Functional Magnetic Resonance Imaging

Scott A. Huettel

Functional magnetic resonance imaging (fMRI) is a method for detecting the functional activity of different areas of the brain. The use of fMRI has grown dramatically over the past 25 years and is now the dominant research method in cognitive neuroscience. Each year, several thousand articles are published that use fMRI to study topics as basic as the structure of memory and as complex as the foundations of moral cognition. What accounts for this remarkable growth? What information does fMRI provide, and how does that information lead to inferences about brain function? What new directions will fMRI researchers pursue in the coming years, as scientists make new discoveries both about the technique itself and about the functional organization of the brain? Box Extension 15.2 describes the development of the fMRI technique and its current and potential future uses.

fMRI: From Physiology to Physics

The development of fMRI followed a complex and winding path. It had long been recognized that arterial (oxygenated) and venous (deoxygenated) blood had different magnetic properties; in particular, venous blood, but not arterial blood, is paramagnetic, which means it can distort a surrounding magnetic field. It was not until the late 1980s, however, that Seiji Ogawa, a researcher at Bell Laboratories, demonstrated that changes in blood oxygenation could be visualized using MRI. Ogawa manipulated the oxygen content of the air breathed by rats—from fully to minimally oxygenated—and found that the blood vessels of the fully oxygenated rats were distinctly visible on a particular type of magnetic resonance image. On such images, the brightness depends on the homogeneity of the local magnetic field; if a compound such as deoxygenated blood distorts the magnetic field in one location, that part of the image will get slightly darker. Recognizing that changes in local neuronal activity might affect the oxygen content of nearby blood vessels, Ogawa speculated that future researchers might use this method to create images of the functioning brain.

Almost immediately, other researchers began exploring ways of translating Ogawa’s discovery—now called blood-oxygenation-level-dependent (BOLD) contrast—into practice. Rather than exogenously manipulating what a human breathed while in the scanner, these researchers took advantage of a fortuitous aspect of neurophysiology: There are endogenously induced changes in the BOLD signal. After neurons are active, they require two metabolites—oxygen and glucose—to provide the energy for restoration of their membrane potentials. These metabolites are not stored locally but are delivered through the vascular system, which sends more blood to regions of the brain that have been more active. Because the blood supplies more oxygen than can be used by the neurons, the net result is a systematic increase in blood oxygenation in more active brain regions, which in turn leads to an increase in these regions' BOLD signals (Figure A). Because most of the brain is at least somewhat active most of the time, many fMRI studies rely on a subtractive analysis of the BOLD signal: A control signal is subtracted from the experimental signal to reveal a net increase in activity in certain brain regions that is specific to the experimental conditions.

Figure A Summary of BOLD signal generation (1) Under normal conditions, oxygenated hemoglobin (Hb) is converted to deoxygenated hemoglobin at a constant rate within the capillary bed. (2) When neurons become active, however, the vascular system supplies more oxygenated hemoglobin than needed by the neurons, through an overcompensatory increase in blood flow. This results in a decrease in the amount of deoxygenated hemoglobin, leading to a brighter magnetic resonance image. (From Huettel, Song, and McCarthy 2009; after Moseley and Glover 1995.)

Mapping brain function

The most remarkable aspect of fMRI was not its ability to create images of the brain. After all, a variety of methods—from computerized tomography to standard MRI—contributed to clinical practice through the creation of images of brain structure. Nor was fMRI unique in its ability to study brain function. Neuroscientists already used single-unit electrophysiological methods to record individual action potentials, scalp electrodes to assess the integrative activity of populations of cells, and positron emission tomography (PET) to map function via the concentrations of radioactively labeled isotopes in the brain. In contrast to these other methods, however, fMRI was remarkable for its accessibility. Its data were not collected from animal models or via invasive injections into the brain, but from human volunteers subjected to noninvasive procedures. The required hardware was readily available, and its outputs were vibrantly colored three-dimensional maps of the human brain (see Figure 15.7 in the book). This new technique also filled a niche within the overall set of techniques available to the neuroscientist (Figure B). It combined passable temporal resolution (on the order of a second) with good spatial resolution (on the order of millimeters), while allowing coverage of the entire brain. In essence, fMRI provides a coarse but comprehensive view of metabolic changes throughout the human brain—and its technology is accessible enough that undergraduate students can use it for class projects.

