Functional magnetic resonance imaging of the human brain: data acquisition and analysis

Exp Brain Res (1998) 123:5±12 Springer-Verlag 1998 Robert Turner Alastair Howseman Geraint E. Rees Oliver Josephs Karl Friston Functional magnetic resonance imaging of the human brain: data acquisition and analysis R. Turner ()) A. Howseman G.E. Rees O. Josephs K. Friston The Wellcome Department of Cognitive Neurology, Institute of Neurology, Queen Square, London WC1N 3BG, UK Abstract It is now feasible to create spatial maps of activity in the human brain completely non-invasively using magnetic resonance imaging. Magnetic resonance imaging (MRI) images in which the spin magnetization is refocussed by gradient switching are sensitive to local changes in magnetic susceptibility, which can occur when the oxygenation state of blood changes. Cortical neural activity causes increases in blood flow, which usually result in changes in blood oxygenation. Hence changes of image intensity can be observed, given rise to the socalled Blood Oxygenation Level Dependent (BOLD) contrast technique. Use of echo-planar imaging methods (EPI) allows the monitoring over the entire brain of such changes in real time. A temporal resolution of 1±3 s, and a spatial resolution of 2 mm in-plane, can thus be obtained. Generally in a brain mapping experiment hundred of brain image volumes are acquired at repeat times of 1±6 s, while brain tasks are performed. The data are transformed into statistical maps of image difference, using the technique known as statistical parametric mapping (SPM). This method, based on robust multilinear regression techniques, has become the method of reference for analysis of positron emission tomography (PET) image data. The special characteristics of functional MRI data require some modification of SPM algorithms and strategies, and the MRI data must be gaussianized in time and space to conform to the assumptions of the statistics of Gaussian random fields. The steps of analysis comprise: removal of head movement effects, spatial smoothing, and statistical interference, which includes temporal smoothing and removal by fitting of temporal variations slower than the experimental paradigm. By these means, activation maps can be generated with great flexibility and statistical power, giving probability estimates for activated brain regions based on intensity or spatial extent, or both combined. Recent studies have shown that patterns of activation obtained in human brain for a given stimulus are independent of the order and spatial orientation with which MRI images are acquired, and hence that inflow effects are not important for EPI data with a TR much longer than T1. Key words Magnetic resonance imaging Brain mapping Statistical parametric mapping Echo-planar imaging methods Introduction Magnetic resonance imaging (MRI) techniques have been adapted in recent years to allow formation of brain images that are sensitive to local changes in blood flow. In identification of cortical areas active for a particular brain task, the techniques permit a spatial resolution of 2 mm, a temporal resolution of 1 s, and the opportunity to rescan a single subject as often as desired. Whilst the underyling phenomenon is simple, in the sense that many MRI scanners were found to have the capability of performing functional MRI (fmri) studies once the technique had been discovered, quantification of the effect in relation to cortical electrical activity has been elusive, and the origins of the functional signal require some explanation. The decay of signal associated with local magnetic field inhomogeneities, called T2* relaxation, was previously considered a nuisance and represented a limitation upon MRI. To mitigate this, either the so-called spin-echo technique was used, in which a second `refocussing' radiofrequency pulse following the initial excitation pulse removed the effects of dephasing, or the time between the radiofrequency excitation pulse and the acquisition of signal was reduced a much as possible, as in FLASH (Fast Low-Angle SHot imaging (Haase et al. 1986). It was only when it was realized that the presence of a paramagnetic substance in the bloodstream could act as a vascular marker, giving useful contrast, that sequences without a refocussing pulse, with a relatively long time between the excitation pulse and data acquisition (20± 80 ms), began to be used. Initially the paramagnetic con-

6 trast agent was exogenous, a non-toxic compound of gadolinium introduced into the bloodstream via a leg vein. A fraction of a millimole of contrast agent per kilogram of body weight is sufficient to give a loss of signal from tissue surrounding cerebral blood vessels of perhaps 40% as a bolus of contrast agent passes through. The time integral of the signal attenuation during the first pass of the contrast agent through the brain gives a quite accurate estimate of the cerebral blood volume (CBV) and the Central Volume Theorem (Stewart 1894) can be used to obtain local cerebral blood flow as the ratio of CBV to mean transit time. Early studies (Villringer et al. 1986; Belliveau et al. 1988) examined the passage of contrast agent through the brains of rats and later dogs, making increasing use of echo-planar imaging (EPI) (Mansfield 1977), which acquires a complete image in less than 100 ms and thus allows ªsnapshotsº of the contrast agent distribution as it passes rapidly through the brain. The technique was applied eventually to functional activation studies in humans (Belliveau et al. 1991). The subjects were given visual (ªphoticº) stimulation while a bolus of contrast agent was injected into a leg vein, and single-slice images of the brain in the plane of the calcarine fissure were obtained at 0.75-s intervals to monitor the bolus passage. By integrating the time course of image intensity, estimates of relative blood volume were obtained, and compared (by image subtraction), with those obtained when the subjects were at rest in darkness. Consistent increases (up to 30%) of blood volume in primary V1 visual cortex were observed. Subsequent developments rapidly overtook this pioneering work. Working independently, Ogawa et al. 1990) and Turner et al. (1991) showed in laboratory animals that similar changes in MRI image contrast extending around the blood vessels could be obtained simply by changing the oxygenation state of the blood. This comes about because deoxyhaemoglobin is more paramagnetic than oxyhaemoglobin (observed by Pauling and Coryell 1936; which itself has almost exactly the same magnetic susceptibility as tissue. Thus deoxyhaemoglobin can be seen as nature's own contrast agent. Interventions to the state of the brain that create an imbalance between oxygen uptake and blood flow will thus inevitably cause a change in MRI signal around the cortical vessels, if MRI sequences are used that are sensitive to magnetic field inhomogeneity. This development culminated in the work of Kwong et al. (1992) and Ogawa et al. (1992) who succeeded in showing that the change in deoxyhaemoglobin in human visual cortex, while the subject viewed a bright light, was sufficient to cause measurable changes in gradientecho MRI images of a slice passing through the calcarine fissure. The technique was dubbed ªBlood Oxygenation Level Dependent (BOLD) contrastº. Thus the way was opened to functional mapping studies of the human brain without use of contrast agent, no radiation dose, and with the high spatial resolution of MRI. Imaging techniques For human brain activation studies it is highly desirable to obtain image data quickly. There are several reasons for this, apart from the obvious need to avoid experiments lasting many hours. The first is that many perceptual and cognitive tasks of interest can be continued only for a few minutes without habituation, fatigue or boredom. Secondly, the spatial resolution is generally on the order of 1±3 mm, which means that effective head immobilization is essential. The longer a subject is lying in an uncomfortable position inside the MRI magnet, the greater the chance of large movement, for which it is difficult to correct. Thirdly, it is important to sample the activation state of the whole brain as synchronously as possible. Since MRI usually obtains image data one slice at a time, and 20±30 slices are necessary to cover the entire brain, this implies acquisition of a slice in a time very short compared with the haemodynamic response time of the cerebral vasculature of 6±8 s. The only successful MRI technique capable of this speed with reasonable spatial resolution and good signal-to-noise ratio is EPI (Stehling et al. 1991) as previously mentioned. Compared with slower MRI techniques, the spatial resolution of EPI is somewhat impaired (to about 2 mm), while the temporal resolution is improved to about 100 ms. Furthermore, the rate of information capture, that is, the signal-to-noise per unit time, is highest for EPI and its variants [such as single-shot GRASE (Feinberg and Oshio 1991) and single-shot spiral EPI (Ahn et al. 1986)]. However, EPI images (especially the gradientecho version of EPI) suffer more than the other techniques from distortion and signal loss arising from magnetic field inhomogeneities in the brain, which often derive from the natural susceptibility differences between brain and air and are thus not amenable to correction by improved shimming (adjustment of the static field homogeneity with a set of correction coils). The specification of commercially available fmri scanners can constrain experimental design. The Siemens Vision system can acquire up to 25000 images during a single run before pausing for data processing and disc storage. An image volume can comprise any number of axial slices (64 slices representing whole-brain acquisition) at variable repeat times. Not many experimental paradigms will require continuation of scanning longer than the 40±50 min this allows, although this may be significant for paradigms concerned with sustained physiological or psychological phenomena. Characteristics of BOLD contrast The difference in susceptibility between fully oxygenated and deoxygenated blood is small (about 0.02 10 Ÿ6 cgs units), and thus image intensity changes in BOLD contrast studies are generally small-less than 15% at a magnetic field of 2 T even when the haemoglobin oxygen sat-

