NATUREVol 45312 June 2008 REVIEWS potentials>Similar dissociations between spikes and CBF had thickness",the overall length of neuronal processes remains rela been demonstrated earlier and very recently in a number of studies tively constant,with axonal length being approximately 4km mm using other techniques6-ss and dendrite length 0.4 km mm.Overall,synaptic density and the The findings are in close agreement with a number of older auto- ratio of excitatory to inhibitory synapses also remain constant. radiography studies,also showing that regional glucose utilization is Given these neuro-statistical data,what are the actual contents of a directly related to neuronal synaptic activity3.For example,the neuroimaging voxel?An examination of the 300 top-cited cognitive greatest 2-DG uptake occurs in the neuropil (that is,in areas rich fMRI studies suggests that the commonly used in-plane resolution is in synapses,dendrites and axons,rather than in cell bodies).During 9-16 mm2,for slice thicknesses of 5-7mm.The average voxel size orthodromic and antidromic electrical microstimulation,only before any pre-processing of the data is thus 55 ul(or 55 mm3).Often orthodromic microstimulation,which involves presynaptic term- the effective size is 2-3 times larger due to the spatial filtering that inals,increases glucose consumption.Similarly,the highest density most investigators apply to improve the functional SNR.Less than of cytochrome oxidase(an enzyme of the respiratory chain)is found 3%of this volume is occupied by vessels and the rest by neural in somato-dendritic regions that are adjacent to axon terminals. elements(see Fig.3)A typical unfiltered fMRI voxel of 55 ul in size Finally,as mentioned earlier,presynaptic activity increases metabol- thus contains 5.5 million neurons,2.2-5.5 X 1010 synapses,22 km of ism even if the output is inhibited (that is,the spiking activity is dendrites and 220km of axons. abolished). This 'large population view'is in contrast to the scope of the Despite all this evidence,some discussion still concentrates on the traditional microelectrode recordings.It would be nice if we could importance of the firing rate of action potentials of projection neu- monitor every relevant neuron in the cortex during intracortical rons in the generation of the haemodynamic responses,perhaps microelectrode recordings,but this is practically impossible stemming from the fact that important early studies of neural corre- Instead,the typical electrophysiological measurements in behaving lates of behaviour took the mean spiking rate to be the gold standard animals report only on the properties of most active large neurons for quantifying neuronal activation.These discussions,however, that constitute a minority.The strong selection bias during extracel- often suffer from a certain amount of contention seeking where none lular recordings is partly due to practical limitations (for example, is warranted.In many cases,spikes do indeed correlate with LFPs, injury or simply size bias2)and partly to the physiological properties and they will also correlate with the BOLD signal.In addition,unusu- of neurons and/or the organizational principles of neural networks. ally high correlations between multiple unit activity and BOLD signal In fact,many different types of electrical and optical measurements (or LFP and multiple unit activity)may result from excessive signal- provide evidence that a substantial proportion of neurons,including smoothing owing to sampling rates of several seconds rather than a the cortical pyramidal cells,might be silent.Their silence might fraction of a second,as well as inter-subject averaging when simultan- reflect unusually high input selectivity or the existence of decoding eous physiology and fMRI measurements are not possible (see ref.55 for discussion). schemes relying on infrequent co-spiking of neuronal subsets.Most Predicting neural activity from the fMRI signals.Functional MRI important for the comparison of neuroimaging and electrophysiol- signals are presumed to result from changes in the activity of the ogy results is the fact that lack of measurable neuronal spiking may neuronal populations responsible for the functions in question(for not necessarily imply lack of input and subthreshold processing. example,stimulus-or task-selective neurons).This assumption is A direct analogy between neuronal spiking as measured in animal mainly based on decades of electrophysiology research with record- experiments and the fMRI signal obtained in human recording is thus simply unrealistic and might often lead to incorrect conclusions. ings from isolated single neurons in experimental animals,in which particular sensory stimuli that the animal perceives or tasks that it It is hardly surprising that most studies so far relying purely on BOLD performs were found to increase the firing rate of certain cells but not fMRI have failed to reveal the actual neural properties of the studied of others.The psychologist or cognitive neuroscientist who finds area,at least those properties (for example,selectivity to various cortical area X to be activated by the task at hand implicitly or expli- visual features)that were previously established in electrophysio- citly assumes that-if an electrode were placed in the subject's logical studies. brain-an increase in the spiking rate of those specialized neurons An example is cortical area V5(or MT)that has been extensively underlying the subject's behaviour would be observed.This might studied in the context of motion processing and perception465 well be true in some cases,but not in all.When attempting to inter- Electrophysiology has shown that the vast majority of the V5 neurons pret the fMRI signal by modelling,or when comparing the results of in monkeys are direction