Figure B Neuroscience techniques differ in their spatial resolution, temporal resolution, and invasiveness The vertical axes illustrate spatial resolution in terms of distance (left) and the corresponding brain structures (right). The horizontal axis illustrates temporal resolution. The relative invasiveness (i.e., how much potential damage or risk the procedure involves) is indicated in height; relatively noninvasive techniques are shown as coming out of the page, while invasive techniques are recessed. fMRI provides a good balance of spatial and temporal resolution and thus is appropriate for a wide range of experimental questions. However, other approaches, including electrophysiology, lesion studies, and drug manipulations, can provide important complementary information. ERPs, event-related potentials; MEG, magnetoencephalography; TMS, transcranial magnetic stimulation; EEG, electroencephalography; PET, positron emission tomography. (From Huettel, Song, and McCarthy 2009.)

Most early fMRI studies collected “proof-of-concept” data. They confirmed properties of brain function that had already been established through other methods; for example, that visual perception involved the occipital lobe, and that word generation recruited prefrontal and parietal cortex in the left hemisphere. Yet as the capabilities of fMRI became apparent, researchers soon began exploring new frontiers. A good example can be seen in the investigation of how new information is encoded into memory. If someone studies a list of 100 words, they will remember only some of those words at a test a week later—but it is very challenging to identify the brain processes that determine which words will be remembered and which forgotten. To overcome this challenge, researchers presented a series of words to participants being scanned using fMRI. They then compared the BOLD response to each word with the likelihood that the same word would be remembered in a later memory test. This approach, which essentially evaluates the fMRI signal as a potential predictor of subsequent memory, revealed that increased activity in regions of the medial temporal lobe was associated with successful encoding of a word into memory. Note that this type of study would not have been possible using other techniques of neuroscience—the verbal stimuli could not have been used in animal models, and the analysis method could not have been applied to PET data.

Researchers now use fMRI to study essentially every topic of interest to the cognitive scientist: from basic processes of perception and cognition to complex aspects of decision making and social behavior. Modeling of the BOLD signal has become much more sophisticated as well, going well beyond the simple “blocked design” approaches (e.g., alternating 30 seconds of Task A with 30 seconds of Task B) of early studies. Most current experiments record moment-to-moment changes in cognitive processing evoked by the subject's experimental task. This allows analysis of the fMRI data according to stimulus characteristics, participant behavior (e.g., the subsequent memory effect described earlier), or even a hypothesized computational model for cognition. The use of increasingly large sample sizes (often more than 30 individuals) allows researchers to study group and individual difference effects on brain function. The use of fMRI, accordingly, has now expanded into developmental and clinical neuroscience domains.

Limitations

Some of the limitations of fMRI are immediately apparent; for example, collecting fMRI data is expensive. These costs mean that fMRI experiments are not as common—or as large in sample size—as experiments that study behavior itself. Moreover, the collection and analysis of fMRI data require substantial technical training. Many of the top fMRI analysis packages are available free of charge, and researchers can download publicly available data sets to practice their analysis skills. But the complexity of these methods can still be daunting for many individuals; it is very easy to make mistakes in data analysis that render one’s results uninterpretable. For these and related reasons, fMRI has been frequently criticized in the popular media as a modern form of mindreading—it takes data from a few individuals, applies an incomprehensible set of analyses, and generates a speculative conclusion.

The limitations described above are real, but not insurmountable. The vast majority of current fMRI research follows rigorous experimental practices, in that it collects data from a reasonably large sample of individuals and applies the appropriate statistical corrections to ensure conservative but still meaningful conclusions. The primary limitations of fMRI are much more subtle and pernicious, reflecting the difficulty of linking experiments to cognition. If the experimental setup fails to manipulate the cognitive process of interest, the experiment cannot provide useful information about that process, regardless of the quality of the hardware or the competence of analyses. As researchers investigate increasingly complex aspects of cognition (e.g., moral judgment), it becomes even more important to have well-designed experiments that are closely linked to behavior.

