Neuroimaging is an emergent method of investigation for studying the human brain in healthy and impaired populations. An increasing number of these investigations involve topics important to rehabilitation. Thus, a basic understanding of the more commonly used neuroimaging techniques is important for understanding and interpreting this growing area of research. Included in this article is a description of the signal source, the advantages and limitations of each technique, considerations for study design, and how to interpret cortical imaging data. Particular emphasis is placed on functional magnetic resonance imaging because of its ubiquitous presence in rehabilitation research.
Mapping brain function has been a pursuit of scientists for more than 2 centuries. Phrenologists proposed the first popular method in the early 19th century. They believed that the amount of brain tissue devoted to a cognitive function determined its influence on behavior.1 Although they were unable to measure cortical volume directly, they assumed that increases in brain size would translate into measurable bumps on the skull.
Although phrenology lacked scientific rigor, it did introduce the concept of brain function localization, in which different aspects of human behavior are represented primarily in specific locations of the brain. Subsequent advances were made by studying the effects of damage to the human brain, such as the localization of language in patients with stroke by Paul Broca,2 or by studying the effects of stimulation on brain areas during neurosurgery, such as the localization of the motor cortex by Fritsch and Hitzig.3 By the end of the 20th century, the study of functional localization took a major step forward with the development of new neuroimaging technologies. These noninvasive or minimally invasive methods allow the human brain to be studied in both healthy people and those affected by disease or injury. In addition, these approaches allow for repeated measurements in the same subject. Among these techniques, positron emission tomography (PET), transcranial magnetic stimulation (TMS), magnetoencephalography (MEG), and functional magnetic resonance imaging (fMRI) are used most commonly.
To date, the fields of neuroscience and psychology have dominated the neuroimaging landscape, but a growing number of researchers are using these techniques to investigate rehabilitation-related questions. It is only through elucidation of the scientific underpinnings of diseases, disorders, and recovery that effective treatments can be developed. Neuroimaging is a powerful tool that may be harnessed by rehabilitation specialists to improve understanding of the mechanisms by which treatments improve function. Therefore, the purpose of this article is to provide a brief overview of the commonly used neuroimaging techniques, with a particular emphasis on the technique most broadly used in the field of rehabilitation, fMRI. This article was written for people without an extensive physics or imaging background. We describe the foundation of fMRI, what to consider when one is designing or conducting an fMRI experiment, and how a clinician may interpret the results of research with fMRI.
The following investigative techniques all have advantages and disadvantages that are based on the strengths and limitations of each. We discuss the different techniques in terms of the degree of invasiveness, tasks that can be performed, spatial resolution, and temporal resolution. Spatial resolution refers to how accurately the measured activity is localized within the brain, and temporal resolution refers to how closely the measured activity corresponds to the timing of the actual neural activity.
It is important to appreciate that none of the techniques discussed in this article allows researchers to interpret brain activity at a scale that is available to invasive techniques, which measure responses from single neurons.4 However, given the relative advantages of these techniques over such invasive measures for studying brain activity in humans, this type of research will remain a staple for many years.
Magnetoencephalography is a technique that relies on the naturally occurring magnetic fields in living organisms. These fields are found in association with the electric currents that are part of normal brain function. As ionic currents change during brain activity, there is an associated change in the detectable magnetic fields. With MEG, it is believed that the detectable neuronal signals originate from dendritic activity in the pyramidal cells of the cerebral cortex.5 Thus, this technique provides a direct measure of cellular activity.
With the MEG technique, sensors arranged on the surface of the head measure the magnetic fields. Typically, the sensors are fixed in a helmet design that is placed on the head of the subject in a seated or supine position. The measurements made by each sensor reflect the synchronous activity of thousands of neurons contained in the brain areas under the sensor, particularly the activity that occurs in the sulci of the brain.6 This pattern is attributable to the nature of the magnetic field. The field follows the “right-hand rule,” whereby with the thumb pointed in the direction of the electric current, the magnetic field wraps around the current in the direction of the fingers. If the current is perpendicular to the surface of the head, then the magnetic field will be parallel to the surface of the head and will not thread through the detection coils. If the current is parallel to the surface, then the magnetic field will be oriented ideally to exit and enter the head (Fig. 1). This induced field is about 1 billion times smaller than the earth's magnetic field, approximately 50×10–15 to 500×10–15 T.6 For a detectable measurement to be obtained from this current, the subject being studied and the MEG instrument must be in a magnetically shielded room.
