We report an endogenous signal that has a widespread cortical distribution and is time-locked to the termination of a sustained state of task-readiness. In three event-related functional magnetic resonance imaging (fMRI) experiments, subjects saw an arrow cue that predicted either the direction of motion or the location of a subsequent test stimulus. A reactivation of the BOLD (blood oxygenation level-dependent) signal occurred at the termination of the state of readiness in occipital regions that were transiently activated by the cue and in frontal-parietal regions that maintained an attentional set over the trial. Moreover, a delayed activation occurred in prefrontal and temporo-parietal regions that did not initially respond to the cue and that have been implicated in re-orienting attention to novel sensory events. These latter regions may have generated control signals that ended the state of readiness in regions active during the cue period. These results indicate that terminating a state of readiness produces a widely distributed cortical signal and suggest that areas involved in a preparatory state may be maintained as a network which can be modulated as a whole.
While performing a task, the brain passes through a series of states or processing stages that reflect different phases of the task. For example, if subjects are told to respond to a test object that will appear on their left, they may first orient their attention in that direction in expectation of the object and then process the sensory information from that object when it appears. There has been much interest in how these task states are prepared and maintained. Behavioral studies have examined this issue by measuring the performance costs produced by switching between task sets (Allport et al., 1994; Rogers and Monsell, 1995; Meiran, 1996).
In this paper we report an endogenous signal that is related to the termination of a task state. Several event-related functional magnetic resonance imaging (fMRI) studies have shown that a cue to orient attention to a location generates preparatory signals, prior to the presentation of the test stimulus, in parietal and occipital cortex (Kastner et al., 1999; Shulman et al., 1999; Corbetta et al., 2000; Hopfinger et al., 2000). Here we show that if the cue period is sufficiently long and is not followed by a test stimulus (i.e. the trial ends following the cue period), regions initially activated by the cue show a reactivation of the BOLD (blood oxygenation level-dependent) response when the trial is ended.
A striking aspect of the reactivation was its widespread cortical distribution. Reactivations were observed in virtually all areas initially activated by the cue, even those that responded transiently to the onset of the cue and had returned to baseline by the end of the trial. Therefore, the reactivation phenomenon may provide insight into how task states are maintained as a network that can be modulated as a whole, as well as how those states are terminated.
The reactivation phenomenon has been observed in three experiments. Some of the data from two of the experiments (experiments 2 and 3) have been previously published (Corbetta et al., 2000; Shulman et al., 2001). Experiments 1 and 2 were very similar. Both involved the presentation of an arrow cue during a cue period, which predicted the direction of coherent motion during a subsequent test period. The two experiments differed in terms of whether the duration of the cue period was constant (experiment 1) or variable (experiment 2) within a scan. Experiment 3 involved a paradigm in which the presentation of an arrow cue predicted the location of a stimulus during the subsequent test period.
Stimuli were generated by an Apple Power Macintosh computer and projected onto a screen at the head of the bore by a Sharp LCD projector. The screen was viewed through a mirror attached to the head coil. Behavioral responses (accuracy and reaction times, RTs) were recorded by a fiber-optic light-sensitive keypress held in the subject's right hand. During the behavioral session (see below) stimuli were presented on a Macintosh monitor and responses recorded by a Carnegie Mellon interface box.
Ten subjects were tested with event-related fMRI in a cued motion detection task. On each trial, subjects were shown a stationary arrow cue, superimposed on a circular patch of random dots, which indicated the direction in which a subsequent test stimulus (during the test period) would move (Fig. 1A). The arrow cue was presented for 1600 ms and then removed for the remainder of the cue period, in which only the stationary dots were present. In seven scans, the duration of the cue period was 4.72 s (two MR frames, where each frame is 2.36 s) while in another seven scans, the duration of the cue period was 9.44 s (four MR frames). Some trials ended following the completion of this cue period. As described previously (Shulman et al., 1999), these cue trials were necessary to isolate the signals present during the cue period. On the remaining trials, following the cue period, the stationary dots were randomly replotted each display frame (30 ms/frame) producing dynamic noise. On cue + noise/motion trials, a percentage of the dots were coherently moved for 300 ms (always in the direction of the previous arrow cue), 400–1800 ms after the onset of the dynamic noise. Subjects indicated that they had detected the motion by pressing an MR compatible keypress with their right hand. Following the coherent motion, all dots were randomly replotted for the remainder of the test period, which totaled one MR frame (2360 ms). On cue + noise trials, only dynamic noise was presented throughout the test period and subjects had to withhold a response. For all three trial types (cue, cue + noise, cue + noise/motion), the end of the trial was signaled by a change in the color of the fixation point from green (white in Fig. 1) to red (black in Fig. 1). The color changed from red back to green 1 s before the start of the next trial. Half of the trials were cue + noise/motion, a quarter were cue + noise and a quarter were cue, with the three types randomly mixed.
