Functional magnetic resonance imaging

Functional magnetic resonance imaging or functional MRI (fMRI) is a type of specialized MRI scan used to measure the hemodynamic response (change in blood flow) related to neural activity in the brain or spinal cord of humans or other animals. It is one of the most recently developed forms of neuroimaging. Since the early 1990s, fMRI has come to dominate the brain mapping field due to its relatively low invasiveness, absence of radiation exposure, and relatively wide availability.

Background
Since the 1890s it has been known that changes in blood flow and blood oxygenation in the brain (collectively known as hemodynamics) are closely linked to neural activity. When neural cells are active they increase their consumption of energy from glucose and switch to less energetically effective, but more rapid anaerobic glycolysis. The local response to this energy utilization is to increase blood flow to regions of increased neural activity, which occurs after a delay of approximately 1–2 seconds. This hemodynamic response rises to a peak over 4–6 seconds, before falling back to baseline (and typically undershooting slightly). This leads to changes in local cerebral blood volume and local changes in the local concentration of oxyhemoglobin that are detectable through their paramagnetic effects.

History
Blood-oxygen-level dependence (BOLD) is the MRI contrast of blood deoxyhemoglobin, first discovered in 1990 by Seiji Ogawa at AT&T Bell labs. Ogawa and colleagues had recognized the potential importance of BOLD for functional brain imaging with MRI, but the first successful fMRI study was reported by John W. Belliveau and colleagues in 1991 using an intraveneously administered paramagnetic contrast agent (Gadolinium). Using a visual stimulus paradigm, localized increases in blood volume (32 +/- 10 percent, n = 7 subjects) were detected in the primary visual cortex. In 1992, three papers were published using endogenous BOLD contrast MRI. One was submitted by Peter Bandettini at the Medical College of Wisconsin on February 5, revised March 31, accepted March 31 and published in the June 1992 issue of Magnetic Resonance in Medicine (MRM). The second by Kenneth Kwong and colleagues also applied BOLD to image human brain activities with MRI and was submitted on March 26 and published in the June issue of PNAS in 1992. In the same year, Dr. Ogawa submitted their result on March 31 and published in July issue of PNAS. In the following year, Dr. Ogawa published the biophysics model of BOLD contrast in Biophysical Journal. Dr. Bandettini also published a further paper in 1993 demonstrating quantitative determination of functional activation maps.

Physiology
As neurons do not have internal reserves for glucose and oxygen, more neuronal activity requires more glucose and oxygen to be delivered rapidly through the blood stream. Through a process called the hemodynamic response, blood releases glucose to neurons and astrocytes at a greater rate than in the area of inactive neurons. It results in a surplus of oxyhemoglobin in the veins of the area and distinguishable change of the local ratio of oxyhemoglobin to deoxyhemoglobin, the "marker" of BOLD for MRI.

Hemoglobin is diamagnetic when oxygenated (oxyhemoglobin) but paramagnetic when deoxygenated (deoxyhemoglobin). The magnetic resonance (MR) signal of blood is therefore slightly different depending on the level of oxygenation. Higher BOLD signal intensities arise from increases in the concentration of oxygenated hemoglobin since the blood magnetic susceptibility now more closely matches the tissue magnetic susceptibility. By collecting data in an MRI scanner with sequence parameters sensitive to changes in magnetic susceptibility one can assess changes in BOLD contrast. These changes can be either positive or negative depending upon the relative changes in both cerebral blood flow (CBF) and oxygen consumption. Increases in CBF that outstrip changes in oxygen consumption will lead to increased BOLD signal, conversely decreases in CBF that outstrip changes in oxygen consumption will cause decreased BOLD signal intensity. The signal difference is very small, but given many repetitions of a thought, action or experience, statistical methods can be used to determine the areas of the brain which reliably show more of this difference as a result, and therefore which areas of the brain are active during that thought, action or experience.

Almost all current fMRI research uses BOLD as the method for determining where activity occurs in the brain as the result of various experiences, but because the signals are relative and not individually quantitative, some question its rigor. Other methods which propose to measure neural activity more directly have been attempted (for example measurement of the Oxygen Extraction Fraction (OEF) in regions of the brain, which measures how much of the oxyhemoglobin in the blood has been converted to deoxyhemoglobin or direct detection of magnetic fields generated by neuronal currents ), but because the electromagnetic fields created by an active or firing neuron are so weak, the signal-to-noise ratio is extremely low and statistical methods used to extract quantitative data have been largely unsuccessful as of yet.

Neural correlates of BOLD
The precise relationship between neural signals and BOLD is under active research. In general, changes in BOLD signal are well correlated with changes in blood flow. Numerous studies during the past several decades have identified a coupling between blood flow and metabolic rate; that is, the blood supply is tightly regulated in space and time to provide the nutrients for brain metabolism. However, neuroscientists have been seeking a more direct relationship between the blood supply and the neural inputs/outputs that can be related to observable electrical activity and circuit models of brain function.