The MEG technique allows for a wide variety of behavioral tests encompassing visual, auditory, somatosensory, and motor studies.5–7 Baseline studies with no task also can be performed to assess resting brain activity. The data that are obtained must be processed before areas of task-related brain activity can be determined. One consideration of MEG analysis is the “inverse problem.” Although it is possible to calculate unambiguously the fields outside the head given the currents within the head, the reverse problem has no unique solution. To restrict the result to one solution representing the most likely brain activity pattern, one must add information in the form of assumptions of simplicity or data from other studies, such as other neuroimaging or neurophysiology studies.
The MEG technique provides excellent temporal resolution, on the order of 1 millisecond, thereby allowing determination of the temporal sequence of brain activity.6 In addition, it potentially can provide a spatial resolution of 5 mm.6 The processed MEG data can be presented in several ways, including magnetic field contour maps of the surface of the head or areas of brain activity displayed on an image of the brain obtained by magnetic resonance imaging (MRI). This technique can be used to answer many motor, cognitive, and sensory questions, particularly when the timing of neural activity is of interest. This technique has potential in the field of rehabilitation, as demonstrated by studies of stroke8 and aphasia.9
Positron emission tomography is a technique in which a radioactively labeled molecule is used (for a review, see Cherry and Phelps10). The label is a positron-emitting isotope. In PET studies, isotopes of carbon, nitrogen, oxygen, and fluorine (used as a hydrogen analog) commonly are used. These isotopes have a short half-life, ranging from 2 minutes to 2 hours (because of the short half-life of the isotopes, a device to produce the isotopes, a cyclotron, must be on-site or nearby, depending on the isotope). With PET, one of these isotopes is incorporated into a biologically relevant molecule and injected into or inhaled by the subject. It then concentrates in brain areas according to the molecule's chemistry and the brain's metabolic and blood flow demands. These areas are detected when a positron is emitted from the molecule. This positron is annihilated upon interaction with a nearby (1–2 mm) electron, producing 2 gamma rays almost perfectly 180 degrees apart. The PET scanner detects the 2 oppositely directed gamma rays when these rays hit 2 photodetectors coincidently and collinearly. These types of detection form the basis of the PET image.
Because a wide range of molecules can be used, PET can provide measurements of blood flow, blood volume, brain metabolism (especially glucose), and neuroreceptor or neurotransmitter chemistry. The spatial resolution of PET is approximately 4 to 5 mm.10 In functional studies, in which brain activity during task performance is compared with baseline activity, the temporal resolution is typically 1 to 2 minutes and is limited by both the technique and the metabolism of the molecule. An advantage of PET studies is the variety of data that can be obtained with the various labeled molecules that are available, such as 11C-labeled raclopride to study the dopamine D2 receptor, 15O-labeled water to study blood flow, and 2-fluoro-2-deoxy-d-glucose to study cerebral metabolic activity. Because this technique exposes the subject to radiation, a disadvantage is that repeated scanning is not possible. This disadvantage limits task repetition and the performance of multiple tasks in the same session. As with MEG, the data can be presented in several ways, including maps of the distribution of the labeled molecule or maps of task-related changes in brain activity. These maps also can be displayed on a brain image obtained by MRI.
Transcranial magnetic stimulation is unique among the techniques used to investigate function in the human brain (for a review, see Pascual-Leone et al11). Unlike methods of neuroimaging that provide measures of naturally occurring brain activity, the TMS technique uses a magnetic field to induce an electric current in underlying brain tissue, thereby stimulating the neurons. This technique also can be used to map brain function, elucidate brain areas involved in task performance by manipulating brain activity, or potentially provide treatment as a therapeutic intervention.
With TMS, a coil of wire (in either a circular or a figure eight configuration) is placed above the area of interest on the head of the subject. An electric current is passed through this wire, creating a magnetic field that passes painlessly (in most cases) and unimpeded through the scalp and skull into the brain below. This field induces an electric current, thereby stimulating the neurons in the brain. The spatial resolution is dependent on the type of coil, distribution of cerebrospinal fluid, and other factors and varies in the centimeter to millimeter range.4 The temporal resolution of the technique must be thought of differently, as activity is being induced rather than measured.