The coherence level (the percentage of dots that moved coherently on a single frame) for each subject was determined in a behavioral pre-session so that the hit rate for detecting motion targets was ~70–85%. During this pre-session, subjects also received blocks of trials involving both arrow cues and neutral cues (e.g. a plus sign cue that did not provide any information concerning the direction of coherent motion during the subsequent test period). As reported previously (Shulman et al., 1999), subjects detected more motion targets and responded faster following an arrow cue than a neutral cue, indicating that subjects were using the directional information provided by the arrow cue.
Twelve subjects were tested. There were several changes in the procedure. The cue period duration could be two, three, or four MR frames (for cue period durations of 4.72, 7.08, or 9.44 s), with the durations mixed within a scan. The test period was two MR frames (4.72 s) rather than one frame. On 1/3 of cue + noise/motion trials, the motion target was presented during the first MR frame (between 400 and 1860 ms from the start of the frame), on 1/3 of the trials, the target was presented during the second MR frame and on 1/3 of the trials, no target was presented. Each subject received 16 scans. Subjects were again tested in a pre-session to select an appropriate coherence percentage and verify that the cue information was being used to aid performance.
Thirteen subjects performed a modified Posner visual orienting task. Figure 1B shows the structure of different trials and their timing. The display consisted of a fixation cross (each arm of the cross subtended 16 min of visual angle) flanked on either side by two square boxes (box size 1°, eccentricity 3.3°). The beginning of a trial was marked by a change in the color of the fixation point from red to green, while the end of a trial was marked by a change in color back to red. Four types of trials were randomly presented. On ‘cue’ trials, subjects were presented with a small foveal arrow for 2.36 s. The arrow pointed either toward a left or right box, located at 3° on each side of the fixation point. The trial ended after the fixation point reverted to red. On ‘valid’ trials, an arrow was presented as on cue trials. The offset of the cue was followed by a random 1.5–3.0 s interval, after which a target (an asterisk flashed for 100 ms) was presented in the box indicated by the arrow cue (left box target for leftward arrow cue). On ‘invalid’ trials, the target was presented at the location opposite that indicated by the cue (right box for leftward arrow cue). Subjects were instructed to fixate, pay attention to the arrow cue and respond as fast and accurately as possible with a right hand key-press to the onset of the target. On ‘delay’ trials, the offset of the cue was followed by a 4.72 s delay in which no target was presented. The four types of trials were presented as follows: cue trials 20% of total number of trials; delay 20%; valid 44%; and invalid 16%. Each subject received 16 scans.
In all three experiments, the intertrial interval separating the end of one trial from the start of the next trial varied randomly between two, three and four frames. As described in our previous work (Shulman et al., 1999; Ollinger et al., 2001a), this random variation was introduced in order to estimate the BOLD response without assuming a particular hemodynamic response function or counterbalancing of trials. As a result of this procedure, the number of trials varied slightly on different scans. Trial onset was synchronized with the onset of the MR frame, rather than being evenly distributed (Price et al., 1999). It has been shown empirically (Miezin et al., 2000) that finer sampling intervals yield only a small improvement in signal estimation.
fMRI scans were collected on a Siemens 1.5 T Vision system, using an asymmetric spin-echo echo-planar sequence sensitive to BOLD contrast (T2*; frame duration = 2.36 s, T2* evolution time = 50 ms, flip angle = 90°) (Ogawa et al., 1990). During each scan, 128 frames of 16 contiguous 8 mm axial slices were acquired, with 3.75 × 3.75 mm in-plane resolution (Conturo et al., 1996). Structural images were collected with a sagittal MP-RAGE T1-weighted sequence (TR = 9.7 ms, echo time TE = 4 ms, flip angle = 12°, inversion time TI = 300 ms) and a T2-weighted spin-echo sequence (TR = 3800 ms, TE = 90 ms, flip angle = 90°).
Functional data were realigned within and across scans to correct for head movement, using six-parameter rigid-body realignment. A whole-brain normalization was applied to each scan to correct for changes in signal intensity between scans. Differences in the time of acquisition of each slice within a frame were compensated by sinc interpolation so that each slice was aligned in time to the beginning of each frame. For each subject, an atlas transformation (Talairach and Tournoux, 1988) was computed based on an average of the first frame of each functional run and the T2 and MP-RAGE structural images.