While current data indicate that local field potentials, an index of integrated electrical activity, form a marginally better correlation with blood flow than the spiking action potentials that are most directly associated with neural communication , no simple measure of electrical activity to date has provided an adequate correlation with metabolism and the blood supply across a wide dynamic range. Presumably, this reflects the complex nature of metabolic processes, which form a superset with regards to electrical activity. Some recent results have suggested that the increase in cerebral blood flow (CBF) following neural activity is not causally related to the metabolic demands of the brain region, but rather is driven by the presence of neurotransmitters, like glutamate, serotonin, nitric oxide, acetylcholine, dopamine and noradrenaline.

Some other recent results suggest that an initial small, negative dip before the main positive BOLD signal is more highly localized and also correlates with measured local decreases in tissue oxygen concentration (perhaps reflecting increased local metabolism during neuron activation). Use of this more localized negative BOLD signal has enabled imaging of human ocular dominance columns in primary visual cortex, with resolution of about 0.5 mm. One problem with this technique is that the early negative BOLD signal is small and can only be seen using larger scanners with magnetic fields of at least 3 Tesla. Further, the signal is much smaller than the normal BOLD signal, making extraction of the signal from noise more difficult. Also, this initial dip occurs within 1–2 seconds of stimulus initiation, which may not be captured when signals are recorded at long repetition (TR). If the TR is sufficiently low, increased speed of the cerebral blood flow response due to consumption of vasoactive drugs (such as caffeine ) or natural differences in vascular responsiveness may further obscure observation of the initial dip.

The BOLD signal is composed of CBF contributions from larger arteries and veins, smaller arterioles and venules, and capillaries. Experimental results indicate that the BOLD signal can be weighted to the smaller vessels, and hence closer to the active neurons, by using larger magnetic fields. For example, whereas about 70% of the BOLD signal arises from larger vessels in a 1.5 tesla scanner, about 70% arises from smaller vessels in a 7 tesla scanner. Furthermore, the size of the BOLD signal increases roughly as the square of the magnetic field strength. Hence there has been a push for larger field scanners to both improve localization and increase the signal. A few 7 tesla commercial scanners have become operational, and experimental 8 and 9 tesla scanners are under development.

Technique
BOLD effects are measured using rapid volumetric acquisition of images with contrast weighed by T1 or T2*. Such images can be acquired with moderately good spatial and temporal resolution; images are usually taken every 1–4 seconds, and the voxels in the resulting image typically represent cubes of tissue about 2–4 millimeters on each side in humans. Recent technical advancements, such as the use of high magnetic fields and multichannel RF reception,  have advanced spatial resolution to the millimeter scale. Although responses to stimuli presented as close together as one or two seconds can be distinguished from one another, using a method known as event-related fMRI, the full time course of a BOLD response to a briefly presented stimulus lasts about 15 seconds for the robust positive response.

fMRI studies draw from many disciplines
fMRI is a highly interdisciplinary research area and many studies draw on knowledge in several fields:
 * Physics: Physical principles underlie fMRI signals and many studies require an understanding of these underlying principles.
 * Psychology: Almost all fMRI studies are essentially cognitive psychological, cognitive psychophysiological, and/or psychophysical experiments in which the MRI scanner is used to obtain an extra set of measurements in addition to behavioral or electroencephalographic measurements.
 * Neuroanatomy: The fMRI signals can be put into the context of previous knowledge only with an understanding of the neuroanatomy.
 * Statistics: Correct application of statistics is essential to "tease out" observations and avoid false-positive results.
 * Electrophysiology: Familiarity with neuronal behavior at the electrophysiological level can help investigators design a useful fMRI study.

Advantages and Disadvantages of fMRI
Like any technique, fMRI has advantages and disadvantages, and in order to be useful, the experiments that employ it must be carefully designed and conducted to maximize its strengths and minimize its weaknesses.

Advantages of fMRI

 * It can noninvasively record brain signals without risks of ionising radiation inherent in other scanning methods, such as CT or PET scans.
 * It has high spatial resolution. 2–3 mm is typical but resolution can be as good as 1mm.
 * It can record signal from all regions of the brain, unlike EEG/MEG which are biased towards the cortical surface.
 * fMRI is widely used and standard data-analysis approaches have been developed which allow researchers to compare results across labs.
 * fMRI produces compelling images of brain "activation".

