Debating the Meaning of fMRI

A three-dimensional magnetic resonance image of a macaque monkey head. Inset: A schematic of the combination of cortical field maps of tactile stimulation obtained using fMRI (red and green squares) and electrophysiological recording techniques (cross-hatched regions). Functional magnetic resonance imaging (fMRI) experiments are, no doubt, incredibly intriguing: Researchers put volunteers inside a huge, harmless magnet that takes detailed pictures of the brain, expose those people to some sort of sensory stimulus, see what regions of the brain light up, and try to draw some conclusions about brain activity. Such experiments have become increasingly popular in neuroscience and psychology since their advent in the early 1990s. They are more commonplace for basic research studies than costly, time-consuming experiments done with positron emission tomography (PET). Magnetic resonance machines have been used in everything from psychiatry studies investigating how the brain reacts as food is digested to social psychology studies asking how the brain responds to people of different races. Yet scientists using fMRI still cannot say for certain to what extent they are actually measuring neuronal activity. Strictly speaking, fMRI machines measure the level of oxygen in the blood through a technique called blood oxygenation level-dependent (BOLD) contrast. The two different states of hemoglobin--oxygen-rich oxyhemoglobin and oxygen-poor deoxyhemoglobin--differ in their magnetic properties. The huge fMRI magnets are sensitive to changes in the concentration of deoxyhemoglobin. As neural activity increases, blood flow to the vasculature of the brain presumably increases, altering this concentration. But how closely coupled are the blood flow changes detected by fMRI and the spiking activity of neurons? Scientists know that the fMRI signal is somehow related to neural activity, but have yet to show a precise relation of the signal and the long-time accepted "gold standard" of neuronal activity, namely the firing rate of individual neurons. "We need to be able to bridge the gap between the fMRI measurements we get in the human brain and the more conventional electrophysiological measures of spiking activity that physiologists have been recording for 100 years," says David Heeger, an associate professor of psychology and neuroscience at Stanford University and an fMRI specialist. "If that relationship weren't very tight, then some results, maybe many results, would have to be reinterpreted." For years, several labs have tried to make the connection, yet definitive answers remain elusive. "Everybody doing fMRI is hoping that what they're doing is a reflection of neuronal activity," says Bennett Shaywitz, a professor of pediatrics and neurology at Yale University who's using fMRI to study the neural mechanisms of dyslexia. "But we all know that what we're essentially measuring now is blood flow." Because recording directly from individual neurons in living humans is generally not an option, investigators have turned primarily to the macaque monkey for electrophysiological insights and fMRI comparisons. A recent study in macaque monkeys actually suggests that the 1.5 Tesla fMRI machine, the typical power level for most fMRI studies (much more powerful machines exist but are not commonplace), is not quite as reliable as is commonly thought.1 When researchers compared cortical maps based on multiple fMRI images of macaques with maps based on electrophysiological recordings of those same macaques, they found that the fMRI signal related to the actual neural signal only 55 percent of the time. The results, however, are not quite as dire as they appear. Stronger magnets, ample subjects, and ample scans would likely help solve the problem. More importantly, notes the published study's senior author Leah A. Krubitzer, an associate professor of psychology at the University of California, Davis, the plane of the greatest variability is relatively easy to predict. The region of greatest variance was consistently in the plane perpendicular to the major blood vessels that supply the region. Theoretically, imagers could identify and weed out the unreliable data. Krubitzer emphasizes that her lab has much invested in the abilities of fMRI and that she was disappointed to see the relatively low level of congruence, but she was not looking to dethrone fMRI as one of the greatest neuroscience developments of the 1990s, the so-called "decade of the brain." Nevertheless, some scientists were skeptical. "We had a hard time getting that paper published," says Krubitzer. "One reviewer would say 'this is a heroic paper' and 'this is wonderful,'" and another would be really, really angry. She contends that doubters may have "put all their eggs in that fMRI