Analysis of BOLD Responses
The BOLD signal in each subject was analyzed with a linear regression model (within-trial model) that estimated separate time-courses during the cue and test periods for each trial type without making any assumption about the shape of the hemodynamic response (Shulman et al., 1999; Ollinger et al., 2001a,b). A second linear model (between-trial model) estimated separate time-courses on cue and cue + test trials (cue + noise and cue + noise/motion trials in experiments 1 and 2; valid, invalid and delay trials in experiment 3). Time-courses for a single period (e.g. the cue or test period) were therefore determined from the within-trial model, while time-courses over an entire trial (e.g. a cue + test trial) were determined from the between-trial model. In practice, the time-course for the cue period from the within-trial model was similar to the time-course for cue trials from the between-trial model, while the time-course for the test period from the within-trial model was similar to the difference between the time-courses for cue + test and cue trials from the between-trial model. Both the within- and between-trial models also included terms for each scan for an intercept and linear trend.
The consistency of the estimated BOLD responses across subjects was evaluated by ANOVAs on both voxels and regions of interest (ROIs). The parameter estimates from the regression model for each subject were entered into the ANOVA and subject was treated as a random effect. When ANOVAs were conducted on ROIs, those ROIs were defined from voxels identified by an independent statistical test, thresholded at a z-score of 4.0. For example, ROIs for an analysis testing the difference between two conditions were first defined from the main effect of MR frame in a within-subject ANOVA on the time-course values for the two conditions. These ROIs were therefore not biased toward either condition. Regional ANOVAs were conducted using a significance level of P < 0.01. Ten time-points were entered into the ANOVAs for experiments 1 and 2, while eight time-points were entered for experiment 3, which involved a shorter trial duration.
Since experiments 1 and 2 involved very similar paradigms, the data from the two experiments were pooled to determine the ROIs. In order to decrease the effects of anatomical variability, the time-courses for each individual were put into atlas space (in which each voxel is 2 × 2 × 2 mm) and then smoothed to 4-voxel FWHM (8 mm). The main effect of MR frame in this analysis selects those voxels that, averaged across the two experiments, show a time-course that is significantly different from a flat line. This image was then thresholded at a z-score of 4.0 (P < 0.00003 uncorrected). ROIs were defined from this thresholded image and a second ANOVA, with frame and cue period duration as factors, was conducted on the time-course values for each condition averaged over each ROI.
Some of the ANOVAs described in the paper were conducted using this two-step procedure, in which ROIs were first defined from the appropriate main effect of frame in a within-subjects voxel-based ANOVA and multi-factor within-subject ANOVAs were then conducted on the resulting ROIs. In other cases, multi-factor within-subject ANOVAs were conducted at the voxel level.
Description of the Reactivation Phenomenon
As reported in earlier work, a set of frontal, parietal and occipital regions was activated during the cue period (Shulman et al., 1999). Figure 2 shows the time-course of the BOLD response during the two-frame and four-frame cue periods (within-trial model) of experiment 1 in several regions defined from the main effect of MR frame in that experiment. In both the two-frame and four-frame conditions, an initial response was observed, reflecting the onset of the cue, which peaked at frame 3 and then fell off. However, in the four-frame condition a second response was observed which peaked at frame 7 (shown by the black arrow), three frames following the end of the cue period.
This second response, which we call a reactivation, is an endogenous component. The only sensory event that occurred at the end of the cue period on cue trials was the change in the color of the fixation point from green to red. The same change occurred in the two-frame and four-frame condition, yet the reactivation in the two-frame condition, which would occur on frame 5, was virtually absent (e.g. R vIPs).
Two hypotheses were formed concerning the cause of the reactivation in the four-frame condition. First, since the duration of the cue period in experiment 1 was blocked, there was little uncertainty concerning the onset of the test stimulus. Therefore, when the cue period was long in duration, cue regions might have been reactivated toward the end of the cue period in anticipation of the test stimulus. This hypothesis suggested that a reactivation would not develop in experiment 2, in which cue duration was mixed and subjects were unsure of when the test stimulus would occur. Secondly, the reactivation might have reflected the termination of a state of readiness maintained during the cue period. The reactivation was initiated following the end of the trial. Analyses described below provide support for this latter hypothesis.