Disadvantages of fMRI

 * The images produced must be interpreted carefully, since correlation does not imply causality, and brain processes are complex and often non-localized.
 * Statistical methods must be used carefully because they can produce false positives. One team of researchers studying reactions to pictures of human emotional expressions reported a few activated voxels in the brain of a dead salmon when no correction for multiple comparisons was applied, illustrating the need for rigorous statistical analyses.
 * The BOLD signal is only an indirect measure of neural activity, and is therefore susceptible to influence by non-neural changes in the body. This also means that it is difficult to interpret positive and negative BOLD responses
 * BOLD signals are most strongly associated with the input to a given area rather than with the output. It is therefore possible (although unlikely) that a BOLD signal could be present in a given area even if there is no single unit activity.
 * fMRI has poor temporal resolution. The BOLD response peaks approximately 5 seconds after neuronal firing begins in an area. This means that it is hard to distinguish BOLD responses to different events which occur within a short time window.  Careful experimental design can reduce this problem.  Also, some research groups are attempting to combine fMRI signals that have relatively high spatial resolution with signals recorded with other techniques, electroencephalography (EEG) or magnetoencephalography (MEG), which have higher temporal resolution but worse spatial resolution.
 * fMRI has often been used to show activation localized to specific regions, thus minimizing the distributed nature of processing in neural networks. Several recent multivariate statistical techniques work around this issue by characterizing interactions between "active" regions found via traditional univariate techniques.
 * The BOLD response can be affected by a variety of factors, including: drugs/substances; age, brain pathology; local differences in neurovascular coupling; attention; amount of carbon dioxide in the blood; etc.

For these reasons, Functional imaging provides insights into neural processing that are complementary to insights of other studies in neurophysiology.

Scanning in practice
Subjects participating in a fMRI experiment are asked to lie still and are usually restrained with soft pads to prevent movement from disturbing measurements. Some labs also employ bite bars to reduce motion, although these are unpopular as they can be uncomfortable. Small head movements can be corrected for in post-processing of the data, but large transient motion cannot be corrected. Motion in excess of around 3 millimeters results in unusable data. Motion is an issue for all populations, but most especially problematic for subjects with certain medical conditions (e.g. Alzheimer's Disease or schizophrenia) or with young children. Participants can be habituated to the scanning environment and trained to remain still in an MRI simulator.

An fMRI experiment usually lasts between 15 minutes and an hour. Depending on the purpose of study, subjects may view movies, hear sounds, smell odors, perform cognitive tasks such as n-back, memorization or imagination, press a few buttons, or perform other tasks. Researchers are required to give detailed instructions and descriptions of the experiment plan to each subject, who must sign a consent form before the experiment.

Safety is an important issue in all experiments involving MRI. Potential subjects must ensure that they are able to enter the MRI environment. The MRI scanner is built around an extremely strong magnet (1.5 teslas or more), so potential subjects must be thoroughly examined for any ferromagnetic objects (e.g. watches, glasses, hair pins, pacemakers, bone plates and screws, etc.) before entering the scanning environment.

Related techniques
Aside from BOLD fMRI, there are other related ways to probe brain activity using magnetic resonance properties:

Diffusion based functional MRI
Neuronal activity produces some immediate physical changes in cell shape that can be detected because they affect the compartment shape and size for water diffusion. A much improved spatial and temporal resolution for fMRI data collection has now been achieved by using diffusion MRI methodology that can detect these changes in neurons. The abrupt onset of increased neuron cell size occurs before the metabolic response commences, is shorter in duration and does not extend significantly beyond the area of the actual cell population involved. This technique is a diffusion weighted technique (DWI). There is some evidence that similar changes in axonal volume in white matter may accompany activity and this has been observed using a DTI (diffusion tensor imaging) technique. The future importance of diffusion-based functional techniques relative to BOLD techniques is not yet clear.

Contrast MR
An injected contrast agent such as an iron oxide that has been coated by a sugar or starch (to hide from the body's defense system), causes a local disturbance in the magnetic field that is measurable by the MRI scanner. The signals associated with these kinds of contrast agents are proportional to the cerebral blood volume. While this semi-invasive method presents a considerable disadvantage in terms of studying brain function in normal subjects, it enables far greater detection sensitivity than BOLD signal, which may increase the viability of fMRI in clinical populations. Other methods of investigating blood volume that do not require an injection are a subject of current research, although no alternative technique in theory can match the high sensitivity provided by injection of contrast agent.

Arterial spin Labeling (ASL)
Arterial Spin Labelling (ASL), also known as arterial spin tagging, is an MRI technique capable of measuring cerebral blood flow (CBF) in vivo. ASL is capable of providing cerebral perfusion maps, without requiring the administration of a contrast agent or the use of ionising radiation, as it uses magnetically-labelled endogenous blood water as a freely-diffusible tracer. It was first proposed in 1992 and has since benefited from a number of modifications aimed at improving its robustness. ASL can monitor changes in CBF with activation and fMRI studies can therefore be conducted using ASL instead of relying on the BOLD effect. ASL fMRI is less popular than BOLD, as it suffers from a lower signal to noise ratio, can be less sensitive to weak stimuli and its temporal resolution is poorer than in BOLD studies. On the plus side, it can provide quantitative measures of a single well-defined parameter, CBF, whose baseline value can also be determined in the same experiment. It has also been found to outperform BOLD in terms of stability to slow signal drifts and localization of the activation area. The ASL activation signal is believed to be dominated by changes in the capillary bed of the activated area of the cortex, wheareas the BOLD signal is likely to be dominated by changes in the oxygenation of nearby veins.

