NATUREVol 453|12 June 2008 REVIEWS time,high magnetic field scanners are likely to require even tighte tateof arousalof te subjectHigh-speed MRImethods,cap I by t or the response in such experr t deal of and ach s have distinct.non-over ppin ropi)wh 正 a hu n).Wi blocked subtraction paradig ly b night.at least to some tacke the probl with the xpectation that itwil ventual ugh fo cou sing [MRI nd the Afer adaptation the su to read a n r with a mi y s in th the to the h striosomes of basal on must know a great deal ab to th ing ofac ibuted l system.su ch as that unde one have been widely used in co and the w油10 d eiving are ch more rtant ver. and dit 0.7×9 far hav exploit prove optimal for the vastm tions b en erro terms owing to physiolo al n sal at d and h mos Anothe s to tak into account eminiscentof y or att all diffo be task or stimulus con es tha are no man savin In s underlyi ng th patterns. The presence, ion,with an emphasis subtraction des be pote ally detected by ly on pu that into a task ithout aff different types of neural population sumpti What do activation maps repre ent? design c ogaitinveleeL Does the a ctivation of an are a n tion ation in the studied b iour.But do we?It isu the determine their interactions ion,in tantiated in the patterns ith the hat the ularcogniticproockd& stis indeed being isolate ceptualized as information with an input, (that is signal difference between test and e2008 Macmillan Publishers L All rights reserved facilities in the future, and this should definitely improve the quality of data obtained in human magnetic resonance studies. At the same time, high magnetic field scanners are likely to require even tighter interaction between magnetic resonance physicists and application scientists, as the much larger inhomogeneity of both B0 (main static field) and B1 (the field generated by the excitation pulses) at high field will demand a great deal of expertise and experimental skill to achieve the desired image quality. All in all, MRI may soon provide us with images of a fraction of a millimetre (for example, 300 3 300 mm2 with a couple of millimetres slice thickness or 50035003500 mm3 isotropic), which amount to voxel volumes of about two–three orders of magnitude smaller than those currently used in human imaging (see ‘Developments and perspectives’ in the Supplementary Information). With an increasing number of acquisition channels such resolution may ultimately be attained in whole-head imaging protocols, yielding unparalleled maps of distributed brain activity in great regional detail and with reasonable—a couple of seconds—temporal resolution. Would that be enough for using fMRI to understand brain function? The answer obviously depends on the scientific question and the spatial scale at which this question could be addressed—‘‘it makes no sense to read a newspaper with a microscope’’, as neuroanatomist Valentino Braitenberg once pointed out. To understand the function￾ing of the microcircuits in cortical columns or of the cell assemblies in the striosomes of basal ganglia, one must know a great deal about synapses, neurons and their interconnections. To understand the functioning of a distributed large-scale system, such as that underlying our memory or linguistic capacities, one must first know the archi￾tectural units that organize neural populations of similar properties, and the interconnections of such units. With 1010 neurons and 1014 connections in the cortex alone, attempting to study dynamic inter￾actions between subsystems at the level of single neurons would prob￾ably make little sense, even if it were technically feasible. It is probably much more important to understand better the differential activity of functional subunits—whether subcortical nuclei, or cortical columns, blobs and laminae—and the instances of their joint or conditional activation. If so, whole-head imaging with a spatial resolution, say, of 0.7 3 0.7 mm2 in slices of 1-mm thickness, and a sampling time of a couple of seconds, might prove optimal for the vast majority of ques￾tions in basic and clinical research. More so, because of the great sensitivity of the fMRI signal to neuromodulation (see below and Supplementary Information). Neuromodulatory effects, such as those effected by arousal, attention, memory, and so on, are slow and have reduced spatiotemporal resolution and specificity7,8. Designs and analyses. Many studies initially used block designs, reminiscent of earlier positron emission tomography (PET) paradigms. These designs use time-integrated averaging procedures, and usually analyse the data by means of subtraction methods. The central idea is to compare a task state designed to place specific demands on the brain with an investigator-defined control state. Under these conditions, both enhancements and reductions of the fMRI signal are observed. In the early cognitive fMRI studies the prevailing block design was cognitive subtraction, with an emphasis on serial subtraction designs9 . Such designs rely strictly on pure insertion, which asserts that a single cog￾nitive process can be inserted into a task without affecting the remain￾der, an assumption that all too often is not tenable (see ‘On pure