fMRI: The Perfect Imperfect Instrument

Courtesy of Chloe Hutton, Functional Imaging Lab THICK OR THIN? The cerebral cortex thickness metric can be used to study the progression of diseases such as Alzheimer disease, epilepsy, mental retardation, and schizophrenia, and to investigate how brain function is affected in the abnormal regions. According to legend, functional neuroimaging can trace its roots to the stroke of noon on a day in the late 19th century, when Italian physiologist Angelo Mosso observed a sudden increase in brain pulsation in his test subject, Bertino the peasant. Using an elaborate contraption, Mosso had been measuring the pulsations coming from a soft spot in Bertino's skull, the result of a head injury. Intrigued by the sudden pulsing, Mosso asked Bertino if the chiming of the local church bell had reminded him of his forgotten midday prayers. When Bertino said yes, his brain pulsated again. Then Mosso asked Bertino to multiply 8 by 12. Again, Bertino's brain pulsated. Thus was borne the notion that blood flow in the brain is related to cognition. That notion remains the basis of what some neuroscientists consider the most powerful in vivo brain imaging technology in use today, functional magnetic resonance imaging (fMRI). Since its emergence in 1992, fMRI has dominated the field of human brain mapping and has been featured in thousands of published papers--nearly an order of magnitude more than for any other functional imaging technique. For noninvasive whole-brain coverage in humans, fMRI's spatial resolution cannot be beat. And, unlike its historical predecessor, positron emission tomography, fMRI does not depend on dangerous ionizing radiation. Martin Lauritzen, professor of clinical neurophysiology at Denmark's Glostrup Hospital and the University of Copenhagen, says, "In a way, I think the fMRI is a perfect machine." But there are things that even a perfect instrument cannot do, such as record neuronal events in real time. More profoundly, the technique is limited by interpretational challenges; handling the enormous amount of data that even a single scan produces requires statistical skill, both to uncover deeper truths and to prevent erroneous conclusions. ©Functional Imaging Laboratory FINGER TRIGGER? In this experiment, subjects were imaged while performing finger movements. Before the images were taken, part of the brain controlling movement of the hand was stimulated repetitively. The figure shows that there is an increase in brain activity at the site of stimulation (primary motor cortex) and in other areas of the brain that control movement (premotor cortex). The graphs show the relative size of the changes in activity during movement (M) and rest (R) after real stimulation (red), as compared to pretend (blue). Much work, therefore, lies ahead. Karl Friston, director of the Functional Imaging Laboratory at University College London, predicts that the next chapter in fMRI's history will see an exciting role for computational neuroscience and the parallel development of sophisticated statistical techniques for understanding the vast information the technique produces. Some scientists are looking beyond fMRI to find the next great imaging technique, while others debate exactly what it is that fMRI measures (see Hot Papers | Caution: Brain Working). Meanwhile, says Nikos Logothetis, professor of neuroscience and director of the Max Planck Institute for Biological Cybernetics in Tübingen, Germany, fMRI is the best game in town. "Right now, nothing even comes close to the sophistication and quality of the fMRI machine," he says. "It's a pretty exciting technology for both clinical applications and research." TIME TRIALS Functional MRI produces colorful computer images that show where neuronal activity is occurring in the brain when a person is asked to perform a certain task, such as remembering directions or identifying an object. The images are derived from a magnetic recording of changes in the local concentration of deoxygenated hemoglobin that occur when neurons are activated. Work by Logothetis and others has alleviated much of the initial concern that fMRI signals may not accurately reflect what really happens in the brain.1 Still, debate on this topic lingers. These blood-based changes occur in one to two seconds, which is far longer than the few hundred milliseconds required to convert a sound into meaning. This is not a hardware problem--the scanners can take pictures every 50 to 60 milliseconds--but rather a biological one. "In terms of fMRI as we know it, it probably doesn't need to be improved that much," says Friston. However, by relying on fMRI alone, he says, "The concern is that you are missing out on a rich world of fast temporal dynamics." Poor temporal resolution of fMRI precludes segregating the different stages of the fast conversion from word to meaning, which is why language skill expert Peter Hagoort must use fMRI in combination with electroencephalography (EEG) and magnetoencephalography (MEG) in his efforts to identify those stages. Many neuroscientists dream of a "more perfect" instrument, one that will combine the spatial sensitivity of fMRI with the millisecond temporal acuity of EEG or