1　A very brief review on the evolution of functional brain imaging leading up to fMRI

       Information processing within the brain depends upon the electrical activity of neurons. To assess changes in brain function, we must use a measure that reflects neuronal activity. One possible approach is to measure this activity directly, using electrodes that can measure changes in voltage in neurons’ axons or dendrites. However, implanting electrodes inside the brain is obviously an invasive procedure, so it has been limited in animal research or in a small patient population with clinical needs for the electrodes. For most human electrophysiological studies, therefore, electrical or electromagnetic activity is measured using detectors outside the surface of the scalp, but obviously with limited spatial resolution.

       Given the limitations of direct electrical measurement, scientists have adopted alternative approaches to indirectly measure neuronal activity, for example, through its metabolic correlates. Positron Emission Tomography (PET), which became prominent in the 1980s and was the first imaging technique that could be used in normal human subjects, measures the local quantities of radioactively labeled glucose, oxygen, or other metabolites. Though not requiring implantation of electrodes, PET is still considered an invasive technique because compounds that emit ionizing radiation must be injected into the body before each scan.

       The demand to measure metabolic changes in the brain without any invasive procedures lends motivation to develop MRI methodology to image brain activities. Serendipitously, and as early as in 1936, Linus Pauling and Charles Coryell^[1] found that diamagnetic oxygenated hemoglobin would become paramagnetic when deoxygenated, resulting in significant magnetic field inhomogeneity in its vicinity. This reduces T2^* values, as demonstrated decades later by Thulborn and colleagues^[2]. Thus, it was postulated that, if blood oxygenation varies spatially according to brain function, the brain activation induced magnetic field variation should be measurable using MRI. This work was greatly expanded by Seiji Ogawa and colleagues using animal models at high field MRI^[3]. They found that there was significant signal loss spatially correlated to the presence of deoxyhemoglobin, which was subsequently named as the blood-oxygenation-level dependent (BOLD) contrast to enable measurement of functional changes in brain activity. Continued and rapidly expanding effort in the ensuing years have led to the development of task-activated fMRI^[4,5,6], event-related fMRI^[7,8], resting state connectivity^[9,10], and their vast applications in clinical and basic science. A more detailed understanding of fMRI principles and applications can be found in the textbook by Huettel et al^[11].

2　Review on the current fMRI methodologies

       While fMRI is maturing as a technique, several important limitations remain. The need to address these limitations has been the primary motivation in the development of innovative methodologies. Specifically, these novel methods are aimed to meet the demand for increased spatial and temporal resolutions, the need to maximally extract useful information from the fMRI time series, and the requirement for improved fidelity of our fMRI data to the underlying neuronal activity. The following discussions are centered on these areas of interest.

2.1　Acquisition methodology: meeting the demand for high spatial and temporal resolution

       Since the very beginning, the pursuit of spatial and temporal resolution has been the main driving force behind the many technological breakthroughs in MRI, and fMRI is no exception. The advent of high-field scanners, massive parallel imaging techniques, and innovative pulse sequence developments has pushed the spatial and temporal resolutions to tens of milliseconds and hundreds of micrometers. Indeed, it is now possible to visualize real-time brain activities in different cortical layers and sub-nuclei in deep gray matters.

       Improving the spatial and temporal resolution of fMRI will require more than just collecting smaller voxels and sampling the MR signal more frequently. While such improvements are likely, especially with advances in hardware quality, real improvements in resolution will require fundamental changes in data collection techniques. Images must be collected using pulse sequences that more closely map fMRI signal changes to neuronal activity, both in space and time. Distortions caused by magnetic inhomogeneities must be corrected. And, as will be discussed later in this chapter, additional forms of contrast must be developed and validated to overcome key limitations of BOLD fMRI.

2.1.1　High field imaging

       High magnetic field leads to high SNR, which can be subsequently used to increase imaging speed and spatial resolution. In fact, since fMRI’s beginnings some fifteen years ago, the field strength at which images have been collected has steadily increased. The 1.5 T scanners on which the original human research was carried out is no longer used for fMRI research, the then high-field 3.0 T scanners are the standard workhorse machines for image acquisition. Increasingly, many institutions are installing human MRI scanners at field strengths of 7.0 T, 9.0 T and even 11.7 T.

