Hard Science or "Technicolor Phrenology"?
The Controversy over fMRI

by David Dobbs

from Scientific American Mind, April 2005

______________________________________________

Functional magnetic resonance imaging — or fMRI— has made a bright splash since its development in the early 1990s. Operating at spatial and time scales far finer than previous scanning techniques, it has sparked great excitement for letting us finally “watch the brain at work.” Tens of thousands of fMRI studies have explored everything from the nature of Alzheimer’s to differences in brain activation between adolescents and adults, schizophrenic and normal minds, the impulsive and the methodical, the empathetic and the impassive. They have studied face, object, and word recognition, movements simple and complex, working memory and false memories; they have looked at people anticipating pain, mothers recognizing their children, people pondering ethical dilemmas, and people lying; they have even examined why many people buy Coke even though they really prefer the taste of Pepsi. Psychologists have welcomed fMRI for finally making their soft science hard. And cognitive neuroscientists have used it heavily in the recent explosion of understanding about the brain.

Increasingly, however, the fMRI boom is stirring argument over the reliability of its findings. This debate, at once technical and philosophical, regards both fMRI’s accuracy (mainly because it measures neuronal activity indirectly, by detecting associated increases in blood flow) and the legitimacy of its main use, to link complex mental functions to particular brain areas. Critics also feel that fMRI overlooks the “networked” or “distributed” nature of the brain’s workings, in which, a near-consensus holds, most mental tasks involve several brain areas working together.

“This is a very gross technique,” says critic Steven Faux (pronounced fox), who heads the psychology department at Drake University. “It’s like a blurry photo – better than no photo, but still blurry, with real limitations that are too often overlooked. It’s very easy to overextend this technology.”

Many fMRI practitioners sometimes sound bewildered that this powerful new tool has raised such difficult questions. It’s as if the development of a better telescope provoked argument about whether stars and black holes exist and whether constellations are illusory or real. “It is a huge surprise to me how big this issue has become,” says Marcus Raichle, a Washington University neurologist who has worked in brain scanning for over two decades. “Just a huge surprise.”

A Vague Precision

Brain imaging technique began with an early-20th-century method called pneumoencephalography, a dangerous procedure in which the skull’s cerebrospinal fluid was replaced with air to show the brain more clearly on X-ray. The angiograph, developed in the 1920s, improved on that by using dyes injected into the bloodstream. (Angiography is still used to manage blood-vessel defects and some tumors.) These early methods showed only static structure rather than function. Likewise, the CAT or CT scans (computerized axial tomography) developed in the 1970s, which also exploit X-ray technology, take static pictures, though of far greater detail.

The 1970s also brought the first functional imaging technology – that is , imaging designed to see the brain at work rather than merely its structure – in the form of PET, or positron emission tomography. Like fMRI, PET measures the increased blood flow associated with neuronal activity. In PET’s case, a nuclear scanning device tracks lightly radioactive elements (positrons) injected into or inhaled by the subject. But PET is not only invasive but slow, requiring the better part of an hour for a scan, and it provides a rather broad “temporal resolution” of 60 seconds (meaning it takes that long to measure the blood flow to an area of the brain) and a spatial resolution of about 6 to 9 mm.

The fMRI, in contrast, takes whole-brain scans in less than two seconds and creates images at resolutions of about 3 cubic mm – or, more commonly, in “voxels” (a word messily merging volume and pixel) about 2 mm square and 4-5 mm long, about the size of a grain of rice. FMRI also require no injections, allowing more extensive scanning. In a typical fMRI study, a subject lies in an MRI scanner and is scanned, first at rest with eyes closed, to provide a baseline reading, and then while performing some mental task – identifying faces, threading a computerized maze, engaging in a role-playing game, answering an ethical problem — while the scanner takes multiple images. In the most common fMRI technique, called BOLD (for blood oxygen-level dependent) fMRI, the MRI machine measures increases in blood flow by spotting a change in magnetism that occurs when a blood surge raises the ratio of fresh, oxygenated hemoglobin to “used,” deoxygenated hemoglobin, which has a significantly different charge. (See figure XX, [rough art is in illustration titled BOLDillo]). The areas receiving these surges show as brighter colors on the fMRI images, red changing to yellow as flow rises. Doubts about whether these increases correspond to actual neuronal activity have been answered by several studies tying blood flow directly to neuronal activity, including recent animal studies that used probes to match the firing of individual neurons to heightened flow seen in fMRI.