A stimulus can be given in under 1 millisecond, approximating the timescale of neural activity. The stimulation can be given as a single pulse, as paired pulses to a single brain area or 2 different brain areas, or as repetitive pulses.11 This flexibility in pulses allows the study of single brain areas, multiple brain areas and their interactions, connectivity between areas, inhibitory and excitatory circuits, and the behavioral effects of modulating cortical excitability.4,11 Output measures can be electromyographic responses in the muscles of interest, observable movement, patient reports (such as subtle visual perceptions, ie, phosphenes, with occipital cortex stimulation), and a disruption of task performance. The latter occurs when low-frequency stimulation is given immediately preceding or during task performance, thereby interfering with the normal activity of the brain area under the coil.12
The principal drawback of this technique is the direct stimulation of the brain, which carries the unlikely but still notable risk of a seizure. Therefore, proper investigator training and subject screening are required.13 In addition to the use of TMS as a brain-mapping device, repetitive TMS is under investigation as a treatment modality for a variety of neurological problems, such as depression, paralysis, bradykinesia, writer's cramp, stroke, and more.14–19
As the name implies, MRI uses strong magnetic fields to create images of biological tissues, taking advantage of properties that are intrinsic to tissues of the brain and, as such, does not expose the subject to radiation. MRI can provide information about several different atoms, such as hydrogen, phosphorus, carbon, and sodium atoms. However, in both anatomical imaging and functional imaging, the atom of interest is hydrogen, specifically, the hydrogen attached to water.
The static magnetic field created by an MRI scanner is expressed in the unit Tesla. For comparison, the earth's magnetic field is approximately 0.00005 T.20 The magnetic fields of scanners used for structural MRI are typically 1.5 to 3 T. To create images, the scanner uses a series of changing magnetic field gradients and oscillating electromagnetic fields, known as a pulse sequence, which is adjusted for the properties of hydrogen nuclei.21 The density and environment of hydrogen nuclei in different types of tissues allow pulse sequences to differentiate among tissue types, such as ligaments, tumors, and grey or white matter in the brain.
For an understanding of how this process occurs, knowledge of some physics concepts is required. One water molecule contains 1 oxygen atom and 2 hydrogen atoms (H2O). The hydrogen atom has 1 proton that spins and can be thought of as a small magnet that produces its own magnetic signal. Under normal circumstances, protons are randomly pointing in different directions, so that there is no net magnetic field (Fig. 2A, 1).22 The magnetic field in an MRI machine is always “on,” so that when a subject enters the magnet bore, all of the protons in the body align with the external magnetic field, creating a net internal magnetic field (Fig. 2A, 2). During an MRI, a second external magnetic field is applied via a radio-frequency (RF) pulse, causing the protons to wobble around their axis like a top. This wobbling creates a rotating magnetic field that changes with time and generates an electric current in the receiver (Fig. 2A, 3). This signal has both horizontal and vertical components (Fig. 2A, 4) and is used to create an image. When the RF pulse ends, the protons gradually align back to their original orientation.
Magnetic resonance imaging commonly is directed at measuring 2 processes of relaxation of the proton that are characterized by time constants: T1 and T2 (Fig. 2B). These processes take place as the wobbling protons relax back to their original state after the RF pulse ends. A T1-weighted scan measures the “righting” of the tipped proton as it realigns with the original magnetic field. The rate of this relaxation is influenced by nonexcited molecules in the surrounding tissue and is used for differentiating grey matter from white matter.23 In a T2-weighted scan, the focus is on the “falling out” or dephasing of synchrony of the rotating protons.23 Dephasing occurs quickly and results largely from the loss of energy to spinning nuclei nearby (and is influenced by the quality of the magnet used). These time constants vary in different tissue environments (such as grey matter and blood vessels) and form the basis for differentiating healthy tissues from diseased tissues.