basket" and did not want to hear about drawbacks. Another recent study, taking a different approach, offers more encouraging results. California Institute of Technology (Caltech) researchers set out to relate animal neuron recordings with fMRI-monitored brain activity. They found that the fMRI signal from the human motion area, V5 (or MT), changed in a way that was directly proportional to activity of single neurons previously recorded in a similar area of the macaque brain.2 All subjects were responding to visual moving stimuli. The study does not provide any insight into what links neuronal spiking to BOLD, but simply shows that the fMRI signal is proportional to the mean firing rate of macaque neurons. "I think the significance of this for the wider community is that people can feel fairly confident that their fMRI data is probably reflecting spiking rates," comments lead author Geraint Rees, a Wellcome Advanced Fellow at Caltech, referring to researchers studying the visual system in particular. He adds a few caveats, however, saying that the study's conclusions should not necessarily be applied to all combinations of stimulus, experiment, scanner, and animal. Heeger analyzes the study as well in an accompanying piece to this published research and supports it by showing similar results from a reinterpretation of a study done by his group in which they compared macaque single neuron recordings to human fMRI signals.3 New approaches in monkeys and humans should help further reveal fMRI's secrets. Directly studying individual human neurons in living people has proven very difficult--but not impossible. Christof Koch, the senior author of the V5 study and a professor of computation and neural systems at Caltech, participated in a more recent study that employed an old, though rarely used, technique to get around the ethical and logistical problems of doing so.4 They made their recordings in an special subset of patients already slated to receive treatment for their epileptic seizures. No longer treatable with drugs, the patients had to have a small part of their brain removed to prevent further seizures. In patients in whom the targeted brain area is not easily identified, neurosurgeons insert 10 to 12 electrodes into the brain and wait for a seizure to occur before zeroing in on the desired region. Researchers took this opportunity to record from individual neurons. According to Koch, the data corroborated data from fMRI scans to some extent. They found that certain neurons in the hippocampus (a center for seizure activity) are very selective for individual categories, familiar faces for example. Similar observations have been made in many fMRI studies that have attempted to relate face recognition to brain activity. The research that should provide the biggest step forward is that which considers fMRI measurements and individual neuron recordings simultaneously in the same species. Nikos K. Logothetis, director of the physiology of cognitive processes group at the Max Plank Institute in Tuebingen, Germany, is working on doing precisely that: using a specially designed magnet, in both awake and anesthetized macaques. Unlike the V5 study, the Logothetis research will not rely on statistics to compare two disparate data sets but will find the exact relationship between the two signals. Meanwhile, according to Peter Bandettini, a National Institutes of Health investigator and fMRI specialist who plans on doing similar experiments, properly designing of an fMRI paradigm, and asking reasonable questions based on that paradigm, remains the key for any good fMRI experiment. Says Heeger, referring to fMRI experiments, "Whether there's a problem with the way people have done things in the past, I really don't think that's so important." He adds that the real benefit of building a bridge from electrophysiological underpinnings to fMRI images is the prospect of eventually enabling researchers to "work seamlessly" among fields including neuroscience, cognitive psychology, experimental psychology, molecular biology, and cellular biology. Eugene Russo can be contacted at erusso@the-scientist.com References 1. E.A. Disbrow et al., "Functional MRI at 1.5 tesla: A comparison of the blood oxygenation level-dependent signal and electrophysiology," Proceedings of the National Academy of Sciences, 97:9718-23 Aug. 15, 2000. 2. G. Rees et al., "A direct quantitative relationship between the functional properties of human and macaque V5," Nature Neuroscience, 3:716-23, July 2000. 3. D.J. Heeger et al., "Spikes versus BOLD: what does neuroimaging tell us about neuronal activity?" Nature Neuroscience, 3:631-3, July 2000. 4. G. Kreiman et al., "Category-specific visual responses of single neurons in the human medial temporal lobe," Nature Neuroscience, 3:946-53, September 2000.