uration is reduced to 20% in acute hypoxia (Jezzard et al. 1994). In brain activation studies at 1.5 T (Kwong et al. 1992) the signal change is usually no more than 2±4%. Experiments (Turner et al. 1993) and model calculations (Ogawa et al. 1993; Weiskoff et al. 1994; have shown that Kennan et al. 1994) the change in T2* relaxation rate associated with this signal loss increases with the static magnetic field of the MRI scanner, so that the change in signal is typically about 3 times larger at 4 T than at the more commonly used field strength of 1.5 T, for the same echo time and sequence type. Given the improved signal-to-noise ratio at higher field (roughly proportional to field strength) there is a strong case for the use of high magnetic fields for this type of functional MRI. At 4T, signal changes of up to 30% have been observed for visual stimulation, with typical single-shot background noise of 0.5% or less, giving z-values of 60 or more for some pixels. Of course, for more subtle brain activations in association cortex, intensity changes are generally much smaller ± from 2% to 8% at this field strength ± and proportionately lower at the more commonly used field of 1.5 T. At the lower field strength it is often necessary to perform averaging over images to obtain significant results, although single-subject remain eminently feasible. Data analysis principles In every quantitative science, when experimental data are collected the first and frequently the most successful analytic strategy is to attempt to fit them to a linear combination of the experimentally controllable variables, or to simple transforms of seen. Fast and efficient algorithms have been developed over the last several decades that use the methods of matrix algebra to handle multiple independent and dependent variables. If the data can be assumed, or better still demonstrated, to have normally distributed error (NDE), statistical significance can be assigned to the results in a fully validated way. Image intensity data are no different in principle from any other kind of data, though they are often more copious. In hypothesis-led experimental design and analysis, one or more independent variables are controlled or at least monitored, and the effects on intensity at any voxel in the image are recorded. The experimental design and the model used to test for specific responses to the controlled or monitored variables are embodied in what is known as the ªdesign matrixº. In principle, any non-random and measurable factor that can affect image intensity can be incorporated into the design matrix, and its contribution to intensity variations estimated within the accuracy of the assumption that the effect is linear. Both the magnitude of intensity variation, and the spatial extent of co-activated pixels, can be taken into account straightforwardly in making statistical interferences when the data be regarded as a Gaussian Random Field (GRF) (Adler 1981). The data of fmri are no exception, and provide a suitable arena for the application of these methods, which are commonly described as the use of multilinear regression, or the General Linear Model. Research groups that have correctly used t-tests, z-score maps or correlational analyses to interrogate their fmri data have been implementing, wittingly or unwittingly, special cases of this technique, which encompasses all statistically valid linear analyses of NDE data. Thus the fundamental question relating to fmri data analysis is how ªwell-behavedº, in a statistical sense, the raw data of MRI brain images are. To quality as forming a GRF, data must have the following properties: (a) the autocorrelation function most be twice differentiable, and (b) the spatial correlations must be stationary. Raw MRI data do not have these properties, because (a) the point spread function is narrow (usually not more than two pixels wide, full width at half maximum) ± thus the image is not smooth, sharp boundaries being represented by abrupt image intensity changes; and (b) head motion and physiological pulsations result in highly spatially correlated time-varying changes in image intensity. The strategy of statistical parametric mapping (SPM) when applied to fmri data is to use a series of image transformations, which have known statistical implications, to gaussianize the data, and thereby to allow robust statistical interference. Intrinsic features of fmri data, such as the spectral distribution of noise, favour specific types of experimental design. These reflect the temporal resolution of fmri and the large number of measurements that are typically made. At the same time the design needs to take account of the fact that unlike positron emission tomography (PET), fmri does not measure absolute blood flow changes but relative changes in the oxygenation of venous blood. Design issues also have implications for data processing; thus the normalization and statistical methods used to find significant activations need to address the particular characteristics of fmri data. Sources of statistical confounds in fmri data In time series images of phantoms, apart from slow drifts of image intensity caused for instance by temperature changes in imaging hardware, instrument noise is primarily thermal in origin, and is thus ªwhiteº within the receiver bandwidth, appearing uniformly across the image. In fmri time courses of human brain, additional low-frequency random and non-random noise components are evident. These arise from head motion, slow global variations in blood oxygenation, or physiological cardiac and respiratory pulsations. The spectrum of this noise typically has a ª1/fº characteristic, so called because the spectral power density falls linearly with increasing frequency, with additional peaks corresponding to periodic motions. These two generic types of noise combine to generate the form of spectrum seen in Fig. 1. In summary the noise spectrum of the fmri signal comprises a low-frequency and wideband component. Periodic physiological noise with a short image TR appears as focal peaks in the spec- 7