Beyond the standard MRI technique that is useful for detecting pathology or trauma, different MRI acquisition procedures can be used to generate functional maps. These functional maps describe brain activity occurring in awake and performing humans and can be acquired on the basis of regional changes in tissue perfusion, cerebral blood volume, or the ratio of oxygenated hemoglobin to deoxygenated hemoglobin secondary to neural activity. The former 2 factors have not been used widely in functional imaging studies; therefore, the focus of this article is on the latter, which is referred to as blood oxygen level–dependent (BOLD) contrast. This is the source of the signal change in fMRI. T2 weighting, with an additional factor added to account for the nonhomogeneity of the magnetic field (and thus referred to as “T2* weighting”), is the basis of fMRI.23
BOLD Contrast Method
The BOLD contrast method is based on the hemodynamic response to neural activity and is thus an indirect measure of this activity. Since the late 1800s it has been known that blood flow increases with neural activity (for a historical review, see Friedland and Iadecola24). In 1936, it was discovered that the hemoglobin molecule has magnetic properties that differ depending on whether it is bound to oxygen. Oxygenated hemoglobin (hemoglobin with oxygen attached) has no magnetic properties, whereas deoxygenated hemoglobin (hemoglobin without oxygen attached) does.25 An increase in brain activity alters the ratio between the 2 forms of hemoglobin. In the early 1990s, it was determined that this phenomenon could be used to investigate brain function with what is known as the BOLD signal (T2*-weighted image).26
As explained above, fMRI is an indirect measure of neural activity, but recently the linkage between BOLD contrast and neural activity was explored in both animals27,28 and humans.29 Significant correlations were found between the BOLD signal and both the local field potential27 and the neuronal firing rate.29 Although more work is needed to clarify this relationship, these studies confirmed that the BOLD signal reflects neural activity.
The BOLD contrast in a brain region is dependent on the balance of oxygen consumption and oxygen supply in the microvasculature of the brain. The signal in BOLD contrast images is used to reveal where and with what intensity brain activity occurs during a behavioral task, such as finger tapping or viewing a flashing checkerboard. This technique is based on the same principles as traditional MRI, but the very fast imaging sequence used in BOLD fMRI (called “echoplanar imaging”) is sensitive to blood-based properties. Increased neural activity within a particular brain region results in an increase in blood flow and leads to decreased concentrations of deoxygenated hemoglobin in nearby vessels. This situation is somewhat counterintuitive, but the decrease is attributable to an excessive increase in the supply of oxygenated blood compared with what is required for the increased neural activity. The relative decrease in deoxygenated hemoglobin concentrations results in higher intensities on fMRI scans.
The fast acquisition time of fMRI allows whole brain images to be collected in about 3 seconds, meaning that hundreds of fMRI volumes (ie, one set of slices through the brain) can be collected in a given experiment. However, because of this fast acquisition time and the need for an adequate signal, fMRI does not have the fine degree of spatial resolution that traditional anatomical (T1-weighted) MRI has. After several preprocessing steps, the brain areas showing activity on images from fMRI are coregistered with (overlaid on) anatomical images to visualize anatomical landmarks and provide the differentiation of grey matter from white matter. Figure 3 shows the difference between the 2 types of acquisition in the same subject; Figure 3A shows an example of an axial slice of the brain during a T1-weighted structural scan, and Figure 3B shows an example of a T2*-weighted functional scan. Although the typical spatial resolution of functional scans is 3 to 5 mm, newer techniques with stronger magnets are enabling experimenters to view structures in the submillimeter range by using BOLD contrast.30,31
Considerations in Study Design
Given the dependence of BOLD contrast on blood flow, it is intuitive that there would be a delay between neural activity and the signal that can be detected (Fig. 4).32 This delay, which is referred to as a hemodynamic delay, limits the temporal resolution of this technique and has consequences for study design. Because of the approximate 3- to 5-second hemodynamic delay, fMRI is most useful for measuring the cortical activity associated with tasks that require several seconds to complete. More complex study designs can be used to correlate cortical activity with movement parameters or to allow for fMRI investigations of tasks with shorter durations, but these procedures are used infrequently in rehabilitation research and thus are not discussed here.
Most fMRI studies have a block or “boxcar” design, in which periods of a control (or resting) state are interleaved with periods of task performance, sensory stimulation, or both. The resultant images are generated by examining the statistical difference in BOLD signals between the control and task periods.23 In other words, the control period serves as a baseline activation level during which all of the extraneous stimulation (eg, noise of scanner, arousal state, and general sensory input) is accounted for, so that only the cortical activity that is different from the activity at baseline is considered to be part of the task. This method is known as the “subtractive method.”33 Although this method is used often, it is important to realize the assumptions that are being made with this design. The subtractive method assumes that the task phase adds cognitive processes to the resting phase secondary to the addition of a new behavior. In actuality, however, the neural processing between the control phase and the task phase may not be additive; rather, some cognitive processing may be replaced during the task phase.33
An fMRI study protocol involves both BOLD functional sequences and structural MRI. As mentioned above, the structural scan is necessary to visualize the brain and ultimately to serve as a template upon which to localize the activation that is acquired by fMRI. For the structural procedure, a subject lies down and is moved into the bore of the MRI scanner. No functional experiment is performed during this step because it is for anatomical purposes only. Following this step, the functional part of fMRI is done. During this part of the study, the subject remains in the bore of the magnet but performs a task as images are acquired. Often, instructions for task performance are presented on a small projector screen visible to the subject with a mirror positioned above the eyes. A computer located outside the scanning room is dedicated to presenting the experimental paradigm and recording the responses of the subject. The task conditions that can be used are limited by the space considerations of the magnet bore and the noise of the scanner but are nevertheless numerous. They include sensory or visual stimulation, cognitive tasks, and motor performance. Figure 5 shows an example of a setup in which an MRI-compatible electrogoniometer was used to record finger motion while subjects tracked a sine wave.34
Typically, the fMRI task or tasks are repeated several times. Repeated trials increase the statistical power by increasing the number of comparisons of task and control conditions. It is common to include control periods between task periods to allow blood flow to return to the resting state prior to the next task.