Yet the link between neuronal activity and fMRI’s blood-flow measurement, however real, is decidedly rough. Abigail Baird, a Dartmouth psychologist who uses fMRIs to study brain changes during adolescence, puts it succintly: “Hemodynamic response is a sloppy thing.” For starters, neuronal action takes milliseconds, while the blood surge follows by 2 to 6 seconds; a detected increase in blood flow therefore might be “feeding” more than one operation. In addition, because each voxel encompasses thousands of neurons, thousands or even millions may have to fire to significantly “light up” a region; it’s as if a whole section of a stadium had to shout to be heard. Meanwhile, it’s possible that in some cases a small group of neurons drawing little blood may perform functions as crucial as a larger group elsewhere but would either go undetected or show as minor activity. Likewise, some neurons might operate more efficiently than others, drawing less blood. All these factors could lead the fMRI images to misrepresent actual neurodynamics.

Processing the scan’s gigabytes of raw data so that they become images introduces other caveats and variables. Researchers must choose among and adjust many different algorithms to extract an accurate image, adjusting along the way for variations in skull and brain configuration, movement of subjects in the scanner, “noise” in the data, and so on. This “chain of inferences,” as a recent Nature article called it, offers much opportunity for error.

Finally, most fMRI studies – probably 95% — use “univariate” processing, which critics say shortchanges the distributed nature of neurodynamics. The charges rise because univariate (literally meaning “one variable”) algorithms consider the data coming in from each voxel during a scan as one sum, which makes it impossible to know how the activity in a particular voxel accrued (all at once, for instance, or in several pulses) or how it related sequentially with activity in other voxels. Univariate processing does sees all the parts working – thus the multiple areas lit up in most images — but not in a way that shows how one area follows or responds to another. This makes viewing an fMRI image something like “listening” to a string quartet by hearing, condensed into a single noise after the music has ended, only the total amount of sound each instrument produced during the piece rather than hearing how the players pass melodic lines back and forth and accompany and respond to one another. Statistical methods known as multivariate analysis can break down each voxel’s activity and analyze the interchanges among brain areas, but the complexity of those analyses has so far limited their use.

Obvious and Not So Obvious

For some, these vagaries and limitations make fMRI too rough an instrument for the more ambitious work it’s being used for. “The beautiful graphics fMRI produces imply much more precision than there actually is,” says Drake University’s Faux. “But it’s really a very gross, if not vague, physiological measurement that people are using to try to pin some very complex behaviors. And in too many studies the authors way overinterpret the data. None of that advances the science.”

Faux is too polite to name such studies. But surveying the literature finds them readily. Some are trivial -- a study showing that men’s amygdalas (which play a key role in generating emotion) light up when they view Ferraris. Some, as Faux says, recklessly overinterpret: A study of Democrats and Republicans watching Kerry and Bush videos concluded that heightened activity in the subjects’ emotion-sensitive amygdalas when they viewed the opposing candidate “suggest[ed] the volunteers were actively trying to dislike the opposition.” Yet other studies suffer serious design failures, as did several dozen that claimed to find physiological markers of ADHD-diagnosed kids — but failed to control for the effects of their subject’s Ritalin use.

Such studies, however, don’t prove any fatal flaw in fMRI, says Dartmouth’s Baird, but instead highlight the importance of using careful technique, solid study design, and judicious interpretation. Baird, who likes to check her fMRI studies against similar studies using other methods, likens fMRI interpretation to analyzing skid marks at accident scenes. “Someone who’s done it often, who is careful, and who collects a lot of other evidence will probably draw useful conclusions. Someone who’s inexperienced or who doesn’t check the whole scene will probably read them poorly.”

Even serious, well-crafted studies can be undermined by design failures quite subtle. In a widely cited and publicized study of adolescent emotional responsiveness, for instance, Dr. Yurgelun-Todd of McLean Hospital scanned adolescents as they viewed and characterized the expressions on a series of black-and-white images of fear-struck, middle-aged faces. Compared to adults, adolescents showed less activity in the frontal lobes, where much cognition occurs, and more in the amygdala. The adolescents also scored poorly in characterizing the expressions. This suggested, Yurgelun-Todd told PBS's Frontline, that "the teenager's brain may be responding with more of a gut reaction than an executive or thinking kind of response." But in a follow-up, Baird ran a similar experiment using color photos of adolescent faces and found the adolescent subjects responded and scored much like adults. “They were simply more engaged by more contemporary peer photos in color,” says Baird. “They did well if they cared."

This tale highlights some of fMRI’s most vexing nontechnical difficulties: the danger and ease with which a design flaw can corrupt results; the imagery’s power to sway professionals, press, and public despite those flaws; and the way results can reinforce conventional ideas, such as those regarding teen thinking and behavior. This last problem animates some of fMRI’s most serious critiques. Some critics, including Faux and psychologist William Uttal, argue that many of the “cognitive functions” under study in fMRI work are so abstract and vague that they denote little more than a conceptual nervous system. Faux’s leading bugbear is the central executive measured in so many studies. “That’s a real favorite,” he says, “to measure the ‘central executive.’ Now — what is that?”