The ASL Experiment

The mains steps of a standard pulsed ASL experiment are summarized below. It starts with the application of an inversion pulse to part of the neck (the inversion slab). This step aims to invert (i.e. label) the magnetization of arterial blood. A time TI later, when the labelled blood is expected to have reached the brain, an excitation pulse is applied to the area of the brain where CBF is to be measured (the imaging slice). Some of the signal in the acquired image will have come from the inverted blood that has reached the imaging slice (this is typically 1% of the total signal), but the overwhelming majority will have come from static brain tissue. This image is called the labelled (or tag) image. To remove the contribution from static tissue, an identical image (the control image) is acquired in the absence of the blood labeling step. To mimic magnetization transfer (MT) effects of the blood labeling pulse, however, the control image acquisition needs to be preceded by a pulse with a similar RF deposition to the one used to invert blood. The tag image can be subtracted from the control to yield an image whose signal comes exclusively from labelled blood – this is the perfusion weighted image (PWI). Fast imaging techniques such as EPI are usually employed to allow the perfusion state of the brain to be captured in as short a time as possible. To improve the SNR of the PWI, several pairs of similar control and tag images are usually acquired and then averaged. Finally, an absolute CBF map can be generated from PWIs acquired at one or more inversion times by the application of a suitable model. There is an immense number of variations to this basic experiment, according to the way labeling is performed, the method used to control MT effects and the types of image readout employed. Some of these are discussed below.

Labeling in ASL

Two basic methods exist to perform labeling in ASL: continuous ASL (CASL) and pulsed ASL (PASL). In CASL, a 2-4 second long flow-induced adiabatic inversion pulse is applied at the level of the internal carotid arteries. The CASL labeling scheme has a relatively low inversion efficiency and causes high levels of power deposition. Some of these problems can be minimized using a train of RF pulses to perform a flow-induced adiabatic inversion, a technique called pseudo-continuous arterial spin labeling (PCASL). Blood inversion can also be accomplished using a short inversion pulse applied to a larger (100-150 mm wide) region of the neck. This method, pulsed arterial spin labeling (PASL), shows higher inversion efficiencies (very close to 1), particularly when adiabatic inversion pulses are used. However, imperfect slice selection profiles imply that a gap of 10-30 mm must be left between the inversion slab and the brain area to be imaged. This introduces a delay in the arrival of the inverted blood to the imaging slice, during which it experiences T1 decay, leading to a decreased sensitivity of the experiment when compared to CASL. Another recently proposed labeling technique is velocity-selective arterial spin labelling (VS-ASL), where labelling is performed as a function of spin velocity rather than position. Only spins whose velocity is initially above a given threshold (i.e. arterial blood water spins), but subsequently drops (as it is delivered to the capillary bed) are tagged. This way, all spins traveling in small arteries and arterioles are expected to be labelled, with no constraints on spatial position.

Control of Magnetization Transfer Effects

In PASL and CASL, the labeling pulses are in off-resonance with respect to the imaging slice, as they are applied to the neck region. They can therefore give rise to MT effects, which cause a decrease in the signal of the water protons in the imaging slice and must be cancelled to avoid confounding ASL perfusion measurements. One way of avoiding MT effects is to use a small dedicated coil placed in the neck region for the labeling pulses. No MT effects will take place in the imaging slice due to the localized transmit sensitivity of the labeling coil. This method is most often used for CASL, where MT effects are particularly important due to the high power of the labeling pulses used. In most situations, methods that do not involve non-standard hardware configurations are preferred. These usually mimic the MT effects produced in the labeling process during the control image generation, so that no MT-related signal loss will be present in the subtracted images. These techniques are used both in CASL and PASL, but the remainder of this discussion will focus on PASL, as this is the ASL technique used in all the experimental work described in this thesis. One of the earliest methods devised to balance magnetization transfer effects is EPISTAR, Echo Planar Imaging and Signal Targeting with Alternating Radiofrequency. To acquire the EPISTAR control image, an inversion slab identical to the one used for the tag image acquisition is placed at the same distance from the imaging slice, but distally to it. Assuming that the power spectrum of fat is symmetric around the Larmor frequency, MT effects will be identical in both runs of the experiment. A method similar to EPISTAR where the control inversion pulse is applied non-selectively, called PICORE, Proximal Inversion with a Control for Off-Resonance Effects, has also been proposed. In another MT control scheme, the control experiment is performed applying two contiguous 180º pulses, each with half the power of the labeling pulse to the labeling slab. This method, PULSAR, Pulsed STAR Labeling of Arterial Regions, was described above under the subheading "The ASL Experiment". It is similar in rationale to TILT, Transfer Insensitive Labeling Technique, where two 90º pulses are applied in or out of phase during the tag or control image generation, respectively. Another very popular PASL technique, which is intrinsically MT balanced, is FAIR, Flow-sensitive Alternating Inversion Recovery. In FAIR, a PWI is obtained by subtracting two images acquired following an inversion recovery. One of the images is acquired following a slice-selective inversion, whereas the other follows a non-selective inversion pulse (at the same TI). Even though FAIR images have a very different contrast from the control/tag pairs in other types of ASL, the subtraction image also shows the inverted arterial blood that has reached the imaging slice and is similar to the PWI obtained using other methods.