8 Fig. 1 Sketch of the noise spectrum found using echo-planar imaging (EPI) acquisition with short TR in studies of human brain trum, or when a long image TR is used can be aliased to give an increase in wideband noise. Bulk head motion has been recognized since the earliest days of fmri (Kwong et al. 1992; Ogawa et al. 1992) as potentially the most severe and misleading confound in fmri studies. Fortunately a typical time series of fmri volume data contains adequate information to characterize such motion with an accuracy of perhaps 30 m, when an appropriate algorithm is used (Friston et al. 1994a). Motion effects are removed by realignment of subsequent multislice images in a time series to the first multislice image, using sinc interpolation for greatest precision. If the repeat time is of the order of T1, out-of-plane motion creates a different spin excitation history for different groups of spins within the slice, and thus contaminates the data with motion-correlated image intensity changes (Friston et al. 1996). An ARMA-based algorithm removing any image intensity variations correlated with out-ofplane motion removes most of this source of variance, leaving functionally related activations intact. Physiological noise in brain images (Jezzard et al. 1993) arises almost entirely from T1 effects associated with the pulsatile motions of heartbeat and respiration (Weisskoff et al. 1993), with a small very slow residual component coming from spatially correlated endogenous fluctuations of blood oxygenation (Moskalenko et al. 1977; Biswal et al. 1995). The heart cycle causes both periodic blood flow and pulsatile bulk motion due to induced pressure variation. The blood flow effects are confined mainly to vessels, but the pulsatile motion, while most pronounced in the brain stem, can extend throughout the entire brain. Respiration causes a generalized variation in blood oxygenation, bulk displacement of the head in reaction to breathing, and changes in pressure in the venous sinuses that may be communicated to cerebrospinal fluid spaces. The expected time course of measured physiological noise depends not only on the sources but also on the imaging TR (repetition time). If the TR is short compared with both cardiac and respiratory cycles (i.e. TR<<1 s), then both are seen as straightforward periodic functions. In a frequency spectrum (neglecting harmonics) each appears as a peak with a central ªmeanº frequency and with widths representing departure from strict periodicity. If TR<T1 is used the effect of physiological noise on image intensity is enhanced, because spin excitation history effects become more important, as tissue moves in and out ot the imaging slice. For such short TR measurements it may be worthwhile to apply simple notch filters centred on the mean cardiac and respiratory frequencies (Le 1996). Clearly any activation frequency component within the notch filter stop bands will be discarded along with the noise. If whole-brain multislice EPI, or other slow single-slice techniques, are used the TR is typically several seconds. This is above the Nyquist limit for both the cardiac and respiratory noise and so aliasing will occur. Low-frequency noise components of the fmri timeseries are often due to cardiac and respiratory aliasing. For example, if scans are acquired every 6 s and the subject's respiratory cycle is 6.1 s, then each scan is 0.1 s out of phase. Over a prolonged period of scanning, this asynchrony creates an artefactual signal of low frequency, 0.0027 Hz. Other low-frequency components (Biswal et al. 1995) relate to Meyer's waves and related phenomena that are related to slow periodic haemodynamic changes (probably due to intrinsic autoregulatory feedback of cerebrovascular tone). The aliased spectral density function depends primarily on the spectral width of the noise. If the noise is centred at an arbitrary frequency but extends over a width of DF N >F R then its aliased contribution will extend across the entire measurement frequency spectrum. In this situation (i.e. non-monochromatic noise and TR long compared with noise period) a simple notch filter would not be appropriate. Functionally related time variations of changes in neuronal activity are delayed and smoothed by the haemodynamic time constant t, itself a combination of the time required to inform resistance vessels of the need for increased flow, and the time for the flow increase to reach the venular bed. t varies spatially within a factor of 2, with a typical value of 8 s. Data from a fmri experiment form a time-series that may contain many hundreds of measurements at each voxel. If the data are acquired sufficiently rapidly (e.g. every 6 s or less) this time-series can show temporal autocorrelations or temporal smoothness due to the haemodynamic response function and other sources of physiological noise. The effect of the haemodynamic response function can be thought of as convolving an underlying neural time-series with the response function to give the observed haemodynamic time-series (Friston et al. 1994b). In this context, time-dependent changes in the BOLD signal may be represented, in frequency space, as the sum of several frequency components (where the spectral density is the Fourier transform of the autocovariance function). Each frequency component has a different source; heuristically these sources can be thought of as either (i) signal, related to functionally activated brain areas (i.e. neuronal activity convolved with the haemodynamic response function) or (ii) noise, arising from aliased physiological biorhythms or slow movement artefacts. The successful experimental design is therefore one in which signal and noise in the fmri time series are not confounded ± the designs of the initial studies by Kwong et al. (1992) and Ogawa et al. (1992) already show this.