       The primary advantage of high field scanning comes from increased signal. As the scanner field strength increases, a greater proportion of spins will align parallel with the static field, and thus net magnetization will increase. This increase in raw MR signal is quadratic with field strength, that is, as field strength doubles the raw signal recorded by the scanner quadruples. Mitigating this advantage, however, is a corresponding increase in thermal noise, which increases linearly with field strength. The signal to noise ratio (or SNR) thus increases linearly (i.e., a quadratic increase divided by a linear increase equals a linear increase) with increasing field strength. In principle, the increased SNR can allow the image to be parceled into smaller voxels while maintaining sufficient SNR within each, improving spatial resolution. For example, studies of ocular dominance columns, which are approximately one millimeter in width, have been carried out at 4T^[12].

       It is important to recognize that the quality of fMRI data depends not just upon net magnetization and thermal noise, it also depends upon physiological and neural variability, which are collectively called physiological noise. The strength of the static field has different effects upon thermal noise and physiological noise. While thermal noise increases linearly with increasing field strength, physiological noise increases quadratically with field strength^[13]. So, as field strength increases from 1.5 T to 3.0 T, raw signal will quadruple, thermal noise will double, and physiological noise will also quadruple. Thus, at very high field strengths, physiological noise may become dominant and thus the improvements associated with increasing field strength may be considerably less than linear (i.e., smaller than for raw SNR). Although there have been theoretical suggestions that there may be an asymptotic upper limit for effective field strength in fMRI^[13], those suggestions have yet to be tested due to the current paucity of very high field scanners. Regardless of these arguments, a concrete benefit for fMRI at high fields may be the increased spatial as well as temporal resolution as a direct result of greatly increased SNR.

       It would certainly not be complete to discuss high-field fMRI without mentioning its challenges. Aside from its high cost, there is also mounting technical difficulties: the often extreme non-uniformity of the magnetic field, high every deposition (in the form of specific absorption ratio or SAR), greatly increased susceptibility artifact. Encouragingly, recent technical advances such as parallel transmit methodology can large addresses the transmit field non-uniformity and SAR issues^[14,15], while improved local shimming and innovative pulse sequences can overcome susceptibility-induced image distortions and signal losses^[16,17,18] .

2.1.2　Parallel imaging with SENSE acceleration

       In addition to or in concert with the use of high field MRI scanners, one effective and less costly way to increase spatial and temporal resolution is to collect more data per unit time, using a technique known as parallel imaging. In parallel imaging, several detector coils are used to simultaneously sample the brain to improve spatial or temporal resolution.

       To understand how parallel imaging can improve spatial resolution, assume that normally a single coil is used to acquire a single image. Using four coils, one could collect four images with overlapping fields of view and larger coverage in k-space to reach higher resolution. The k-space data are sub-sampled for individual coils, resulting in severely aliased images. These aliased images must be corrected either in image space (in the case of SENSE, or sensitivity encoding) or in k-space (in the case of SMASH, or simultaneous acquisition of spatial harmonics), restoring original image quality^[19,20]. By incorporating sensitivity maps from individual coils and using an iterative reconstruction process to remove aliasing artifacts, a final image with uniform spatial coverage and high spatial resolution can be achieved. The relation between the number of coils and the matrix size can be expressed more generally in the following way. Assuming the number of receiver coils is M, and the number of voxels desired in the final reconstructed image is N², the matrix size could increases by a theoretical factor up to M for a given acquisition time. Although in practice, the acceleration factor is usually much smaller. Likewise, parallel imaging can have salutary effects upon temporal resolution. The temporal resolution decreases proportionally to the number of receiver coils used; that is, by a factor, theoretically, up to M. Fig. 1 illustrates an example of high-resolution fMRI activation of motor cortex, using a T1-weighted inversion-prepared echo-planar imaging with a SENSE acceleration factor of 2 and an 8-channel head coil array on a 3T MRI scanner (GE Healthcare, Waukesha, WI). With the improved hardware such as a 32-channel array, the spatial and temporal resolution can be further increased.

点击查看大图

Fig. 1  A high-resolution fMRI motor activation map (with t-scores) during a finger-tapping task. An inversion-recovery EPI pulse sequence with SENSE acceleration is used at a spatial resolution of 1.5 mm×1.5 mm×1.5 mm.
Fig. 2  A schematic illustration of the CAIPIRHINA type of simultaneous multi-slice imaging with half FOV shift in slice 1 and 3 to improve the robustness of slice separation (from image on the right to slices below) using parallel image reconstruction along slice selection direction.