The Buxton Model

The perfusion weighted images obtained in ASL are not CBF maps in themselves, as they are arbitrarily scaled and the intensity of the signal in each voxel is determined by a complicated interplay of CBF and arrival and dwelling times of the blood water bolus in each voxel. To obtain a CBF map from one or more PWIs, the signal in each voxel must be processed using a suitable mathematical model. These models can be derived either by modifying Bloch’s equations to account for the perfusion phenomenon or using kinetic theory. Using the kinetic methodology developed by Buxton and colleagues, the longitudinal magnetization difference, ΔM(t)=Mzcontrol(t)-Mztag(t), in a voxel in a PWI can be interpreted as the convolution (conv) of an arterial input function (AIF),which will be denoted by c(t), and a residue function, R(t), such that:

M(t)= 2*IE*M0b*f conv(c(t),R(t)) (1)

In this equation, M0b refers to the longitudinal magnetization of arterial blood, IE to the inversion efficiency of the inversion pulse and f is the temporal rate at which the inverted magnetization reaches the voxel of interest, i.e. f is the cerebral blood flow in units of inverse time. The AIF represents the amount of labelled blood that reaches the tissue, whereas the residue function describes the washout of the bolus due to T1 decay and venous outflow. In PASL, according to Buxton’s model, the arterial input function, which is created by the labeling pulse, has the ideal shape of a hat function. This perfectly well-defined bolus travels upwards in the arterial tree, “decaying” with arterial blood’s relaxation time constant, T1b, until it reaches the perfused tissue at time Δt, the bolus arrival time. The AIF can then be written as:

c(t) = exp (-t/T1b) for Δt < t < \Δt + τ (2)

and c(t) = 0 at all other times. τ is the duration of the bolus, which is related to the spatial width of the inversion pulse and the mean blood velocity at the inversion site and from there to the imaging plane. In Buxton's model, the perfused voxels are modeled by a single compartment, implying that upon reaching the perfused voxel, all labelled water spins immediately exchange with the tissue (i.e. the exchange time, Tex, is null). It is also assumed that the water spins immediately reach equilibrium with the venous system, so that the rate at which they leave the voxel due to venous outflow is f/λ, where λ is the blood-brain partition coefficient of the tissue. Under these assumptions, the residue function in Equation 1 can be written as:

R(t) = exp (-t/T1t) exp (-f t/λ) (3)

where T1t is the relaxation time constant of brain tissue (WM or GM). The bolus decays due to T1 relaxation and, usually to a lesser extent, venous outflow. The latter can be interpreted as an apparent change in the T1 of the tissue, so that R(t) = exp(-t/T1tapp), with 1/T1tapp = 1/T1t + f/λ.

Inserting Equations 2 and 3 into Equation 6.1, ΔM can be written as:

For 0 < t < Δt: ΔM (t) = 0

For Δt < t < Δt + τ: ΔM (t) = 2 M0b IE f exp(-t/T1app)*(exp(-k*Δt)-exp(-k*t))

For t > Δt + τ: ΔM (t) = 2 M0b IE f exp(-t/T1app)*exp(-k*Δt)*(1-exp(-k*τ)) (4)

where k = 1/T1b – 1/T1tapp and t=0 is the time at which the inversion pulse was applied.

It was shown that despite all the simplifying assumptions made, calculating absolute CBF values from an ASL experiment involves using a complicated model requiring the estimate of 7 parameters independent from perfusion: M0b, Δt, τ, IE, T1b, T1t and λ. These can be estimated from literature values or else separately measured. If perfusion weighted images are acquired at different delays following inversion, further increasing the duration of the ASL experiment, some of the unknowns other than CBF can also be estimated from a multi-parameter fit. In practice, this is only possible for one (usually Δt) or two (Δt and τ) additional variables.

QUIPSS-II

To allow for better CBF estimation using PWIs acquired at a single time point following the labeling, Wong and colleagues proposed in 1998 applying a saturation pulse to the distal end of the inversion pulse, at time TIsat, to sharply define the end of the bolus. If TIsat<τand TI> TIsat + Δt, it can be shown that: ΔM(TI) = 2 M0b IE f exp(-TI/T1b) (5) This method, Quantitative Imaging of Perfusion using a Single Subtraction (version II), QUIPSS-II, greatly reduces the number of unknowns needed to estimate perfusion. In theory, QUIPSS-II allows CBF to be measured accurately using a single TI. In practice, however, it is difficult to ensure that the two conditions are met for all areas of the brain, particularly in patients with abnormal hemodynamics.