Joint Analysis of Experiments 1 and 2
A joint voxel-level analysis of experiments 1 and 2 was conducted to determine which regions in the brain showed a reactivation that was consistent across the two experiments. The reactivation was identified by testing for a significant interaction of cue duration (two frames — 4.72 s; four frames — 9.44 s) and MR frame during frames 5–9. These frames reflected the time period in which the reactivation was observed in the four-frame condition, but not in the two-frame condition. The interaction term identified all voxels in which the shape of the hemo-dynamic response over those MR frames was different in the two-frame and four-frame cue conditions. A between-subject factor of cue mode (blocked cue durations — experiment 1; mixed cue durations — experiment 2) was also included to determine if the reactivation depended on whether the end of the cue period could be anticipated.
In order to determine those regions initially activated by the cue, irrespective of cue duration, a voxel-level ANOVA was also conducted on MR frames 1–5. The main effect of MR frame identified those voxels that showed a significant change in the BOLD signal over those frames.
The top row of Figure 3 shows those regions that were significantly reactivated (based on the interaction term of the ANOVA), while the bottom row of Figure 3 shows those regions that showed a significant initial cue response. Almost all regions that were initially activated by the cue showed a reactivation. Although Figure 3 appears to show some exceptions to this generalization, such as L lateral occipital (lat occ), these discrepancies were an artefact of the high threshold needed for statistical significance in the voxel-level analyses (see below). The term ‘reactivation’ is restricted in the rest of the paper to regions that showed this second BOLD response following the initial response to the cue (where this initial response was confirmed by a significant main effect of MR frame over frames 1–5 in a regional ANOVA). However, regions were also observed (e.g. right STg, SMg and dorsal inferior frontal gyrus, and bilateral anterior insula) that did not initially show an increased response to the cue, but showed a delayed activation which occurred on the same MR frame in which the reactivation was observed (see below). In the rest of the paper, activations that occurred at the same MR frame as a reactivation but were not preceded by an initial activation to the cue are called ‘delayed activations’.
Time-course of Regions which showed a Reactivation
These regions were activated during the initial establishment of the attentional set generated by the cue. The left two columns of Figure 4 shows the time-course (within-trial model) of the BOLD response in the intraparietal sulcus (IPs) and its extension into the occipital lobe (vIPs), as well as regions in the occipital lobe (e.g. lateral occipital, MT+; see left column of Table 1 for Talairach coordinates and z-scores). In order to show the reactivation across experiments more clearly, the time-courses in each condition were normalized by their magnitude on frame 3, which therefore was set to 1. The reactivation in the four-frame cue condition peaked at frame 7 (shown by the black arrow), three frames following the end of the cue period, and occurred both when the cue durations were blocked (experiment 1) and when they were mixed (experiment 2).
There also was some tendency in experiment 2 for a reactivation in the two-frame condition on frame 5 (three frames following the end of the cue period, indicated by the gray arrow; see RIPs, RvIPs, and MT+ in Fig. 4), although this effect was variable. This reactivation was shown by the increase of the BOLD signal in the two-frame condition over the four-frame condition at frame 5. An explicit comparison of the BOLD signal on frame 5 of the two-frame cue condition between experiments 1 and 2 in those regions showing a significant reactivation yielded a significant difference only in R vIPs [F(1, 20) = 15.9, P < 0.001].
Almost all regions showing an initial positive cue response also showed a reactivation, indicating that the regions activated as part of the task set were modulated as a whole. This was tested formally by taking all the ROIs defined from the initial cue response image (e.g. the ROIs formed from the image shown in the bottom row of Figure 3) and testing for an interaction of cue duration and MR frame for frames 5–9 (i.e. the interaction term that defines the reactivation image displayed in the top row of Fig. 3). A significant interaction was found in all regions, except in the right anterior fusiform gyrus [F(4, 80) = 1.34, P > 0.2], shown in the top left panel of Figure 5. Although the reactivations in several regions, such as SMA and FEF (not shown), appeared to be stronger in experiment 1 than experiment 2, these differences were not significant.
In summary, regions that were activated during the establishment of a preparatory set were reactivated when that set was terminated. This reactivation was most evident when the set was maintained over a long duration.