MEG. Hagoort, who directs the FC Donders Centre for Cognitive Neuroimaging in the Netherlands, says it is difficult to predict what such an already rapidly changing technology will look like in 10 or 20 years. ©Functional Imaging Laboratory MMMMM, ICE CREAM! The brain activity shown here occurred when subjects were viewing a picture that reliably predicted a vanilla smell. The circled responses are in orbitofrontal cortex (OFC), which is a brain area known to be involved in emotional processing and learning. In the meantime, Arthur Toga, director of University of California, Los Angeles, Laboratory of Neuroimaging, points to optical imaging, which he says "offers you tremendous spatial and temporal resolution and [Toga's emphasis] the ability to actually examine the physiology, which fMRI only infers." But optical imaging, which measures changes in cortical light reflectance, is limited by its invasiveness (it can be done only during surgery) and shallow penetration into the brain. Despite its temporal limitations, there is still a wealth of undiscovered neuronal function at temporal and spatial scales that fMRI can measure, Friston says. So even as research teams worldwide look for alternatives or ways to combine fMRI with EEG or MEG, fMRI will likely remain the superpower of functional neuroimaging for years to come. "We are not at the end of [fMRI's] capacities," Logothetis says. "In fact, we are at the beginning." MORE THAN A PRETTY PICTURE Functional MRI generates maps that are statistical pictures of brain activity, with the colors changing from red to yellow as neuronal activity increases. As Friston explains, in most fMRI studies, changes in brain state are characterized statistically one voxel at a time. (A voxel is a region, usually measuring 3 x 3 x 5 mm.) But statistics being what they are, some false readings are bound to arise. Statisticians typically deal with 95% confidence intervals, in trying to define the probability that a given outcome will occur by chance. With 100,000 or more voxels being compared, one by one, in a typical fMRI study, "just by chance, even if no part of the brain responded, you'd expect 5% of the brain to show activation," says Friston. To protect against these false positives, a statistical framework known as random field theory was introduced into fMRI methodology in the early 1990s, but it's unclear whether the problem has been resolved. Mark Cohen of the University of California, Los Angeles, says false positives are "endemic and uncontrolled in virtually all fMRI analyses." Moreover, even if false positives were not an issue, this so-called "mass univariate" approach is limited by its assumption that voxels are insular units and that different parts of the brain do not interact during neuronal processing. Hypothetically speaking, if voxel A is analyzed independently of voxel B but B affects the responses in A, then a univariate statistical characterization will not accurately reflect everything about the parts of the brain these voxels represent and how their responses are integrated. The notion that brain functions are compartmentalized in a noninterconnected brain has historical roots in the field of phrenology, which predate even Mosso and Bertino. According to David Cox of the Massachusetts Institute of Technology, however, "It's an open question whether the cortex is really as modular as many fMRI studies make it out to be." Cox and Robert Savoy, Massachusetts General Hospital, recently described a new fMRI application of a statistical methodology aimed at identifying, not ignoring, the interactions among voxels from different parts of the brain.2 Courtesy of Chloe Hutton, Functional Imaging Lab GRIPPING DATA: This figure shows activity in the brain during repetitive gripping with the right hand. Each brain represents the activation pattern at different time points over the first six weeks after stroke for one patient. Recovery of function is associated with diminishing brain activation, due to increasingly efficient neural circuitry. This is very similar to what is seen during learning of a new complex motor task in the undamaged human brain In what Cox describes as a "somewhat flashy mind- reading" fMRI experiment, trained observers first scanned volunteer subjects for several sessions while the volunteers looked at certain objects such as baskets, garden gnomes, and teapots. Cox and Savoy used a multivariate statistical approach--one designed to detect interactions among voxels--to identify brain patterns that corresponded to each object. When the volunteers' brains were scanned again later, the observers used the patterns, which covered a large region of the cortex, to guess which objects the volunteers were looking at. Their guesses were up to 85% to 90% accurate. The study highlights the significance of coarse-grained neuronal patterns that extend over large areas, patterns that Cox and Savoy would not have found if they had relied on conventional fMRI statistical methodology. "It's a trickier way to do things," says Cox, "but in the end, it's more powerful." Even though multivariate approaches comprise less than 5% of the thousands of fMRI publications, Friston says they "probably represent the