Look-Locker ASL

As in an inversion recovery experiment, acquiring signal in an ASL experiment at different times following labeling involves repeating the entire experiment for each delay time, making it extremely time consuming. A variant of the ASL experiment that involves acquiring PWIs at several times following a single inversion, in an analogy with the Look-Locker method, has been proposed. In this method, the Buxton model is modified to account for the effect of previous excitation pulses on the tissue’s magnetization. By assuming that the excitation pulses only affect water molecules after they have exchanged into tissue, the tissue longitudinal relaxation time constant can be modified as in a Look-Locker experiment. In the Buxton model (Equation 4), T1tapp should be replaced by:

T1tappeff = 1/T1t + f/λ– ln (cos FA)/ΔTI (6)

and the shape of the PWI time curve shape remains similar to that of the standard Buxton model. ΔTI, the gap between excitation pulses, is usually a well-defined experimental parameter. However, FA, the excitation flip angle, can vary spatially to a great extent, affecting the uncertainty of the experimental measurement of CBF.

Look-Locker ASL has been implemented using the FAIR scheme, taking the name ITS-FAIR, Inflow Turbo Sampling FAIR or LL-EPI, Look-Locker EPI, and also using the PULSAR method, where it was renamed QUASAR, Quantitative STAR Labelling of Arterial Regions. As Look-Locker ASL allows the perfusion curve to be sampled at several time points following a single inversion, scan time can be considerably reduced, at the expense of SNR and the simplicity of the model used.

Other ASL Models

The Buxton model can be complicated further by making more realistic assumptions about the shape of the AIF and placing restrictions and a temporal delay on the exchange of water molecules between the capillaries and tissue. Two-compartment models (tissue + capillaries) are often used to model these effects, usually by including an additional parameter such as the permeability of the capillary wall to water and/or a non negligible exchange time. These models are often analytically quite complex. Finally, a model-free approach similar to the one used in DSC-MRI was recently proposed for ASL by Petersen and colleagues in 2006. In this method, the AIF is estimated from the signal in neighboring arterial voxels and R(t) is calculated directly from a deconvolution of ΔM (Equation 1). CBF can be estimated taking into account that R(0)=1. This method suffers, however, from the problems associated with the deconvolution of noisy data and can only be used if the perfusion signal is sampled at several time points following labeling.

White Matter Perfusion

As mentioned in Chapter 3, white matter is much less metabolically active than grey matter and, in healthy adults, its CBF is 3 to 4 times lower than in the cortex. This places WM perfusion at the limit of detection in a typical ASL experiment. The capillary transit time of white matter is also much longer than that of GM, implying that by the time the bolus reaches white matter tissue, its longitudinal magnetization will have almost fully recovered. This implies that measuring white matter perfusion using ASL is particularly challenging.

Modifications to the Standard ASL Experiment

Some modifications have been proposed to the ASL experiment to minimize its sensitivity and improve its robustness. Some of these modifications will now be discussed.

Vascular Crushers Perfusion measurements in ASL can be confounded by intra-arterial blood signal, which is hyperintense at early time points. As most of this intravascular signal flows through the imaging slice to perfuse more distal brain areas, it should not be taken into account for CBF estimates of the imaging slice. In Look-Locker ASL, arteries can appear even more hyperintense than static tissue, due to flow related enhancement. Although macroscopic vessels can potentially be discarded with the use of an arteriogram, this approach does not take into account the existence of partial volume effects between the arteries and brain tissue or the contribution of the arteries to the tissue signal of neighbouring voxels, due to the definition of the point spread function in MRI. A better strategy is to dephase arterial blood signal using bipolar gradients, similar to the ones employed in phase contrast angiography. These gradients, which in ASL are usually called crushers, are added between the excitation and readout modules with an encoding velocity (vENC) of about 3 cm/s. At this value, it is expected that the blood in capillaries remains virtually undisturbed, whereas signal from macroscopic flow is eliminated. Bipolar gradients are in practice only used in a single direction (usually the foot-head, FH, direction) to avoid a sizable increase in TE and the consequent drop in SNR. Signal from macroscopic in-plane flow will therefore not be eliminated.

Multi-Slice Imaging Most ASL schemes permit magnetization transfer effects to be balanced across the brain, allowing the acquisition of more than once slice. The effective number of acquired slices is limited by the temporal duration of the bolus. Multi-slice ASL usually provides incomplete brain coverage, as gaps must be left between different slices to avoid cross-talk. Ascending slice schemes are often used, as the acquisition order accompanies the bolus arrival order, allowing the perfusion curve to be sampled at approximately the same temporal position across the brain. The tagged blood reaching the upper slices may, however, be affected by the excitation of downstream slices.