2.1.3　Multi-band imaging

       One recent development in parallel imaging is focused on simultaneous multi-slice acquisition, now generally known as multi-band imaging. Compared to the more traditional in-plane acceleration (e.g. SENSE), multi-band imaging has the distinct advantage of gaining multiple folds of imaging speed without sacrificing the SNR at the rate equals to the square root of the acceleration factor. This is because a modulated RF excitation pulse is used to excite multiple slices simultaneously and the aliased slices are then separated using the coil sensitivity profiles along the slice selection direction^[21,22] .

       Since the robustness of parallel imaging reconstruction to remove aliasing effect is dependent on the difference between the coil sensitivity profiles, the individual elements in the coil array need to be distributed along the slice selection, and the gap between the simultaneously excited slices should not be too small. Indeed, it becomes highly challenging to separate the aliased slices when the simultaneously excited slices are closely packed. A technique termed CAIPIRHINA (Controlled aliasing in parallel imaging results in higher acceleration) was developed for improved ability to disentangle the aliased slices (Breuer). In CAIPIRHINA, phase-cycled RF excitation was used in that a 180° phase shift was introduced between adjacent k-space lines in every other slice, leading to a half-FOV shift in alternating slices. This operation takes advantage of the empty space along the y (phase encoding) direction, thus greatly improving the robustness to separate overlapping slices, as shown in Fig. 2.

2.2　Analysis methodology: meeting the demand to extract the most information from the fMRI time-course data

       Although improvements in scanner hardware and imaging technology are important, they are not complete to fully address the many challenges faced by fMRI studies. Even more importantly, with the shear amount of the data we acquire at the present time, the ability to design experimental analysis strategies to get the most information on the underlying neuronal events out of the massive dataset is extremely valuable, which calls for sophisticated experimental design and data interpretation techniques.

2.2.1　Improved experimental design

       Recent technical developments have made it possible to greatly shorten the TR to increase the imaging speed, in fMRI the actual temporal resolution is limited by the intrinsic delay and dispersion in the hemodynamic response. That is, the temporal uncertainty introduced by the fMRI hemodynamic response function will undermine any attempt to make millisecond-level assessments of the underlying neuronal activity. To overcome this problem in the study of one brain system, Ogawa and colleagues have developed a technique that induces either excitatory or inhibitory neural interactions by presenting consecutive stimuli (to the forepaw of a rat) separated by very short intervals, on the order of tens of milliseconds. By manipulating the interstimulus interval, the research team was able to detect the inhibition due to neuronal refractoriness on the response to the second stimulus^[23]. Such an inhibition effect is subsequently manifested in the BOLD signal. With such a clever manipulation, one can achieve improved temporal resolution without physically increasing the sampling interval, as neuronal interactions on the order of milliseconds can lead to very different hemodynamic responses.

       Along the same line, a number of groups have used similar approaches of analyzing the effects of previous stimuli, known as refractoriness, upon the characteristics of the f MRI hemodynamic response. Some early studies demonstrated that the amplitude of the hemodynamic response was roughly proportional to the duration and number of presented stimuli^[24,25]. That is, given the hemodynamic response of a 6 s duration stimulus, one could estimate the response to a 12 s stimulus by adding together two 6s responses^[24]. However, when the separation between stimuli is shorter, there was clear evidence that this simple additive feature was not entirely correct, and a number of subsequent studies confirmed that the fMRI hemodynamic response attenuates with repeated activation of a brain region^{[26,27,28,29]} .

       Indeed, this refractory effect has turned out to be important for a new class of experimental designs that use stimulus adaptation. Studies by Grill-Spector, Malach, and colleagues have investigated whether individual regions of visual cortex are sensitive to changes in higher-order properties of objects^[30]. In which they compared two types of conditions: one in which the same stimulus is presented repeatedly within a block, and another where some aspect of the stimulus changes continually over time. For example, an "identical face" block might show the same view of the same person’s face over and over, while a "size changing" block might present the same face in a variety of sizes on the display. They found that activity in the fusiform gyrus, for example, was greatly reduced to repeated presentations of identical, size-varying, or position-varying faces, compared to repeated presentations of different people’s faces. In contrast, variation in direction of illumination or viewpoint caused a recovery from adaptation. From these results, they concluded that the fusiform gyrus recognizes facial identity over size or position manipulations, but not over illumination or viewpoint manipulations. This approach has since been used by other groups and with other stimulus domains to investigate the attribute specificity of the fMRI hemodynamic response^[31,32,33].