and speed selective.Neuroimaging loca- human neuroimaging to those obtained in monkey physiology lized the homologue of area V5 in humans as an area responding experiments,it is useful to take the following facts into consideration. stronger to moving than to stationary stimuli.Later studies suggested In humans,there are about 90,000-100,000 neurons under 1 mm that human V5 is sensitive to motion direction,and that it may be of cortical surface.This number is relatively constant for all structur- thought of as containing large populations of directionally selective ally and functionally distinct areas,including the somatosensory, units,just like its monkey homologue.The studies of directional temporal,parietal,frontal and motor cortical areas3.An exception specificity exploited the phenomenon of motion after-effect induced is the primary visual cortex of certain primates,including monkey by motion adaptation.After prolonged exposure to a stimulus mov- and human,which has approximately twice as many neurons.The ing in one direction,subjects perceive a subsequent static stimulus to number of cortical neurons under unitary cortical surface is also move in the opposite direction.It is assumed that motion after-effect similar across many species,including mouse,rat,cat,monkey and is due to the fact that the balance of mutual inhibition (opponency) human.Its small variability is the result of a trade-off between cor- between detectors for opposite directions of movement is distorted tical thickness and neural density.The former varies from area to area after adaptation.The sensitivity of the detectors selective for the and from species to species (for example,from mouse to human the adapting direction is reduced,which in turn releases from inhibition cortex becomes approximately three times thicker).Neural density the neurons selective for the opposite direction.Using this phe- varies inversely to cortical thickness.On average,density is 20,000 to nomenon,human studies demonstrated that the fMRI response to 30,000 neurons per mm';it peaks in the primary visual cortex by a a stationary stimulus was greater when the stimulus was preceded by factor of 4,and it is minimal in the motor cortex.Synaptic density a motion-after-effect-inducing,unidirectional adaptation,than ranges from 0.4 to 1 X 10per mm'.Depending on the thickness of when preceded by bidirectional adaptation Given the existing the cortex(2-4 mm),the number of synapses beneath I mm2surface physiology data in the monkey V5,these findings were interpreted is around 10(0.8-4X 10).Although the number of synapses as demonstrating that the BOLD signal directly reflects direction- and the axonal length per neuron increases with increasing cortical selective spiking activity of the area. 875 2008 Macmillan Publishers Limited.All rights reservedpotentials54,55. Similar dissociations between spikes and CBF had been demonstrated earlier and very recently in a number of studies using other techniques56–58. The findings are in close agreement with a number of older auto￾radiography studies, also showing that regional glucose utilization is directly related to neuronal synaptic activity35. For example, the greatest 2-DG uptake occurs in the neuropil (that is, in areas rich in synapses, dendrites and axons, rather than in cell bodies). During orthodromic and antidromic electrical microstimulation, only orthodromic microstimulation, which involves presynaptic term￾inals, increases glucose consumption. Similarly, the highest density of cytochrome oxidase (an enzyme of the respiratory chain) is found in somato-dendritic regions that are adjacent to axon terminals. Finally, as mentioned earlier, presynaptic activity increases metabol￾ism even if the output is inhibited (that is, the spiking activity is abolished). Despite all this evidence, some discussion still concentrates on the importance of the firing rate of action potentials of projection neu￾rons in the generation of the haemodynamic responses, perhaps stemming from the fact that important early studies of neural corre￾lates of behaviour took the mean spiking rate to be the gold standard for quantifying neuronal activation. These discussions, however, often suffer from a certain amount of contention seeking where none is warranted. In many cases, spikes do indeed correlate with LFPs, and they will also correlate with the BOLD signal. In addition, unusu￾ally high correlations between multiple unit activity and BOLD signal (or LFP and multiple unit activity) may result from excessive signal￾smoothing owing to sampling rates of several seconds rather than a fraction of a second, as well as inter-subject averaging when simultan￾eous physiology and fMRI measurements are not possible (see ref. 55 for discussion). Predicting neural activity from the fMRI signals. Functional MRI signals are presumed to result from changes in the activity of the neuronal populations responsible for the functions in question (for example, stimulus- or task-selective neurons). This assumption is mainly based on decades of electrophysiology research with record￾ings from isolated single neurons in experimental animals, in which particular sensory stimuli that the animal perceives or tasks that it performs were found to increase the firing rate of certain cells but not of others. The psychologist or cognitive neuroscientist who finds cortical area X to be activated by the task at hand implicitly or expli￾citly assumes that—if an electrode were placed in the subject’s brain—an increase in the spiking rate of those specialized neurons underlying the subject’s behaviour would be observed. This might well be true in some cases, but not in all. When attempting to inter￾pret the fMRI signal by modelling, or when comparing the results of human neuroimaging to those