Bolus Length Definition As explained above, the bolus temporal length, τ, is defined by the inversion slab's width, but cannot in practice be determined accurately as the velocity of blood at the inversion site is not known. A QUIPSS-II saturation slab applied at time TIsat can be used to define the temporal width of the bolus. If TIsat < ,  can be approximated by the time at which the saturation pulse is applied following the inversion, i.e. τ is now equal to TIsat.

Signal Pre-Saturation ASL is based on image subtraction, making it extremely sensitive to differences between control and tag images which are unrelated to perfusion. These may include subject motion (discussed in more detail later in this chapter), magnetization transfer effects and imperfect slice profiles of the control and tagged pulses. It has been found that applying a pre-saturation pulse to the imaging slice in both control and tag experiments can improve perfusion quantification, as it minimises any potential imbalance between the slice profiles and increases the fraction of the measured signal that can be attributed to perfusion.

Non-EPI Readout ASL can be performed with fast imaging techniques other than echo planar imaging. T1-FFE has been used both in the brain and perfusion studies of other organs, such as the kidney. Some studies have also reported CBF measurements using 3D-GRASE (Gradient and Spin Echo), a 3D fast imaging technique that can be regarded as a hybrid of EPI and a turbo spin echo. 3D-GRASE has the potential to provide better brain coverage than multi-slice EPI, which would in turn permit the implementation of more efficient motion correction schemes, as will be discussed below.

Vessel Selective ASL One of the most exciting possibilities of ASL is its potential to determine the perfusion territories of individual arteries, by selectively labelling the desired arteries only. An analysis of the methods used in these applications can be found in a review article by Paiva and colleagues written in 2007.

Magnetic resonance spectroscopic imaging
Magnetic resonance spectroscopic imaging (MRS) is another, NMR-based process for assessing function within the living brain. MRS takes advantage of the fact that protons (hydrogen atoms) residing in differing chemical environments depending upon the molecule they inhabit (H2O vs. protein, for example) possess slightly different resonant properties (chemical shift). For a given volume of brain (typically > 1 cubic cm), the distribution of these H resonances can be displayed as a spectrum.

The area under the peak for each resonance provides a quantitative measure of the relative abundance of that compound. The largest peak is composed of H2O. However, there are also discernible peaks for choline, creatine, N-acetylaspartate (NAA) and lactate. Fortuitously, NAA is mostly inactive within the neuron, serving as a precursor to glutamate and as storage for acetyl groups (to be used in fatty acid synthesis) — but its relative levels are a reasonable approximation of neuronal integrity and functional status. Brain diseases (schizophrenia, stroke, certain tumors, multiple sclerosis) can be characterized by the regional alteration in NAA levels when compared to healthy subjects. Creatine is used as a relative control value since its levels remain fairly constant, while choline and lactate levels have been used to evaluate brain tumors.

Diffusion tensor imaging
Diffusion tensor imaging (DTI) is a related use of MR to measure anatomical connectivity between areas. Although it is not strictly a functional imaging technique because it does not measure dynamic changes in brain function, the measures of inter-area connectivity it provides are complementary to images of cortical function provided by BOLD fMRI. White matter bundles carry functional information between brain regions. The diffusion of water molecules is hindered across the axes of these bundles, such that measurements of water diffusion can reveal information about the location of large white matter pathways. Illnesses that disrupt the normal organization or integrity of cerebral white matter (such as multiple sclerosis) have a quantitative impact on DTI measures.

fMRI and EEG
Functional MRI has high spatial resolution but relatively poor temporal resolution (of the order of several seconds). Electroencephalography (EEG) directly measures the brain's electrical activity, giving high temporal resolution (~milliseconds) but low spatial resolution. The two techniques are therefore complementary and may be used simultaneously to record brain activity.

Recording an EEG signal inside an MRI system is technically challenging. The MRI system introduces artifacts into the EEG recording by inducing currents in the EEG leads via Faraday induction. This can happen through several different mechanisms. An imaging sequence applies a series of short radiofrequency pulses which induce a signal in the EEG system. The pulses are short and relatively infrequent, so interference may be avoided by blanking (switching off) the EEG system during their transmission. Magnetic field gradients used during imaging also induce a signal, which is harder to remove as it is in a similar frequency range to the EEG signal. Current is also induced when EEG leads move inside the magnet bore (i.e. when the patient moves during the exam). Finally, pulsed blood flow in the patient in the static magnetic field also induces a signal (called a ballistocardiographic artifact), which is also within the frequency range of interest. The EEG system also affects the MRI scan. Metal in the EEG leads and electrodes can introduce susceptibility artifacts into MR images. Care must also be taken to limit currents induced in the EEG leads via the MRI RF system, which could heat the leads sufficiently to burn the subject.

Having simultaneously recorded EEG and fMRI data, the final hurdle is to co-register the two datasets, as each is reconstructed using a different algorithm, subject to different distortions (see EEG-fMRI).