Experiments involving fMRI can be conducted with virtually any MRI scanner, as manufacturers offer this as an option with their equipment. Most hospitals have a 1.5- or 3-T scanner, as 1.5 or 3 T is the field strength used for most fMRI studies, particularly for clinical populations. High-field-strength MRI (4 T or greater) typically is performed at MRI research centers and can be done with field strengths of 4 T, 7 T, and higher. A higher field strength, although associated with greater technical challenges, provides a closer spatial match between neural activity and the resultant record of activation and thus better spatial resolution.30
Preprocessing of Data
The raw fMRI data need to be prepared before statistical comparisons can be performed. An important step in this process is to account for the fact that each slice in an fMRI volume (collection of slices) was acquired at a slightly different time. Thus, slice-timing corrections are performed on the data so that it appears that all voxels (defined below) within one volume were acquired at exactly the same time.35 Additional preprocessing removes various artifacts in the data, such as rifts in the data or movement artifacts caused by subject motion. Numerous brain imaging software packages are available to perform the preprocessing and statistical analysis steps in the experiment.
Quantification of Activation
The basis for detecting brain activation lies in the statistical comparison of the BOLD signals of the task blocks and the control blocks for each voxel within the magnetic resonance image and therefore the brain. A voxel is the unit volume of a magnetic resonance image and defines its spatial resolution. Each magnetic resonance image can be thought of as a checkerboard with rows and columns of boxes with a depth or slice thickness. That is, the dimensions of the box define the size of the voxel. The activation obtained in an experiment can be quantified in terms of either spatial extent (volume), which is expressed in terms of the number of voxels, or the intensity of activation within the voxels.
There are several methods for determining and quantifying cortical activation. Multiple t tests or F tests based on the assumptions of the general linear model generally are used, requiring a voxel-by-voxel comparison throughout the brain (for a more detailed discussion, see Smith35). A cross-correlation analysis also can be used; this analysis involves comparisons of a time course of the signal strength of each voxel with an idealized function of the task and control periods. For most methods, a statistical threshold of activation is used as a cutoff for what is considered active. For example, P=.05 may be used as the threshold, all voxels below this predefined statistical threshold will be considered active, and those above this threshold will not.
A volume analysis is based on a volume measure of activation; that is, the numbers of voxels activated during the experimental task are summed and then localized to a given region. Volume analysis is subject to several difficulties. A primary problem is that a threshold is used to determine what is considered active and what is not, and often this threshold is not subjected to corrections for multiple comparisons; not subjecting the threshold to such corrections could result in false-positive findings.36 Although most investigators set the threshold a priori, it is still an arbitrary number, meaning that the use of a different threshold will produce different results. A second problem with a volume analysis is that a given anatomical area of interest (also called a “region of interest” [ROI]) may not be equivalent between subjects. For example, 2 subjects both may have 200 activated voxels in the primary motor area, but given variations in brain size, that quantity may represent different percentages of the total area available. Volume measures of voxel activation also have been shown to have lower levels of reproducibility than other methods of reporting.37
To address problems of variability, indexes of activation can be calculated to normalize the activation within a subject. A commonly used one is the laterality index.38 This calculation is used to describe to what degree the contralateral or ipsilateral cortical areas are activated relative to the stimulus or movement, as follows: LI=(C−I)/(C+I). In this equation, LI is the laterality index, C is the number of voxels in the contralateral hemisphere, and I is the number of voxels in the ipsilateral hemisphere. This equation produces a range of values from −1 to +1, with a negative number indicating primarily ipsilateral activation and a positive number indicating primarily contralateral activation. This index is particularly useful for describing the control of movement in a subject with a brain lesion.34,39–41
Another method, referred to as the “percent contribution method,” takes into account the general level of excitability or brain activation in a subject. With this method, the total numbers of activated voxels in the areas of interest are summed, and the percentage of the total activation is determined for each area of interest. For example, if the primary sensory cortex has 250 voxels, the primary motor cortex has 100 voxels, the supplementary motor area has 30 voxels, and the premotor cortex has 120 voxels, then the total activation is 500 voxels. The percent contribution of the primary motor cortex therefore is 20% of the total activation. Preliminary work in our laboratory has suggested that this index produces more reproducible results across scanning sessions (unpublished data, 2007).