Removal of noise: experimental design implications Two techniques are used (often in combination) to counter noise. Firstly, noise processes can be modelled and their estimated contribution removed from a measured signal. This technique is used, for example, to remove movement-related signal variation mediated by differential spin excitation histories (Friston et al. 1996). Secondly, an experiment can be designed to take account of the characteristic features of the measurement system, for example the noise and sensitivity frequency spectra. An appropriate example in the present context is the method of ªchopping. Chopping involves alternating between stimulus or task conditions to generate a time-dependent activation with a fundamental frequency sufficiently high to be clear of the 1/f region in the noise spectrum without being so high as to be attenuated by the sluggish a haemodynamic response. Chopping utilizes the high temporal resolution of fmri to compensate for its poorer low-frequency noise performance. This leads to a choice of chopping frequencies between approximately 1 cycle every few minutes (a typical low-frequency noise cut-off point) and 1 cycle every 10 s (above which the signal would be attenuated by the haemodynamic response function). Many researchers have found in practice that it is convenient to design experiments with control and activation periods alternating every 30 or 40 s. In the context of the theoretical analysis presented above, this ensures that activations of interest and physiological noise are confounded as little as possible. The recurring control periods, during which the subject may be performing an engrossing but low-level task, provide a constant baseline that can be used to assess low-frequency baseline drift secondary to physiological noise. Low-frequency effects may be assessed explicitly by comparing control periods at the beginning and end of the experiment, or modelling time effects in the rest periods explicitly. However, it is simpler and more practical to remove slowly varying components from the data by explicitly modelling them as a sum of sinusoidal terms of amplitude to be determined in the multilinear reression. This has the effect of a high-pass filter, with a precisely known reduction in the remaining degrees of freedom in the data set. It is important to set the filter periodicity appropriately with regard to the experimental design in order not to remove signals of interest. For example, a signal of interest of periodicity 60 s can be achieved by alternating periods of 30 s activation (five whole-brain acquisitions with a TR of 6 s) with 30 s rest. Subsequently applying a high-pass filter of periodicity 120 s will remove noise but leave the signal of interest intact. Data analysis The analysis of fmri data differs in some important ways from that of PET data. The time to analyse fmri data is not trivial. Statistical analysis alone, for 1200 16-slice images on an 8-cpu Sun SparcServer, can take approximately 72 h. The fundamental principle on which SPM is based is that a signal depending simultaneously on several variables can be decomposed into contributions from each of those variables, if a sufficient number of readings of the signal are obtained with different combinations of the independent variables. The first-order assumption, widely used in statistical analyses, is that the dependence of the signal on any given variable is linear. Let the departure from its mean, at the time of the ith scan, of the MR brain image intensity in a given voxel j (j=1,2,3, ¼ N) be denoted x ij. The general linear model (GLM) (Friston et al. 1995a) postulates that the value of x ij is related linearly to chosen independent variables (denoted g ik ) affecting the condition of the subject's brain at the time of scan i. Such a variable might be the intensity of some particular sensory stimulus, or the subject's arterial pco 2, or any other controllable variable. Thus we may write: x ij ˆ S k g ik b kj e ij where b kj are a set of coefficients, to be determined by the analysis, relating the change in image intensity to the experimental independent variables, and e ij are random error terms conforming to the normal or Gaussian distribution. In the context of fmri data analysis, the various proposed methods such as t-tests, correlation with a box-car function and F-tests can be viewed as closely related ways of obtaining the coefficients b kj applicable to the experimental conditions, and determining their significance. To convert the data into a spatial Gaussian field they are spatially smoothed, by interpolation to a smaller pixel size, and then convolved using an isotropic Gaussian kernel in space, with a full width at half maximum of the original pixel dimensions. Smoothing in space enhances the signal-to-noise ratio of the data and allows intersubject averaging by blurring differences in gyral anatomy