Time-course of Regions which showed a Delayed Activation
The voxel-level analyses in Figure 3 suggested that some regions that were not initially activated by the cue showed a delayed activation that occurred at roughly the same time as the reactivation. These regions were not activated during the initial establishment of the attentional set generated by the cue. The left two columns of Figure 6 show the time-course (within-trial model) of the BOLD response in six regions showing this pattern (see right column of Table 1 for Talairach coordinates and z-scores). While the initial cue response was weak or absent in these regions, the BOLD signal in the four-frame conditions showed a delayed activation on frame 7 (indicated by the black arrow), three frames following the end of the cue period. Moreover, the signal in the four-frame conditions showed an initial decrease in the BOLD signal that bottomed-out at frame 5 (indicated by the light gray arrow), two frames prior to the peak of the reactivation on frame 7. These observations were confirmed statistically. The initial response in frames 1–3 was only significant in R dIFg [F(2,40) = 6.06, P = 0.005], with a weak effect in R ant Insula [F(2,40) = 4.62, P < 0.05], while the decrease in the signal over MR frames 3–5 was statistically significant for all regions.
The delayed activation was larger in the four-frame condition than in the two-frame condition, similar to what was observed for regions that showed an initial cue response. This observation was confirmed statistically. For the four-frame cue condition, the delayed activation was defined as the difference in the BOLD response from frame 5 to frame 7, e.g. from the trough of the decrease to the peak of the positive response, three frames following the end of the cue period. Analogously, the delayed activation in the two-frame condition was defined as the difference in the BOLD response from frame 3 to frame 5. A regional ANOVA with region and cue duration as factors yielded a significant effect of cue duration [F(1,20) = 16.5, P < 0.001] and a significant interaction of cue duration by region [F(5,100) = 6.4, P < 0.0001]. The effect of cue duration indicated that the delayed activation was larger in the four-frame than two-frame condition, while the interaction with region indicated that the magnitude of this effect varied significantly across regions. Separate tests on individual regions indicated significant effects of cue duration in all six regions: anterior cingulate/pSMA [F(1,20) = 7.39, P < 0.05]; dIFg [F(1,20) = 28.1, P < 0.0001]; R SMg [F(1,20) = 7.7, P = 0.01]; R STg [F(1,20) = 71.7, P < 0.0001]; R ant insula [F(1,20) = 8.81, P < 0.01]; and L ant insula [F(1,20) = 7.0, P < 0.05].
Finally, although the delayed activation was larger in the four-frame condition than the two-frame condition, the delayed activation in the two-frame condition was significant. An overall ANOVA comparing the BOLD signal at frames 3 and 5 of the two-frame cue condition, with region (i.e. the six regions in Fig. 6) and experiment (1, 2) as factors yielded a significant overall reactivation [F(1,20) = 11.2, P < 0.005]. This effect did not interact with region, indicating that the magnitude of the delayed activation was equivalent over the six regions.
In summary, delayed activations were observed for regions that showed a weak or absent initial cue response. In the four-frame condition, these regions showed a significant decrease in the BOLD signal prior to the delayed activation. Moreover, the delayed activation was larger in the four-frame condition than the two-frame condition. However, a significant delayed activation was also observed in the two-frame condition.
Reactivations and Delayed Activations were not Observed in all Active Regions
As noted earlier, the reactivation in a right fusiform region was not significant (Fig. 5). The middle panel of Figure 5 shows the time-course (within-trial model) of the BOLD signal in a left central sulcus region that was active during the subsequent test period. No delayed activation was evident and, correspondingly, the interaction of cue duration (2,4) and MR frames (5–9) was not significant [F(4,80) = 0.58]. The right panel of Figure 5 shows the four-frame data from experiment 1 on both cue trials and cue + test trials (between-trial model), in which the four-frame cue period was followed by a one-frame test period. Since the test period in experiment 2 was two frames and the timing of the BOLD response to a keypress depended on when the keypress was made (i.e. frame 1 or frame 2), the time-courses from experiment 2 have not been presented. Left central sulcus was well activated by the right hand keypress during cue + test trials, even though no delayed activation was observed on cue trials.
Reactivations in the Spatial Attention Task (Experiment 3)
The joint analysis of experiments 1 and 2 enabled the regions showing reactivations or delayed activations to be well-defined from a large sample of subjects. The absence of an interaction with experiment in the joint statistical analysis indicated that there was not a significant difference in the reactivations or delayed activations between experiments 1 and 2. Moreover, the time-courses in the left panels of Figures 4 and 6 showed that the same general effect was qualitatively present in both experiments. However, these results do not constitute an independent replication of the phenomena, which is now presented for the data of experiment 3.