An alternative to expressing activation in terms of the number of activated voxels (volume method) is to examine the intensity of an activated cortical region. This examination can be done for an entire ROI, regardless of the areas of activation within it (thus eliminating the threshold problem), or just for those voxels that were deemed active because they exceeded the given statistical threshold. The intensity method allows the examiner to determine the degree of intensity change during the task. That is, in a comparison of 2 groups of subjects (such as a control group and a treatment group), there may not be a difference in the volume of activation (number of voxels or laterality of activation), but the intensity of the BOLD signal may be stronger in 1 group. Intensity measures have been found to be less variable than volume measures.37,42–44
The intensity of a response is derived from the BOLD signal time course, which provides the raw average level of baseline activation during a control phase and the β weight (or change in signal intensity) during a task phase. The β weight and the level of baseline activation are used to determine the percent change in signal intensity for each task period. The raw numbers can vary greatly between people or testing sessions,37 making direct comparisons difficult. Normalizing the task data to the resting data allows for intersubject and intrasubject comparisons.45,46 The percent change value then can be used in standard statistical comparisons.
This result of examining the intensity of activation in an ROI can be expressed as an intensity index, which is defined as follows: [(task intensity −resting intensity)/resting intensity]×100. In this equation, task intensity is the average raw signal intensity during the task periods and resting intensity is the average raw signal intensity during the resting periods.44 This method is useful for determining whether, within an area of interest, there was a change in the intensity of the BOLD signal between a control phase and a task phase.
All indexes have a limitation of specificity. That is, each time a mathematical manipulation of the data is made, subtleties are lost. This limitation is an issue that researchers need to determine a priori on the basis of their scientific question and the study design.
Localization of Activation
Once activation has been quantified, the specific location of the activation must be determined. Localization can be done on the basis of an individual's anatomy, but it is most commonly performed with Talairach coordinates because of the need for comparisons and averaging of data across subjects.47 During the MRI preprocessing step, a subject's brain is “warped” into Talairach space. The coordinates then can be used to determine where in the brain (eg, Brodmann areas 1–3 or cingulate gyrus) the center of activation is located. Comparisons of groups or tasks can be made by performing statistical comparisons of the anterior and posterior, the right and left, and the superior and inferior coordinates to determine whether the location of activation has shifted within a particular region. For example, a treatment technique may result in a shift of activation in the anterior direction.
Although the normalization step is necessary, it has obvious disadvantages. When a subject's brain is warped into an idealized brain, spatial specificity is reduced. This problem is of particular concern in the examination of a subject with a brain lesion, as the nonlinear transformation associated with the “warp” will be particularly sensitive to drastic image density changes (such as those associated with a brain lesion).48 In addition, brain atrophy secondary to injury or age may distort normal anatomy, making comparisons with a normal coordinate system more difficult. There is currently no widely embraced alternative method; however, given the variability in human brain size and shape, some sort of processing always will be required to allow for comparisons or averaging of data across subjects.
Potential Pitfalls of fMRI
Careful selection of volunteers is the first issue to consider in the design of an fMRI experiment. Because of the high magnetic field, most ferrous metals are not allowed in the scanner, because of the potential for movement or heating of implanted devices. Items that are contraindicated include aneurysm clips, pacemakers, and cochlear implants. Many other items may be contraindicated but, under certain circumstances, may be allowed. Consideration should be given to a broad range of metals, including jewelry, body piercing, dental metal, fragments from injury, occupational metal fragments (such as those encountered by metal workers), and surgical implants or clips. More unusual sources of concern are medication patches and tattoos. These issues require further investigation and vary on the basis of the year in which they were acquired and the materials used. Generally, if a subject has had any prior surgical procedure, it should be investigated, and questions regarding the nature of the surgery should be directed to a physician familiar with MRI experiments.