       In summary, it is noted that new classes of experimental designs take into account the dynamic properties of fMRI time courses. By incorporating both blocked and event-related approaches, and staggering neuronal events with various temporal delays, we can gain substantial improvements in both detection power and estimation efficiency^[34,35,36], allowing characterization of mutual interactions among multiple neuronal events, and estimation of the fMRI time course beyond the hemodynamic time limit.

2.2.2　Data interpretation techniques

       Most of the image analysis methods use variants of the general linear model to test specific hypotheses about brain function. These hypothesis-driven analyses have accounted for much of the growth of fMRI over the past decade, but they cannot address some potentially important research questions. In this section, we will discuss several important and novel techniques for fMRI data analysis that go well beyond the hypothesis-driven general linear model. These techniques borrow technical concepts from many fields, including statistics and mathematics, economics and other social sciences, engineering, and computer science, and are used for boost the power in exploration and prediction. Some techniques explore fMRI data in search of systematic variation, without necessarily adopting an a priori model for that variation. Other techniques reverse the traditional direction of fMRI analyses, in that they use fMRI data to predict variations in behavior, perception, or cognition. These data-driven analyses often use mathematical algorithms that partition the four-dimensional fMRI time series into a set of components that may reflect distinct aspects of brain functioning. Data-driven analyses may identify regularities that are clearly task-related, or they may discover task-unrelated variability that can be eliminated during preprocessing. These analyses are not subject to some of the problems that compromise hypothesis-driven analyses: poorly chosen models, unknown timing of neuronal activity, and variability in the hemodynamic response.

       Here we first consider several techniques that allow researchers to explore their data for potentially meaningful variation——all of which parse the fMRI time series into sets of common components, that consist of groups of voxels and their temporal properties. Importantly, their algorithms attempt to identify acti-vation that is common to a group of voxels, rather than compare the activation of individual voxels with a hypothesized time course. The components are evaluated based on how much of the variation in the fMRI data they explain. These techniques vary both in the algorithms that they use to parse the time series data and, more importantly, in the rules they use to decide on which features are meaningful. Some are completely model-free, in that they extract features without any regard to the underlying experimental paradigm. Others use some information about the task to shape how the components are extracted. These techniques can also differ somewhat in their goals; some are used to identify interesting forms of variability in order to suggest future analyses, while others are used to test simple hypotheses. In recent years, some of these approaches are now also being used to identify uninteresting and non-task-related variability (e.g., for preprocessing). Typical techniques of this kind include principal component analysis (PCA) or independent component analysis (ICA)^[37,38,39].

       Researchers today are certainly not satisfied by only the explorative aspect of the data analysis, indeed many have moved on to harness the predicative power of fMRI analysis. Prediction approaches reverse the typical direction of inference in fMRI research. The regression analyses that we have been used to in the past explicitly treat task events as independent variables (i.e., forming regressors in a design matrix) and the observed fMRI data as dependent variables. In contrast, prediction approaches treat the fMRI data as independent variables that can predict some aspect of behavior, such as subjective experiences (e.g., looking at attractive images), subsequent memory (e.g., remembering a story two weeks later), and simple purchasing decisions (e.g., whether to buy a new iPhone). Naturally, there is a variety of prediction approaches. Some studies label events according to subjects’ behavior, not just the stimulus properties, and these can be analyzed using the traditional regression methods. Others modify the general linear model to include some continuous, behaviorally defined predictor, either across subjects or across trials. Still others attempt to make inferences about the interactions between activity and some behavioral factor, integrating the connectivity methods described above with behavioral methods. Finally, some approaches use novel computational methods that use the joint changes in activation across sets of voxels to make their predictions.

       To summarize, researchers now widely use advanced data analysis approaches, be it hypothesis driven or data driven, exploration or prediction, to investigate many aspects of thought and behavior, including susceptibility to psychiatric disorders, personality traits, complex economic decisions, and even conscious awareness of stimuli. Some techniques do not even fit neatly into any of the aforementioned categories. For example, we will also consider techniques for measuring the connectivity between brain regions, these share both exploration and prediction as goals. In many ways, these new analysis techniques represent a large part of the future fMRI, perhaps more so than advanced acquisition methodologies do.