between subjects. The validity of statistical interferences also depends on the data being sufficiently smooth to constitute of GRF. On the other hand, fmri has high anatomical resolution and so there is a trade-off between the degree of smoothing and the spatial resolution to be considered. Our approach has been to use the least degree of smoothing that is compatible with the smoothness being large enough to ensure the validity of the statistical interferences (i.e. a smoothness of at least 2 voxels). Once these operations have been performed the GLM may be applied to the data, giving unambiguous SPMs of brain activation. In the raw image data, even after head motion correction, the error is clearly not normally distributed, since non-random pulsatile effects account for much of the noise. This problem, which has led some investigators to opt for less powerful non-parametric statistical methods, can be overcome by the use of optimal temporal smoothing (Friston et al. 1995b) convolving the time data with a Gaussian filter having a time width equal to the typical haemodynamic response time t. This device, precisely analogous to the optimal line-broadening strategy used universally in MR spectroscopy, results in an approximately normal 9

10 distribution of error, at the expense of introducing additional temporal correlation. No functionally significant information is lost, and the signal-to-noise ratio can be significantly improved if TR is comparable with t. Analysis of an fmri time series involves fitting a model to the observed data. The model is composed of several different components specified in the design matrix. However, the GLM used for PET data analysis assumed that each scan represents an independent observation. The presence of temporal autocorrelation in fmri time series in a violation of this assumption. The approach we have adopted is to extend the GLM to accommodate these temporal autocorrelations. The variance estimators are recalculated under the assumption that the data represent an uncorrelated set of observations that have been convolved with some smoothing kernel. The resulting statistics are then distributed with the t distribution whose effective degrees of freedom are a function of the smoothing kernel. The effective degrees of freedom are less than the number of scans and correspond roughly to the overall scan time divided by the width of the smothing kernel. This approach assumes the temporal autocorrelation function is known, or can be estimated. In many instances it can be deduced from the temporal smoothing kernel applied to the time series to increase the signal-to-noise ratio. This de-noising device is predicated on the ªmatched filter theoremº and is based on the conjecture that ªinterestingº haemodynamics are the result of convolving an underlying neuronal process with a haemodynamic response function. As such the signal is always ªsmootherº than (or as smooth as) the response function. Clearly the optimum smoothing kernel is related to the haemodynamic response function. ªOptimumº refers to the kernel that maximizes variance in the signal frequencies relative to other frequencies. It is generally accepted that the haemodynamic response function has an associated delay and dispersion of about 6±8 s. This corresponds to a Gaussian kernel of width Ö6=2.45 s to Ö8=2.83 s and it is this sort of kernel that is usually applied to time series with short (<6 s)tr. In many fmri experiments several subjects are studied, and this raises the question of how best to proceed with data analysis. As there are no limits on the volume of data that can be acquired in a single subject (unlike PET), it is possible to perform an experiment was a series of single case studies. The resultant statistical parametric maps can be displayed directly as an anatomical rendering of that subject's brain. Localization of function is therefore undertaken by referring directly to the surface anatomy. Such a solution characterizes the data most fully in terms of both the similarities and differences between the functional anatomy of each subject. Reporting of such data poses problems; a qualitative description of areas activated in common across subjects is necessary rather than a quantitative comparison between subjects. The comparison of results between different groups becomes more difficult as results are not reported in a standard anatomical space. The ability to normalize fmri data into a common stereotactic space, such as that of Talairach and Tournoux, allows the reporting of significant activations in a common reference space. However, the deformations introduced by single-subject normalization may compromise anatomical resolution. Spatial normalization also allows intersubject averaging of data. However, the intersubject variability of gyral anatomy is significantly greater than the anatomical resolution of fmri and so it is likely that