Given the enclosed nature of the MRI scanner, subjects should be screened for claustrophobia, as they may become anxious. Finally, consideration should be given to the medical history of the subject. Many medications have effects on the brain and its vessels, and these effects may influence brain activity, task performance, or both. These medications include antidepressants, anxiolytics, anticonvulsants, antipsychotics, antihypertensives, anticholinergics, and the treatments used in Parkinson disease and Alzheimer disease. Although these medications are not necessarily exclusion criteria, their effects are an issue that each investigator should consider carefully. Also to be considered is whether the disease history varies within the study group and includes disorders that may affect the hemodynamic response, such as vascular disease.
As with any highly technical scientific technique, several issues require attention when experiments are being conducted. Possibly the most vexing of these in fMRI is excessive head motion. Unfortunately, it is quite common for a subject to move his or her head while performing a task in the MRI scanner. This motion may involve a startle reaction to stimulus onset or the beginning of scanner noise. It may involve a tracking motion of the head that is associated with the task,49 or it may simply be attributable to the inability of the subject to remain still for an extended period of time. This problem is of particular concern when patients, older people, or children are being studied, as the ability of these populations to inhibit motion may be reduced.
The problem with unwanted head motion is two-fold. Excessive motion results in a false activation pattern because the fMRI signal changes caused by the hemodynamic response are smaller than the apparent signal differences that result from head movement. Additionally, movement adds to the variance of the signal, which in turn reduces the sensitivity of the statistical tests to detect true activation, resulting in a possible type II error.50 Typically, brain activation related to motion artifacts is most evident on the brain surface (rim activation) and at the interface between cerebrospinal fluid and brain tissue (such as that seen around the ventricles).
Because of the pervasiveness of this issue, fMRI software packages perform motion correction during the preprocessing phase of data analysis. It is important to examine all data for motion and to set an a priori threshold for data that should be excluded. For example, movement greater than or equal to the voxel resolution is discarded. That is, if the voxel resolution is typically 3 mm3, any trial with movement of 3 mm or greater is discarded. In our experience, data for up to 20% of subjects may need to be excluded for this reason.
Several methods can be used to reduce head movement. Head cushioning (such as stabilization with cushions or straps) is a safe deterrent that typically is well tolerated by all subjects. Other methods include a “bite bar” to guarantee stabilization, but this may not be well tolerated by subjects and may lead to difficulty with recruitment. Shorter experimental times or mock scanner training sessions also may reduce the propensity for movement.51 Algorithms also are used during image acquisition to perform real-time motion correction to help minimize this problem.52
In addition to movement of the head, unwanted eye movement is another concern. For example, when cortical activity as a result of motor task performance is being investigated, activation may be attributable to the eye movement associated with the task and not to the target task. Instructing the subject to maintain eye fixation or monitoring eye motion with an eye-tracking device can assist in dissociating the potential confounding effects of unwanted eye motion.53 There are several difficulties with the use of eye tracking in an MRI scanner, but eye tracking is the most quantifiable method for measuring subject eye motion and should be considered if eye motion is a likely confounding variable.
The muscular activation strategy (pattern of muscle use) and mirror movement (movement of the opposite hand [or foot]) during the performance of a task are other issues that require consideration when fMRI experiments are being conducted or when fMRI data are being interpreted. Mirror movement is especially important when patients are being studied, as inhibition of motion may be a problem in this population.54 The muscular activation strategy during task performance is applicable to all studies that require movement. For example, cortical activation pattern changes after a treatment could be attributable to either a change in neural control or a change in muscular control. Ideally, electromyography (EMG) would be used to monitor for muscular activity in both the performing limb and the opposite limb. The use of EMG in the magnetic environment is hampered by difficulties associated with interference between the MRI scanner and the EMG amplifier. Some researchers address this interference by using EMG recording during a mock scanner performance of the task55 or by introducing pauses in the scanning protocol to allow for brief periods of EMG recording.56 Other methods for observing mirror movement involve monitoring gross motion (such as with an electrogoniometer or a joystick)34,57 or visual observation, but these methods are limited also, as there could be extraneous EMG activity of insufficient magnitude to produce observable motion.