Another problem is the breadth of stimuli that can evoke a given cognitive function. Two fMRI experiments may claim to study the same function (e.g., inhibition), but may use tasks that lead to different sorts of neural processes on the part of the subjects. Conversely, experiments that rely on the pattern of brain activation to determine the underlying cognitive process can lead to spurious “reverse inferences,” especially when a given brain region is involved in a wide array of tasks (Figure C). Many fMRI researchers have been exploring meta-analytic methods that aggregate across many fMRI studies to improve the validity and generalizability of conclusions. In summary, the core challenges for fMRI research arise not from problems in measuring brain function, but from linking those measurements to specific cognitive processes.

Figure C fMRI research involves making a chain of inferences: from experimental stimuli (1), to associated cognitive processes (2), to changes in the activity of neurons (3), to metabolic changes in the BOLD signal (4), and finally to statistical maps of activation (5) Usually, inference goes from left to right along this chain; for example, “because our experimental conditions differ in the amount of visible biological motion, the observed contrast in parietal cortex activation must reflect the role of that region in the processing of biological motion.” More problematic, though, is reversing the direction of inference, or drawing conclusions about what cognitive processes are involved in a task, based on the observed pattern of activation (blue arrow). See Poldrack 2006 for additional discussion. (From Huettel, Song, and McCarthy 2009.)

Future directions: Moving beyond localization of function

A longstanding criticism of fMRI—and indeed of many neuroscience techniques—is that it only provides a picture of “spots on a brain.” To someone outside the field, it may not seem that important to create a map of brain function; after all, we already know that each function is represented somewhere in the brain. In a very practical sense, this criticism is misplaced. Knowing how functions are localized to brain regions has considerable real-world clinical importance: for guiding neurosurgery, for anticipating the consequences of particular disorders, and for interpreting the effects of brain trauma. Advanced fMRI methods such as multivariate pattern analysis (MVPA) and fMRI-adaptation, moreover, can provide new insights into exactly what sort of information is being coded by a particular brain region. Thus the goal of localizing brain functions will remain central to fMRI research, and to cognitive neuroscience more generally, for the foreseeable future.

Even so, fMRI now extends well beyond simple mapping of functions to brain regions. Probably the most exciting new direction for fMRI research lies in mapping how sets of brain regions interact using what are often called “functional connectivity” analyses. These analyses are technically complex but have a simple idea at their core: that two regions whose activation follows similar patterns over time are likely to both contribute to the same brain function. Using this approach, researchers have mapped out the structure of large-scale networks in the brain in a manner not possible using other techniques. The most commonly studied of these is the default-mode network, a set of brain regions that tend to be most active when individuals are not performing a goal-directed task but instead are mind-wandering or otherwise engaged in unconstrained thought. Functional connectivity analyses can also provide new insights into how one region might control another. For example, a region in the medial frontal lobe is thought to shape decision making by acting much like a switch in a train yard—it changes how other regions interact based on the type of decision that is about to be made. These and other new methods of analysis point out the true power of fMRI: It gives neuroscientists the flexibility to ask many (but by no means all) questions about brain function while providing sufficient resolution for compelling answers.

References

Huettel, S. A., A. W. Song, and G. McCarthy. 2009. Functional Magnetic Resonance Imaging, 2nd ed. Sinauer, Sunderland, MA.

Moseley, M. E., and G. H. Glover. 1995. Functional MR Imaging: Capabilities and Limitations. In J. Kucharczyk, M. E. Moseley, T. Roberts, and W. U. Orrison (eds.), Functional Neuroimaging, pp. 161–191. W. B. Saunders, Philadelphia, PA.

Poldrack, R. A. 2006. Can cognitive processes be inferred from neuroimaging data? Trends Cog. Sci. 10: 59–63.

Copyright 2016 Sinauer Associates