The big problem is that the brain’s neurons live in a world where firing rates often exceed hundreds of times per second. Combined with the fact that voxels are millimeters in size, exactly what is being imaged when a big blotch appears can be difficult to interpret precisely.

Getting accurate temporal data is further complicated by the fact that the machine is not actually measuring neural activity. It is measuring blood flow. There is usually about a 5-second delay between neuronal firing and observable blood-flow changes. The bottom line? Many fMRI images only display large groups of neurons whose cumulative firing efforts resulted in blood flow changes that are only observable long after triggering stimulus has exited. Assessing temporal activities can be very tricky business indeed!

Item to Keep in Mind #5

Every brain is wired differently.

Brain volumes can vary quite a bit from one person to the next, as can the absolute locations of specific brain structures. The wiring patterns of neural networks, which can include both structural and functional issues, also vary from one individual to the next. Because learning always involves changes in such patterns, and no 2 people learn the same things the same way, one can expect a wide variation in reactions in the brain to identical stimuli. All these issues must be taken into account when designing imaging experiments.

To overcome these individual variations, researchers usually recruit more than 1 person in the subject pool for imaging. And they take lots of images of each one of them. When the examinations are finished, the researchers line up all the images they receive, combine the data, and average what they see. At publication, the image obtained is usually representative not of one subject’s brain activity, but of the averaged brain activity of the entire experimental cohort.

Item to Keep in Mind #6

Be mindful of the dangers of reverse inference.

There is a great temptation to use activation profiles obtained with fMRI to infer a specific mental state. “This region of the brain is active, therefore this mental state must be occurring” is a mistake commonly made in the popular press and even occasionally by neuroscientists. It is usually called reverse inference, a convention that can be habit forming simply because in many cases, it actually reveals something useful. Broca and Wernicke speech centers really do light up when auditory information is being processed. Getting a stroke in those areas debilitates the function. Ergo, when these areas are active, speech is being processed.

Such reverse inferences may be fine for hyperspecialized regions like those of Broca and Wernicke, but there are many regions of the brain whose activation profiles are far more complex. The right prefrontal cortex lights up like a Christmas tree whenever the brain is trying to solve a hard problem. But it is also involved in impulse control, planning, foresight, and even the apprehension of mathematics. If the right prefrontal cortex lights up, is the brain working on solving a second order differential equation or reigning in the impulse to punch former math teachers in the nose? The amygdala, which is powerfully involved in mediating anxious emotions, is also involved in smelling popcorn—and in feeling sexually aroused. If it lights up, is its “owner” feeling afraid, hungry, or horny? When you have a series of choices but only one brain image, which mental state do you wish to infer?

There are many other objections that space here does not permit to describe. Yet, even considering just these 6, the cautious lessons about remembering contexts are very obvious here. One must be willing to constantly train a critical eye on the experimental conditions under which the images were obtained. One must also be willing to be skeptical about any claims concerning mental states. Any activation may narrow down the choices, but it does not a priori reveal our psychological interiors.

ENDING ON THE POSITIVE

There are ample reasons not to end this series on a negative note. Many careful neuroscientists are leapfrogging over the inherent limitations of the technology to obtain meaningful results. One of the most promising approaches uses fMRI in combination with other technologies. Some researchers, for example, are using transcranial magnetic stimulation experiments to temporarily ablate activity in previously stimulated regions of the brain and then looking for the presence or absence of observable behavior. Others are combining their imaging work in people with electrophysiological recordings in animals. A typical experiment might involve following up human data with monkey data, for example, using single neuron recordings in the animals to verify a given observation. Still others are using more sophisticated statistical models and co-opting analytical tools originally derived from research into machine-based learning. This allows researchers to shift the focus on trying to apprehend brain region–specific activation profiles to a specific task to answering more global questions about brain processing in the presence of the given stimulus.

I would like to end our entire series by briefly describing one such hopeful approach, illustrated in the accompanying Figure.

In a standard imaging experiment, researchers create an average of the fMRI activation profiles for adjacent voxels. This averaging makes things a bit easier to detect variations between experimental conditions alluded to previously (say you are having the subject first view the face of a famous movie star, then a boat). To do this averaging, you have to assume that neurons activated within specific voxels (all 10 gazillion of them) behave in an identical fashion. They don’t, most certainly. But there are statistical tools that, when used properly, can ferret out relevant information from these activation profiles and obtain meaningful results. Some of the most powerful of these tools are the so-called multivariate pattern classifiers. These classifiers produce finer-grained images by detecting activation patterns across many individual voxels without averaging any of them. As a result, they can detect signal differences that would normally escape conventional fMRI analysis.

The example in the Figure is an experiment that looked at how speakers process non-native language sounds.

CONCLUSIONS

The 3 articles in this series have attempted to summarize some of the basic science behind fMRI. Our journey took us from quark to voxel, and for a particular reason—to outline the strengths and limitations of noninvasive imaging technologies. One hopes that such wisdom will permit the clinician to train a critical eye on any image derived from these powerful magnets.

Despite these 6 caveats, I do not wish to leave you with the impression that fMRI images are not worth examining. I think fMRI, which can properly marry structure to function, represents one of the most powerful weapons in a cognitive neuroscientist’s arsenal. When used properly, fMRI takes on the powerful mantle of cartographer-in-chief—a valuable position in any expeditionary enterprise.

But just as there are limits to what a map can tell you about a country’s interior function, there are many limits to what these scans can tell us about a brain’s interior function. As long as we remember this, we will obtain a more nuanced opinion about fMRI images. This may help us get a clearer, ultimately more realistic view about how the brain actually processes information.

And it may also save a few airplane seats from being ripped apart at the seams when occupied by frustrated bioengineers.