obtained in monkey physiology experiments, it is useful to take the following facts into consideration. In humans, there are about 90,000–100,000 neurons under 1 mm2 of cortical surface. This number is relatively constant for all structur￾ally and functionally distinct areas, including the somatosensory, temporal, parietal, frontal and motor cortical areas16,59. An exception is the primary visual cortex of certain primates, including monkey and human, which has approximately twice as many neurons. The number of cortical neurons under unitary cortical surface is also similar across many species, including mouse, rat, cat, monkey and human. Its small variability is the result of a trade-off between cor￾tical thickness and neural density. The former varies from area to area and from species to species (for example, from mouse to human the cortex becomes approximately three times thicker). Neural density varies inversely to cortical thickness. On average, density is 20,000 to 30,000 neurons per mm3 ; it peaks in the primary visual cortex by a factor of 4, and it is minimal in the motor cortex59,60. Synaptic density ranges from 0.4 to 1 3 109 per mm3 . Depending on the thickness of the cortex (2–4 mm), the number of synapses beneath 1 mm2 surface is around 109 (0.8–4 3 109 ). Although the number of synapses and the axonal length per neuron increases with increasing cortical thickness61, the overall length of neuronal processes remains rela￾tively constant, with axonal length being approximately 4 km mm23 and dendrite length 0.4 km mm23 . Overall, synaptic density and the ratio of excitatory to inhibitory synapses also remain constant. Given these neuro-statistical data, what are the actual contents of a neuroimaging voxel? An examination of the 300 top-cited cognitive fMRI studies suggests that the commonly used in-plane resolution is 9–16 mm2 , for slice thicknesses of 5–7 mm. The average voxel size before any pre-processing of the data is thus 55 ml (or 55 mm3 ). Often the effective size is 2–3 times larger due to the spatial filtering that most investigators apply to improve the functional SNR. Less than 3% of this volume is occupied by vessels and the rest by neural elements (see Fig. 3) A typical unfiltered fMRI voxel of 55 ml in size thus contains 5.5 million neurons, 2.2–5.5 3 1010 synapses, 22 km of dendrites and 220 km of axons. This ‘large population view’ is in contrast to the scope of the traditional microelectrode recordings. It would be nice if we could monitor every relevant neuron in the cortex during intracortical microelectrode recordings, but this is practically impossible. Instead, the typical electrophysiological measurements in behaving animals report only on the properties of most active large neurons that constitute a minority. The strong selection bias during extracel￾lular recordings is partly due to practical limitations (for example, injury or simply size bias62) and partly to the physiological properties of neurons and/or the organizational principles of neural networks. In fact, many different types of electrical and optical measurements provide evidence that a substantial proportion of neurons, including the cortical pyramidal cells, might be silent63. Their silence might reflect unusually high input selectivity or the existence of decoding schemes relying on infrequent co-spiking of neuronal subsets. Most important for the comparison of neuroimaging and electrophysiol￾ogy results is the fact that lack of measurable neuronal spiking may not necessarily imply lack of input and subthreshold processing. A direct analogy between neuronal spiking as measured in animal experiments and the fMRI signal obtained in human recording is thus simply unrealistic and might often lead to incorrect conclusions. It is hardly surprising that most studies so far relying purely on BOLD fMRI have failed to reveal the actual neural properties of the studied area, at least those properties (for example, selectivity to various visual features) that were previously established in electrophysio￾logical studies. An example is cortical area V5 (or MT) that has been extensively studied in the context of motion processing and perception64,65. Electrophysiology has shown that the vast majority of the V5 neurons in monkeys are direction and speed selective. Neuroimaging loca￾lized the homologue of area V5 in humans as an area responding stronger to moving than to stationary stimuli. Later studies suggested that human V5 is sensitive to motion direction, and that it may be thought of as containing large populations of directionally selective units, just like its monkey homologue. The studies of directional specificity exploited the phenomenon of motion after-effect induced by motion adaptation. After prolonged exposure to a stimulus mov￾ing in one direction, subjects perceive a subsequent static stimulus to move in the opposite direction. It is assumed that motion after-effect is due to the fact that the balance of mutual inhibition (opponency) between detectors for opposite directions of movement is distorted after adaptation. The sensitivity of the detectors selective for the adapting direction is reduced, which in turn releases from inhibition the neurons selective for the opposite direction66. Using this phe￾nomenon, human studies demonstrated that the fMRI response to a stationary stimulus was greater when the stimulus was preceded by a motion-after-effect-inducing, unidirectional adaptation, than when preceded by bidirectional adaptation67. Given the existing physiology data in the monkey V5, these findings were interpreted as demonstrating that the BOLD signal directly reflects direction￾selective spiking activity of the area. NATUREjVol 453j12 June 2008 REVIEWS 875 ©2008 Macmillan Publishers Limited. All rights reserved