2.2.3　Brain connectivity and connectome analysis

       In most conventional fMRI studies, the brain regions involved in the behavioral /cognitive performance of interest are identified through mathematically comparing the voxel-wise time-course profiles with the experimental behavioral / cognitive paradigm, reflecting the functional segregation emphasized by the localizationism in neuroscience research^[40]. In contrast to localizationism, the connectionism emphasizes more on how multiple brain regions are functionally connected, through neuronal networks, even when the subjects are not actively participating in a specific behavioral / cognitive task (e.g., in resting-state fMRI scans). In the past few years, an increasingly more research has been performed to study the brain connectivity. In comparison to the voxel-based analysis, the brain connectivity analysis better reveals the global network properties, such as the global efficiency of neuronal networks^[41]. Furthermore, recent studies demonstrated that the resting-state fMRI based connectivity network mapping may serve as an imaging biomarker^[42], revealing the mechanistic basis of neurological diseases that target distinct large-scale brain networks.

       The functional connectivity of brain networks has been studied with different modalities (such as multiunit recording^[43], PET (Friston 1993)^[44] and fMRI (Biswal 1995)^[9], based on the temporal correlations of signals measured from distant brain areas. In 1995 Biswal et al. first showed that, based on a temporal coherence of activity in the low frequency (<0.08 Hz) component of the BOLD signal, the sensorimotor networks can be identified from resting-state fMRI data. Further studies demonstrated that the low-frequency signal components, although not attributable to specific task or external stimuli, contain valuable information regarding brain activity in multiple neuronal networks, such as visual, auditory, task-negative/default mode, hippocampus or episodic memory, language, dorsal attention and ventral attention systems^[45].

       Mapping of intrinsic connectivity networks with resting-state fMRI does not require subjects’ active engagement in cognitive or neuropsychological tests, and thus can be widely applied to imaging different patient populations^[45,46], even anesthetized individuals^[47]. Recent studies have produced very encouraging results, illustrating that the neurological diseases can be characterized based on patterns of intrinsic connectivity networks, measured by resting-state fMRI^[42]. It has also been demonstrated that the connectivity measures derived from resting-state fMRI are quantitatively reproducible^[48], making the brain connectivity mapping suitable for longitudinal studies (e.g., monitoring the disease progression in clinical trials). It should be noted that, even though functional connectivity studies based on spontaneous signal fluctuations are frequently performed during a continuous resting state, the spontaneous BOLD signal fluctuations can also be measured from off-task periods in block-design fMRI, as well as from event-related fMRI after suppressing the task evoked signal changes^[49,50] .

       A major research direction that has been emerging in the past few years is the mapping of the connectome^[51]. Inspired by the genome projects, multiple connectome projects have been conducted to map the structural and functional connectivity of brains. For example, an ongoing human connectome project^[52,53] aims to measure the anatomical and functional connectivity, using diffusion tensor imaging (DTI) and resting-state fMRI, respectively, among all cortical areas and subcortical structures of human brains in vivo. It is expected that the brain connectivity and connectome mapping will continue to play a major role in basic neuroscience research in the next few years.

2.3　Contrast mechanisms: meeting the demand for improved fidelity of the fMRI signal to the underlying neuronal activities

       To date most of the fMRI studies assume that the BOLD signal is an accurate representation of neuronal activity. Indeed there is substantial empirical evidence demonstrating that neuronal activity is associated with the sorts of physiological changes necessary to evoke a BOLD response-thousands of fMRI studies have provided operational evidence that matches previous electrophysiological evidence. Yet, it was generally accepted that BOLD signal is not neuronal activity itself, and is rather the result of increased local oxygenation arising form the uncoupling of blood flow and oxygen metabolism^[54], although its exact underlying neuronal mechanism has been under debate and investigation^[55,56,57] .