BOLD signal activation has been shown to be affected by emotional status, such as depression58,59 or anxiety.60,61 As such, a comparison of groups of people who may differ in these affective domains requires consideration. Additionally, within the same person, changes in affect from one testing session to another may be reflected in changes in the magnitude of the BOLD signal. To increase the probability that the changes seen are attributable to task performance, emotional status needs to be documented prior to image acquisition. Although changes in emotional status (such as anxiety or depression) would not nullify neuroimaging findings per se, such changes should be taken into account during post hoc analysis. This objective could be achieved by categorizing subjects on the basis of emotional status and searching for differences between groups, perhaps by using the Beck Depression Inventory62 or by adding emotional status as a covariable.
There is growing evidence that task load, or the degree of task demand during an experiment, may influence the magnitude or extent of brain activation.63–65 In a comparison of BOLD signal data from a healthy population with data from a patient population, task load can be of particular concern, because a given task may be less demanding for people who are healthy. Indeed, even imagination of a movement can produce significant levels of activation.57,66–68
It is impractical to control completely for task load a priori, but pilot data on task performance alone (without neuroimaging) would assist in determining whether performance varies between groups. This situation also can be addressed by using measures of accuracy or performance time for post hoc covariables in the analysis. Another method of equalizing task load is to “titrate” the task, that is, modify task difficulty so that all participants maintain the same level of performance (eg, a 75% accuracy of responses).69 Previous work34,44,57 has minimized the problem of task load through the use of age-matched control subjects in studies of subjects with stroke, thereby removing the effect of age on performance. In investigations of the cortical effects associated with a treatment or technique, a control group allows a comparison of groups. A significant difference in a controlled design lends confidence that the results are not attributable to variability of the testing measure but rather represent a true difference between groups.
Functional MRI is an expensive research tool. The high costs are justified by the technical maintenance and highly trained personnel required to keep the device operating. These high costs often result in another problem: a small number of subjects. A small number of subjects, the inherent variability of fMRI, and the variability associated with patient populations (such as stroke location and size and atrophy because of age) can make finding significant results difficult. Is a negative result truly representative of no difference, or is it attributable to a type II error that could be solved with more power? This problem is not unique to fMRI but requires caution in the interpretation of experiments with negative results.52
Although it is true that several issues can lead to artifacts (errors) in signals, most of them can be addressed with careful planning of an experimental design. Specifically, for examining a treatment approach in a pretest-posttest design, it is imperative to use a randomized control group. This group should be matched to the treatment group with regard to age, sex, and handedness. Comparisons of the 2 groups then can be made with confidence, because the error associated with task performance can be assumed to be equally distributed across the groups. Likewise, for examining changes within a subject (versus group data), a multiple-baseline design may be considered to account for the variability that will occur from one testing session to the next, independent of any treatment effect. Functional MRI always will have error, but it is not always convenient or even possible to measure all potential sources of error. Careful planning of an experimental design can lead to greater confidence that any effect seen is attributable to the experimental conditions and not to ubiquitous artifacts.
It is tempting to conclude from “lit up” areas of activation in the cortex that the location of activation responsible for the observed behavior has been identified. This conclusion is not entirely correct. In the interpretation of neuroimaging data, it is important to remember that activation in a given brain region means that that region is associated with the performance of that task. The observed activation alone likely is insufficient to produce movement, and it cannot be determined whether the observed activation is attributable to the activation of excitatory or inhibitory neurons.70 It is easier to keep this cautious interpretation in perspective when one remembers the multiple constraints associated with neuroimaging, such as the indirect measurement of neural activity, “warping” of the brain into Talairach space, the use of a threshold for activation, and the averaging of signals over a period of many seconds. Additionally, some of the effects of error may be even more pronounced in the interpretation of fMRI investigations of patient populations. Indeed, more work needs to be done to optimize this technique for patient populations.
The silver lining in the cloud outshines the grey, however. Neuroimaging already has yielded numerous noteworthy works in the field of rehabilitation, as highlighted in this Special Series. With proper study design and widespread critical interpretation, neuroimaging will become an increasingly powerful tool that will help to shape the understanding of mechanisms of recovery and influence physical therapy intervention of the future.
For more about this PTJ Special Series on the role of neuroimaging in rehabilitation, read the editorial by Richard K Shields on page 639.
Both authors provided concept/idea/project design and writing.
- Received June 1, 2006.
- Accepted January 10, 2007.
- Physical Therapy