a significant degree of spatial smoothing will be required to detect areas activated in common across subjects. This inevitably compromises the spatial resolution of fmri ± one of its most attractive features. It is likely that the use of these related approaches will depend on the nature of the question being studied. For example, studies requiring high anatomical resolution may rely more on single-subject analyses. Results of SPM analysis: slice order effects A controversy that has preoccupied a number of fmri researchers is the question of whether the functionally correlated signals observed with BOLD techniques arise from oxygenation changes in cortical capillaries, in the same location as the active neurons, or from larger vessels draining active cortical areas, where activation may artefactually appear some distance away. The question of whether the MRI sequences used are sensitive to blood flow velocity, in addition to blood susceptibility changes, is crucial in this debate. Changes in flow velocity arising in an active cortical vascular bed may inded propagate downstream in the venous drainage by some distance, and even be amplified as a result of the overall reduction in total venous crosssectional area, as a simple calculation will readily demonstrate. Thus flow-sensitive techniques, such as that of Segebarth and colleagues can demontrate apparent activation (ªinflow effectº) in the form of substantial changes in blood flow velocity, in large veins centimetres away the cortex thought to be active. By contrast, changes in blood oxygenation arising from localized brain activity are rapidly diluted downstream from the activation, as additional veins carrying unchanged blood enter the drainage network. The distance downstream over which one may expect oxygenation changes to propagate will depend on the fraction that happens to be activated of the watersheld of the vein under consideration. Thus BOLD effects distant from the cortical actiation can only arise when large areas of cortex are excited, as in epileptic seizures (Penfield 1993). Under normal circumstances the evidence is that the downstream ªleakageº is no more than a few millimetres. Since inflow effects can produce extremely large image intensity changes (10±20%) it is vital to ensure that the fmri imaging sequence to be used is insensitive to flow velocity (Frahm et al. 1994). One convenient method for confirming this for multislice EPI sequences with long TR is to repeat studies on the same subject with the same paradigm but varying slice order and orientation. BOLD effects arising from isotropically distributed small vessels can be expected to be independent of each of these variables, but inflow effects, if they are significant, will depend

11 Fig. 2 Orthogonal sections of human brain showing regions activated by a visual stimulus. Each EPI data set, 120 repeats of 64 consecutive slices, was acquired with a different slice orientation as shown. Differences in activation patterns can be accounted for entirely by the different image distortion entailed by the different direction of the phase encoding gradient used in each orientation Fig. 3 Orthogonal sections of human brain showing regions activated by a visual stimulus. Each EPI data set, 120 repeats of 64 consecutive slices, was acquired with a different slice order as shown. Statistical parametric maps of the differences in acquisition patterns between these three slice orderings showed no significant areas

12 on slice order and also on slice orientation, as magnetically saturated blood moves from slice to slice. Figure 2 shows data from the same subject, analysed using SPM, from an experiment using reversing checkerboard visual stimulation while slice orientation was varied. A Siemens Vision MRI scanner, at 2 T magnetic field, was used to generate gradient-echo echo-planar images, with slice thickness 3 mm, 64 64 pixel resolution, 192 mm FOV, TE 40 ms. Each volume of 64 slices took 6 s to acquire, and 120 volumes were acquired on each occasion. The susceptibilityinduced image distortions characteristic of EPI are predictably different for the three slice orientations, but it is clear that the activated regions in each case are highly similar. Figure 3 shows results from a similar experiment where slice order was varied while the same subject experienced the same visual stimulation protocol. Here the correspondence between the activation patterns is even more precise. Statistical analysis of the differences in these patterns revealed no significant areas, confirming the freedom from inflow effects found in these types of acquisition. Conclusions Consideration of the relationship of blood oxygenation level-dependent contrast and sequence design, and use of well-validated image processing and statistical analysis techniques, results in reliable maps of human brain activity that will assist the rapid development of cognitive