       One simple example is on the timing of the BOLD response: discrepancies in BOLD time course at different brain regions cannot be simply attributed to differences in neuronal timing or lags in vascular delivery. More sophisticated questions arise on the basis of neuronal signal——is the BOLD signal related to the synchrony of neuronal firing or the total spike activity? These are actual challenges in front of many fMRI researchers today. In this section, we evaluate recent research that investigates the relation between BOLD and neuronal activity, and we continue our discussion of ways to improve the fidelity of the fMRI signal in light of some recent technical development.

2.3.1　The complex relationship between the Bold signal and the underlying neuronal activity

       It is conceivable that simple experiments can be carried out to compare and validate the relation between neuronal and BOLD activity. However, such comparisons have been difficult to make in front of the complex neuronal system and the daunting technical challenges. In one of the most comprehensive studies, Logothetis and colleagues recorded both fMRI and electrophysiological data simultaneously within the primary visual cortex of the monkey^[58]. The monkeys viewed a rotating visual checkerboard stimulus while being scanned in a 4.7-T scanner using gradient-echo echo-planar imaging, during which time concurrent single-unit, multi-unit, and local field potential data were recorded. When the visual stimulus was presented, there was a transient increase in BOLD signal that then persisted until its offset. This pattern of activity was also present in the local field potential data (i.e., the summation of integrative activity in dendrites), but was only weakly or not at all evident in the multi-or single-unit activity (i.e., the axonal firings of individual neurons). Converging evidence has been obtained by studies of cerebellar cortex that compared multi-unit activity, local field potentials, and blood flow via laser Doppler^[59]. Based on these findings, one would conclude that the BOLD contrast mechanism seems to reflect the dendritic inputs and intracortical processing in a given area, rather than its axonal outputs associated with action potential firing. This is consistent with analyses of the energy budget of the brain and how that budget is met through increased vascular supply of metabolites. Because the primate brain has a much higher synaptic density, it has been estimated that dendritic activity consumes about three-quarters of its energy^[60], and thus by inference the BOLD activity within a region should reflect the amount and timing of dendritic activity.

       Further studies have been recently conducted in humans to compare the BOLD fMRI and electrophysiological measures in human subjects^[61]. Changes in local field potentials were measured in patients awaiting neurosurgery who had implanted electrodes, while BOLD changes were measured in neurologically normal volunteers. As in the Logothetis study described above, there was a strong correspondence between the amplitude and duration of local field potentials and the duration of the BOLD response in brain regions around primary visual cortex. However, in other visual cortical regions, this relation was less clear. In the lateral occipitotemporal cortex, neither the field potentials nor the BOLD response were systematically altered by stimulus duration. In contrast, within the fusiform gyrus, there was a clear dissociation between the measures: as stimulus duration increased, the BOLD response increased but field potential activity did not. Such results suggest that other aspects of neuronal activity, such as the synchronicity across neurons within a region, may contribute to the BOLD response. Future studies directly comparing fMRI and electrophysiological measures will be necessary to extend the results discussed in this section to additional brain regions, stimulus conditions, and subject populations.

2.3.2　Multi-modal imaging and non-Bold contrasts to improve the neuronal relevance

       Given the ongoing debate on the complex neuronal and hemodynamic origins of the BOLD contrast, one begs the question whether it is possible to develop more direct MRI measures of neuronal activities.

       One natural solution to improve the neuronal relevance of the BOLD signal is multi-modal imaging integrating electrical recordings [e.g. electroencephalography (EEG)], event-related potential (ERP) with fMRI. If we know the precise spatial configuration and strength of the equivalent dipoles active at a particular instant, then it is possible to compute the exact scalp distribution of potential that would result. Furthermore, we also established that the temporal sequence of neuronal events, even those separated by mere milliseconds, changes the scalp distribution of the electric fields in predictable ways. If fMRI can tell us the exact location of all brain regions activated by a stimulus, and if we know the orientation of the active neurons in those regions, and if we know the precise time course of activity in those regions, then we should be able to calculate the resulting scalp electric fields using forward solution techniques. On the other hand, time courses from electrical recordings can serve as dependent variables to guide the fMRI signal analysis, thereby improving its neuronal relevance.

       Another form of multi-modal imaging is concerned with transcranial magnetic stimulation (TMS). It differs from EEG, which is focused on cortical recording, in that TMS offers a relatively non-invasive means for cortical stimulation. TMS generates a strong magnetic field, typically lasting less than a millisecond, that extends through the skull and into the brain. This produces an electrical current by electromagnetic induction, and in turn disrupts neuronal processes within its reach. One of the most promising uses of TMS has been to determine whether regions activated in fMRI studies are essential for task performance. Taken with the fMRI results, TMS data can provide important converging evidence that a given brain region is a critical in performance of a certain task (e.g. timing, accuracy). While combined or parallel use of TMS and fMRI is still at an early stage, such studies will become increasingly important in coming years.