Nuclear neuroimaging
Before the advent of fMRI functional neuroimaging was typically performed with positron emission tomography (PET) scanners or more rarely with SPECT scanners. Niels A. Lassen and his coworkers lead the earliest efforts of functional neuroimaging, using radioactive gases to construct images of the working brain.

These nuclear imaging techniques do not use the nuclear magnetic resonance property and employ entirely different scanners.

Approaches to fMRI data analysis
The ultimate goal of fMRI data analysis is to detect correlations between brain activation and the task the subject performs during the scan. The BOLD signature of activation is relatively weak, however, so other sources of noise in the acquired data must be carefully controlled. This means that a series of processing steps must be performed on the acquired images before the actual statistical search for task-related activation can begin.

For a typical fMRI scan, the 3D volume of the subject's head is imaged every one or two seconds, producing a few hundred to a few thousand complete images per scanning session. The nature of MRI is such that these images are acquired in Fourier transform space, so they must be transformed back to image space to be useful. Because of practical limitations of the scanner the Fourier samples are not acquired on a grid, and scanner imperfections like thermal drift and spike noise introduce additional distortions. Small motions on the part of the subject and the subject's pulse and respiration will also affect the images.

The most common situation is that the researcher uses a pulse sequence supplied by the scanner vendor, such as an echo-planar imaging (EPI) sequence that allows for relatively rapid acquisition of many images. Software in the scanner platform itself then performs the reconstruction of images from Fourier transform space. During this stage some information is lost (specifically the complex phase of the reconstructed signal). Some types of artifacts, for example spike noise, become more difficult to remove after reconstruction, but if the scanner is working well these artifacts are thought to be relatively unimportant. For pulse sequences not provided by the vendor, for example spiral EPI, reconstruction may have to be done by software running on a separate platform.

After reconstruction the output of the scanning session consists of a series of 3D images of the brain. The most common corrections performed on these images are motion correction and correction for physiological effects. Outlier correction and spatial and/or temporal filtering may also be performed. If the task performed by the subject is thought to produce bursts of activation which are short compared to the BOLD response time (on the order of 6 seconds), temporal filtering may be performed at this stage to attempt to deconvolve out the BOLD response and recover the temporal pattern of activation.

At this point the data provides a time series of samples for each voxel in the scanned volume. A variety of methods are used to correlate these voxel time series with the task in order to produce maps of task-dependent activation.

There are many software packages available for analysing fMRI data.

Reconstruction of MRI data needs to be tested, calibrated and confirmed. MRI can suffer from numerous artifacts that include, geometric distortions, Nyquist ghosting, and signal dropout. Medical Imaging Phantoms are used to provide a consistent geometrical source for calibration and testing purposes. Minute tumor changes can require recalibration by use of a phantom to quantify the change.

Cost of fMRI
The major cost of an fMRI experiment is the MR scanner itself. New 1.5 tesla scanners often cost between $1,000,000 USD and $1,500,000 USD. New 3.0 tesla scanners often cost between $2,000,000 and $2,300,000 USD. Construction of MRI suites can cost $500,000 USD.

MRI procedures themselves can vary considerably in cost but generally fall somewhere between $400 and $3,500, depending on the facility and which region of the body is being scanned. Extremity scans (feet, hands, etc.) tend to be lower in price while body scans (including the brain) tend to be higher.

Commercial use
Most fMRI scans are for research or clinical use. Commercial use is limited. However, a few companies have been set up that attempt to sell fMRI specific hardware or services for research or clinical use.

At least two companies have been set up to use fMRI in lie detection (No Lie MRI, Inc and Cephos Corporation ).

In using fMRI techniques for use in lie detection, activated areas of the brain are observed while the subject is making a statement. Depending on what regions are the most active, the technician might determine whether a subject is telling the truth or not. Since a specific combination of brain functions are needed in order to tell a lie, the simultaneous activation of these regions often indicates deception. This technology is in its early stages of development, and many of its proponents hope to replace older lie detection techniques.

In clinical trials, the usage of fMRI as a method of lie detection has appeared reliable, with studies from 2005 by Kozel et al. indicating a 90% to 93% success rate.

However, there is still a fair amount of controversy over whether these techniques are reliable enough to be used in a legal setting. Some studies indicate that while there is an overall positive correlation, there is a great deal of variation between findings and in some cases considerable difficulty in replicating the findings.

Textbooks

 * Scott A. Huettel, Allen W. Song, Gregory McCarthy, Functional Magnetic Resonance Imaging, Sinauer Associates, 2004, ISBN 0-87893-288-7
 * Richard B. Buxton, An Introduction to Functional Magnetic Resonance Imaging: Principles and Techniques, Cambridge Univ Press, 2002, ISBN 0-521-58113-3
 * Roberto Cabeza and Alan Kingstone, Editors, Handbook of Functional Neuroimaging of Cognition, Second Edition, MIT Press, 2006, ISBN 0-262-03344-5