area in which most advances will be made in the future." The extent to which different parts of the brain interact during neuronal processing remains a major unsolved problem in neuroscience. Still, it is unlikely that the multivariate approach will displace the conventional mass univariate approach and its accompanying assumption that cognitive functions are localized in the brain. Friston likens the latter to making an aerial map of cities and the former to identifying the transportation routes that connect the cities: Do the people move by road, rail, or air? "You need to know where A and B are before you can see how they interact. These are not competing approaches," he says. "They are complementary." While Friston hopes that more sophisticated statistical approaches will lead to more accurate maps of neuronal activity, Cohen is worried that focusing too much on statistics might occur at the cost of a better understanding of the underlying physiology and cognitive science of fMRI signaling. After the notion was first proposed more than a hundred years ago, neuroscientists are finally convinced that fMRI blood-based signals are associated with synaptic processing. Says Lauritzen: "Now the question is, what does that mean?" Leslie Pray (lpray@nasw.org) is a freelance writer in Northampton, Mass. References 1. E. Russo, "New magnet clarifies fMRI signals," The Scientist, 15[16]:27, Aug. 20, 2001">The Scientist, 15[16]:27, Aug. 20, 2001. 2. D.D. Cox, R.L. Savoy, "Functional magnetic resonance imaging (fMRI) 'brain reading': Detecting and classifying distributed patterns of fMRI activity in human visual cortex," Neuroimage, 19:261-70, 2003. THE MEETING OF MIND AND MOLECULE Functional neuroimaging is not just for brain cartographers. It also plays a vital role in the clinical arena. Noninvasive imaging techniques such as fMRI have dramatically altered attitudes toward mental illness by providing a physical context--the brain, as opposed to the ethos--for understanding schizophrenia, depression, and Alzheimer disease. Now, combined with the power of genomics, the future of fMRI promises even more. Daniel Weinberger of the National Institute of Mental Health in Bethesda, Md., uses the term "imaging genomics" to describe the combined use of the two approaches. By matching fMRI signals to genes with known impacts on how the brain processes information, researchers are hoping to find innovative pharmacological ways to disrupt the molecular cascades that contribute to mental illness and neurodegenerative disease. Researchers traditionally use personality questionnaires and similarly subjective tests to look for associations between genetic variation and human behaviors. By using fMRI (or EEG or MEG) instead, they get an arguably more objective view of what the brain is doing when people process cognitive or emotional information. Further, by associating fMRI brain responses with candidate genes and the molecular pathways they encode, researchers can identify sites for potential pharmacological intervention. "Any of these genes could give you an entry point into the cell processes, and without an entry point you have no way of modifying their effects," says Weinberger. Using fMRI, Weinberger and his colleagues have performed three genomics proof-of-principle imaging experiments. Most recently, they measured the neuronal activity of a set of volunteers performing a simple declarative memory task (e.g., remembering simple pictures that had been shown to them previously) and tested its association with a BDNF (brain-derived neurotrophic factor) polymorphism (Val66Met) that has a known effect on memory-related activity.1 They found that 25% of the variation in peoples' memory is predicted by variation in the BDNF gene. The results suggest that BDNF is a key player in human declarative memory and, as the study's authors conclude in their paper, "BDNF signaling may be a propitious target for interventions to enhance declarative memory." These proof-of-principle experiments raise the real possibility that imaging genomics can improve the diagnosis, prognosis, and treatment of mental illness and neurodegenerative diseases. In terms of looking for a physical basis of mental illnesses, says UCLA's Cohen, "Imaging genomics may provide a much needed window to guide future interventions in these cases." 1. A.R. Hariri et al., "Brain-derived neurotrophic factor Val66Met polymorphism affects human memory-related hippocampal activity and predicts memory performance," J Neurosci, 23:6690-4, July 30, 2003. function sendData() { document.frm.pathName.value = location.pathname; result = false if (document.frm.score[0].checked) result = true; if (document.frm.score[1].checked) result = true; if (document.frm.score[2].checked) result = true; if (document.frm.score[3].checked) result = true; if (document.frm.score[4].checked) result = true; if (!result) alert("Please select") return result; } .myradio{background-color : #94bee5} Please indicate on a 1 - 5 scale how strongly you would recommend this article to your colleagues? Not recommended 1 2 3 4 5 Highly recommended Please register your vote

fMRI: The Perfect Imperfect Instrument

Interested in reading more?

Become a Member of