neuroscience. References 1. Adler RJ (1981) The geometry of random fields. Wiley, New York 2. Ahn CB, Kim JH, Cho ZH (1986) High-speed spiral-scan echo planar NMR imaging. IEEE Trans Med Imag 5:2 3. Belliveau JW et al. (1988) Real-time proton susceptibility-contrast imaging of cerebral physiology. Proc Soc Magn Reson Med 222 4. Belliveau JW et al. (1991) Functional mapping of the human visual-cortex by magnetic-resonance-imaging. Science 254:716± 719 5. Biswal B, Yetkin FZ, Haughton VM, Hyde JS (1995) Functional connectivity in the motor cortex of resting human brain using echo-plannar MRI. Magn Reson Med 34:537±541 6. Feinberg DA, Oshio K (1991) GRASE (gradient-echo and spinecho) MR imaging: a new fast clinical imaging technique. Radiology 181:597 7. Frahm J, Merboldt K-D, Hanicke W, Kleinschmidt A, Boecker H (1994) Brain or vein: oxygenation or flow? On signal physiology in functional MRI of human brain. NMR Biomed 7:45±53 8. Friston KJ, Ashburner J, Frith CD, Poline J-B, Heather JD, Frackowiak RSJ (1994a) Spatial registration and normalization of images. Hum Brain Mapping 2:165±189 9. Friston KJ, Jezzard P, Turner R (1994b) Analysis of functional MRI time series. Hum Brain Mapping 1:153±171 10. Friston KJ et al. (1995a) Statistical parametric maps in functional imaging: a general linear approach. Hum Brain Mapping 2:189±210 11. Friston KJ, Holmes AP, Poline J-B, Grasby PJ, Williams SCR, Frackowiak RSJ, Turner R (1995b) Analysis of fmri time-series revisited. Neuroimage 2:45±53 12. Friston KJ, Williams SCR, Howard R, Frackowiak RSJ, Turner R (1996) Movement related effects in fmri time series. Magn Reson Med 35:346±355 13. Haase A, Frahm J, Matthaei D, Hänicke W, Merboldt K-D (1986) FLASH imaging: rapid NMR imaging using low flip angle pulses. J Magn Reson 67:258 14. Jezzard P et al. (1993) An investigation of the contribution of physiological noise in human functional MRI studies at 1.5 tesla and 4 tesla. Proc SMRM, 12th Annual Meeting, New York 1993, p 1392 15. Jezzard P et al. (1994) Comparison of EPI gradient-echo contrast changes in cat brain caused by respiratory challenges with direct simultaneous evaluation of cerebral oxygenation via a cranial window. NMR Biomed 7:35±44 16. Kennan RP, Zhong J, Gore JC (1994) Intravascular susceptibility contrast mechanisms in tissues. Magn Reson Med 31:9±21 17. Kwong KK et al. (1992) Dynamic magnetic resonance imaging of human brain activity during primary sensory stimulation. Proc Natl Acad Sci USA 89:5675±5679 18. Le TH, Hu X (1996) Retrospective estimation and correction of physiological artifacts in fmri by direct extraction of physiological activity from MR data. Magn Reson Med 35:290±298 19. Mansfield P (1977) Multiplanar image formation using NMR spin echoes. J Phys C 10:L55±L58 20. Moskalenko S et al. (1997) Biophysical aspects of cerebral circulation. Pergamon Press, Oxford 21. Ogawa S, Lee T-M, Nayak AS, Glynn P (1990) Oxygenationsensitive contrast in magnetic-resonance image of rodent brain at high magnetic-fields. Magn Reson Med 14:68 22. Ogawa S et al. (1992) Brain magnetic-resonance-imaging with contrast dependent on blood oxgenation. Proc Natl Acad Sci USA 89:5951±5955 23. Ogawa S et al. (1993) Functional brain mapping by blood oxygenation level-dependent contrast magnetic-resonance-imaging: a comparison of signal characteristics with a biophysical model. Biophys J 64:803±812 24. Pauling L, Coryell CD (1936) The magnetic properties and structure of hemoglobin, oxyhemoglobin and carbomonoxyhemoglobin. Proc Natl Acad Sci USA 22:210±216 25. Penfield W (1993) The evidence for a cerebral vascular mechanism in epilepsy. Ann Intern Med 7:303±310 26. Segebarth C, Belle V, Delon C, Massarelli R, Decety J, Le Bas JF, Decorps M, Benabid A (1994) Functional MRI of the human brain: predominance of signals from extracerebral veins. Neuroreport 5:813±816 27. Stehling MK, Turner R, Mansfield P (1991) Echo-planar imaging: magnetic-resonance-imaging in a fraction of a second. Science 254:43±50 28. Stewart GN (1894) Researches on the circulation time in organs and on the influences which affect it. J Physiol (Lond) 15:1 29. Turner R, Le Bihan D, Moonen CTW, Despres D, Frank J (1991) Echo-planar time course MRI of cat brain deoxygenation changes. Magn Reson Med 22:159 30. Turner R et al. (1993) Functional mapping of the human visualcortex at 4 and 1.5 tesla using deoxygenation contrast EPI. Magn Reson Med 29:281±283 31. Villringer A et al. (1986) Dynamic imaging with lanthanide chelates in normal brain: contrast due to magnetic-susceptibility effects. Proc Soc Magn Reson Med 1:21 32. Weisskoff RM et al. (1993) Power spectrum analysis of functionally-weighted MR data: what's in the noise? Proc SMRM, 12th Annual Meeting, New York 1993, p 7 33. Weisskoff RM, Zuo CS, Boxerman JL, Rosen BR (1994) Microscopic susceptibility variation and transverse relaxation ± theory and experiment. Magn Reson Med 31:601±610