insertion’ in the Supplementary Information). Even if an experimental design could satisfy this assumption at the cognitive level, the assump￾tion would be condemned to fail at the level of its neuronal instantia￾tion10 owing to the highly nonlinear nature of most brain processes. To overcome this kind of problem and ensure better interpretation of the neuroimaging data it is necessary to perform a detailed task analysis to determine subtraction components and their interactions. Yet most neuroimaging studies provide noformal task analysis thatwould ensure that the particular cognitive process of interest is indeed being isolated by the subtraction11. Traditional block designs have excellent functional contrast-to-noise ratio (that is, signal difference between test and control epochs, normalized to the mean signal of all epochs), but they are usually long (from 20 to 60 s), and may be confounded by the general state of arousal of the subject. High-speed fMRI methods, capa￾ble of whole-brain imaging with a temporal resolution of a few seconds, enabled the employment of so-called event-related designs12. The time course of the response in such experiments is closer to the underlying neural activity. The block designs discussed so far may reveal differential patterns of activation only in those cases in which different stimulus attributes or different cognitive processes have distinct, non-overlapping spa￾tial organizations. Overlapping networks of neurons subserving dif￾ferent functions are likely to go unnoticed owing to the spatial averaging that characterizes the blocked subtraction paradigms. Functional MRI adaptation designs were conceived as tools that might, at least to some extent, tackle the problem of spatially over￾lapping neural networks13. In this experimental design, a stimulus is presented repeatedly with the expectation that it will eventually induce response adaptation in neurons selective for its various prop￾erties. In general, repetition of an identical stimulus does indeed produce a reduction in the fMRI signal. After adaptation, the subject is presented with a stimulus that is varied along one dimension (for example, the direction of a moving pattern or the view of a human face) and the possibility of a response rebound is examined. If the underlying neural representation is insensitive to the changes in the stimulus then the fMRI signal will be reduced, similar to the reduc￾tion produced by the repetition of identical stimuli. Alternatively, if the neurons are sensitive to the transformation, the signal will show a clear rebound to its original, pre-adaptation level. Functional MRI adaptation designs have been widely used in cog￾nitive neuroscience, but they also have shortcomings, as any area receiving input from another region may reveal adaptation effects that actually occurred in that other region, even if the receiving area itself has no neuronal specificity for the adapted property13. Moreover, the conclusions of experiments relying on adaptation designs strongly rely on existing electrophysiological evidence, which itself may hold true for one area and not for another72. Finally, clever analysis is required to exploit clever design. Most studies so far have used voxel-based conventional analyses of MRI time series from one or more subjects14. The approach is predicated on an extension of the general linear model that allows for correla￾tions between error terms owing to physiological noise or correla￾tions that ensue after temporal smoothing. The method is reliable and, when well implemented, offers the best analysis strategy for most studies. Another approach is to take into account the full spatial pattern of brain activity, measured simultaneously at many loca￾tions15. Such multivariate analyses or pattern-classification-based techniques (decoding techniques) can often detect small differences between two task or stimulus conditions—differences that are not picked up by conventional univariate methods. However, this is not equivalent to saying that they unequivocally reveal the neural mechanisms underlying the activation patterns. The presence, for instance, of voxels selective to two different stimulus attributes could be potentially detected by modern classifiers, yet the existence of two types of patterns does not necessarily imply the existence of two different types of neural populations72. What do activation maps represent? Does the activation of an area mean that it is truly involved in the task at hand? This question implies that we understand what neural acti￾vity in a given area would unequivocally show its participation in the studied behaviour. But do we? It is usually alleged that cognitive capacities reflect the ‘local processing of inputs’ or the ‘output’ of a region, instantiated in the patterns of action potentials, with their characteristic frequency and timing. In principle, brain structures can be conceptualized as information processing entities, with an input, a local-processing capacity, and an output. Yet, although such a scheme may describe the function of subcortical nuclei, its implementation NATUREjVol 453j12 June 2008 REVIEWS 871 ©2008 Macmillan Publishers Limited. All rights reserved