Defenders have an answer: they say the central executive is a supervisory function residing in a network of areas in the prefrontal cortex and anterior cingulate cortex (a small area tucked into the space between the two frontal lobes) that monitor other brain functions and regulate priority-setting and decision-making. Besides pointing to imaging studies that show these areas activating in various combinations when the mind sets priorities, encounters new situations, and makes decisions, they can cite a well-recorded history of patients who have lost the ability to do those things (especially making decisions and setting priorities) after suffering frontal lobe damage.

If this seems slippery ground – well, it is, by nature. Faux and Uttal are protesting the arbitrary nature of terms that must of necessity be abstract, and the argument inevitably leads to judgment calls about the reality of an unseen thing. As fMRI defenders note, everyone considers it a necessary reach when physicists and astronomers study unseen structures that are largely inferred but that are useful because they explain behavior (in those cases, atoms and galaxies). “You can’t see or measure subatomic particles directly,” says John Darrell Van Horn, who heads the fMRI Data Sharing Project at Dartmouth University. “But they’re useful, well-supported models we can refine based on experiment. I think many of these functions are quite similar.” Yet as Van Horn notes, the “central executive” pushes the limit for many, including him; he considers it more metaphor than model. Only more evidence is likely to resolve some of these fuzzier issues of nomenclature.

A Wider View

It’s not happenstance that fMRI controversies concern matters both conceptual and tangible. This duality lies inherent in the attempt to connect ephemeral mind to the corporeal brain. It particularly infects the criticism that fMRIs steer cognitive neuroscience into the old error of “localizing” brain function by tying specific mental tasks and processes to particular regions.

Virtually no researchers, of course, seriously believe that brain functions are discretely localized. As Raichle says, “No rational person would suggest there’s a single ‘emotion’ spot, for instance.” Yet the localization charge carries some weight partly because, all rhetoric notwithstanding, most fMRI studies have indeed focused on how particular mental processes activate specific areas. This has provoked the biting accusation that fMRI studies comprise “the new phrenology,” as Uttal’s book on the subject is titled.

But the phrenology charge ultimately overstates the extent to which today’s cognitive neuroscientists view the brain’s modularity. Most fMRI work seeks not to “localize” brain function but to map the parts of a system that act in different combinations for different tasks. While this may suggest a localization mindset, it mainly indicates the early, first-survey nature of the mapping being done; it’s only natural to make a simple map before you make a complex one and to place cities before roads. And things are progressing. Even compared to three years ago, fMRI studies today more often identify and discuss relations between several activated areas. No neuroscientist views the brain as some office headquarters walled off into incommunicative departments. “It’s more like an orchestra,” says Marcus Raichle, with the different sections playing at times, volumes, and timbres depending on the effect needed, interacting in endless combinations to create an infinite variety of music.

What Next?

Yet it remains that today’s fMRI technology misses much of this music. To hear it more fully, it must advance.

Potential for such progress lies in further developing the multivariate algorithms that can track interactions among regions. The NIH’s James Haxby, MIT’s David Cox and Mona Spiridon, Columbia’s Christian Habeck, and others have successfully used multivariate processing to reveal interactions between brain areas during scan studies. Cox’s study, in fact, found that volunteers looking at different objects produced patterns so distinctive that an observer could quickly learn to examine a series of scanned images and correctly guess what the subject had been viewing. Expanding and refining such multivariate protocols should let fMRI reveal far more about how the brain’s regions work together.

Will such improvements end the controversies about fMRI and other brain imaging? Perhaps partly. More standardized processing protocols and peer pressure should reduce methodological blunders, and advances in fMRI will almost certainly answer most of today’s technical concerns. Researchers are already working on combining fMRI’s spatial acuity with the tighter temporal resolution of electroencepholography (or EEG) and magnetoencephalography (MEG), which measure neuronal activity by detecting, respectively, the minute electrical and magnetic activity they produce. Such innovations, or others not yet foreseen, should someday measure neural activity with more spatial and temporal precision.

Yet such advances seem unlikely to resolve the philosophical anxiety that brain imaging provokes. The attempt to identify the “neural correlates of consciousness,” as one paper puts it, rouses the long insistence, first fully articulated by Rene Descartes, that our minds are more than our brains. To put it another way, we resist the notion, as novelist Jonathan Franzen more recently put it when contemplating his father’s Alzheimer’s, of “the mind as meat.” Neuroscientist Antonio Damasio, who calls this resistance “Descartes’ Error,” argues we will eventually tie the complexities of thought and emotion to specific neural operations without any sense of loss. Yet most people feel discomfort at having our ideas and feelings — what seem to be our characters and identities — reduced to pixelated images of tangible operations. As technology makes it easier to answer scientific questions about our brains, this metaphysical unease may only grow.