       Aside from relying on other modalities for convergent evidence, effort has been made within the fMRI discipline to improve its neuronal relevance. Some of these approaches are centered on simplifying the sources of hemodynamic signal by focusing exclusively on the blood flow or blood volume using MRI. Three such techniques are perfusion imaging, vascular-space-occupancy (VASO) imaging, and diffusion-weighted imaging. By using arterial spin labeling (ASL), images can be collected that are sensitive to blood flow from upstream arterial networks into the local vascular system around an area of function. This flow is known as perfusion. Perfusion imaging is sensitive to capillary activity, where oxygen exchange takes place, but is less sensitive to changes in large veins because of T1 recovery effects. Therefore, optimized perfusion contrast is believed to have better functional resolution than BOLD contrast^[62,63]. In VASO imaging, signal from within blood vessels themselves is eliminated, so that images are only sensitive to changes in extravascular signal^[64]. Such images can be used to measure blood volume changes (without having to inject contrast agents), which are greater in smaller vessels like capillaries than in larger vessels like veins. A third approach, diffusion-weighted imaging, collects images that are sensitive to the amount of motion of spins across space. As spins (i.e., protons within water molecules) diffuse more quickly through large vessels, diffusion weighting can be used to selectively attenuate large vessel contributions to the BOLD signal——a principle that was also used in VASO imaging. In addition, measurement of differences in the apparent diffusion coefficient (ADC) across space can provide information about vascular (including capillary) source distribution^[65] .

       One of the most recent advances in fMRI is among the most intriguing: the use of MRI to directly image neuronal activity. This has been and will remain extremely challenging, due to the transience, weakness, and inhomogeneity of the electrical activity of neurons. As an example, an early study used spin-echo imaging to measure changes in spin phase associated with the field perturbations of electrical currents in human nerves^[66]. However, these results have not been confirmed by further studies, and the application of this technique to neurons in the brain may be more challenging than to peripheral nerves. Recognizing that rapidly changing magnetic fields may indicate electrical activity of neurons, Bodurka and Bandettini attempted to selectively detect rapidly changing magnetic fields while suppressing slowly changing magnetic fields^[67]. The initial concept was tested on a phantom with implanted wires. The timing of transient currents in the wires was modulated relative to a 180° excitation pulse. Very small phase differences were detected, demonstrating the feasibility of the approach. To illustrate the potential importance of this effect, the magnetic field changes measured were as short as only 40ms in duration and as small as only 2 × 10^-10 T (200 pT). These magnetic changes are about ten billion times smaller than typical static field strengths, illustrating the profound technical requirements of direct neuronal imaging. Alternative approach involves the use of signal amplification, such as the contrast mechanism based on Lorentz forces, which arise from the movement of charged particles (e.g. ionic flow) through a magnetic field. This force acts to displace the particles, and through amplification of synchronized gradient oscillations, the surrounding water molecules could lose sufficient phase coherence and lead to detectable signal. Proof-of-concept results have been reported in phantoms and in vivo^[68], and similar effects may be possible to be observed in the brain.

       Despite the extraordinary difficulties in direct MRI of neuronal electrical activity, it is still perhaps one of the most important areas for future fMRI research. Continued effort is being made that hopefully one day will provide a direct and non-invasive measure of neuroelectric activity at high spatial and temporal resolution.

3　Summary

       The continued development of new methods represents one of the most exciting aspects of the field of fMRI. Our discussions here on some of the representative technological developments are only a partial reflection of this exciting field. Many other methods, although not mentioned here, could have significant applications in fMRI. For example, compressed sensing^[69] could greatly increase the spatial and temporal resolutions for fMRI. In the coming decade, we are confident that great progress will be made to push the boundaries of current fMRI practice. It can be anticipated that, with advanced methodologies, fMRI will enjoy ever expanding applications in a wide range of disciplines ranging from clinical and basic neuroscience, social and political sciences, and even economics, permeating the scientific communities as well as our daily lives.