These analyses determined if the reactivations and delayed activations observed during experiments 1 and 2 were also observed in the same regions during experiment 3 (i.e. the regions applied to experiment 3 were defined from experiments 1 and 2), which involved a different group of subjects as well as a different task (attending to location as opposed to motion). The results shown below indicate that significant reactivations and delayed activations were present in this experiment and were most prominent on delay trials. On these trials, the cue period was followed by a blank two-frame period in which the subject maintained a state of readiness at a location for a target that was never presented (Fig. 1B). This pattern was consistent with that observed for experiments 1 and 2, in which the most robust reactivations and delayed activations on cue trials occurred following a long cue delay (e.g. the four-frame conditions).
Time-course of Regions that Reactivate following an Initial Cue Response
The right two columns of Figure 4 show the time-courses (between-trial model) from experiment 3 on cue trials and delay trials at the same reactivation regions defined from experiments 1 and 2. On delay trials, the time-course showed an initial response to the arrow cue that peaked at frame 3. This first peak was followed by a decrease in the response and then a second peak or reactivation at frame 6, three frames following the end of the test period. Therefore, the signal increase on frame 6 of delay trials was a reactivation whose relative time-course was the same as that observed in experiments 1 and 2. The reliability of the reactivation was tested by an ANOVA that examined the main effect of MR frame on delay trials for frames 4–8, where frame 4 is the first frame following the end of the target period. This interval was analogous to the interval (MR frames 5–9) in which the reactivation was observed for the four-frame cue condition of experiments 1 and 2. This ANOVA indicated whether the time-course over these frames differed from a flat line and was applied to each of the reactivation regions determined from experiments 1 and 2. Significant reactivations were observed in all regions. While a robust reactivation was observed on delay trials, the reactivation on cue trials (shown by the gray arrow) was weaker and more variable.
Time-course of Regions that show a Delayed Activation and Little or No Initial Cue Response
The two right columns of Figure 6 show that during cue trials, the time-course (between-trial model) of the BOLD signal in most of these regions was relatively flat, with little or no delayed activation (the peak of the delayed activation during cue trials should have been evident on frame 4, shown by the light gray arrow). In contrast, the signal on delay trials showed a delayed activation that peaked on frame 6, three frames following the end of the test period. Moreover, as in experiments 1 and 2 (see left columns of Fig. 6), the peak of the delayed activation on delay trials was preceded by a decrease in the BOLD signal.
These observations were confirmed statistically. An analysis of the BOLD signal over frames 1–4 during delay trials, with MR frame (1–4) and region (i.e. the six regions in Fig. 6) as factors yielded a main effect of MR frame [F(3,36) = 3.81, P < 0.05], indicating that the decrease in the BOLD signal over those frames was reliable, and a region by frame interaction [F(15,180) = 5.83, P < 0.00001], indicating that the magnitude of this change differed across regions. Analyses at individual regions indicated significant changes in dIFg [F(3,36) = 12.4, P < 0.00001] and R STg [F(3,36) = 9.36, P < 0.0005]. The decreases in R SMg and the L and R ant insula showed the same trend but were not significant. Although decreases in these three regions were significant in the joint analysis of experiments 1 and 2, the sample size in that analysis was larger (n = 22 for experiments 1 and 2, n = 13 for experiment 3).
The delayed activation on delay trials, defined by the effect of MR frame over frames 4–8, was significant in all regions: anterior cingulate [F(4,48) = 11.1, P < 0.00001]; R dIFg [F(4,48) = 15.2, P < 0.00001]; R SMg [F(4,48) = 13.4, P < 0.00001]; R STg [F(4,48) = 16.5, P < 0.00001]; L ant insula [F(4,48) = 14.3, P < 0.00001]; and R ant insula [F(4,48) = 21.7, P < 0.00001].
The peak of the delayed activation on delay trials occurred on frame 6, three frames following the end of the test period. The relative timing of this delayed activation was identical to that shown on the four-frame cue trials of experiments 1 and 2 and indicated that it was timelocked to the end of the trial. This hypothesis was strongly supported by a comparison of the BOLD signal on delay trials and valid trials, in which subjects responded to a target in the cue location during the test period. On average, this target was presented midway through the two-frame test period. The left panel of Figure 7 shows that the time-course of the BOLD signal on valid trials (between-trial model) peaked one frame prior to the signal on delay trials, consistent with a response timelocked to the onset of the target on valid trials and a response timelocked to the end of the delay period on delay trials.
In summary, the strongest reactivation and delayed activation in experiment 3 occurred at the end of the delay period rather than at the end of the cue period. This pattern was consistent with that observed for experiments 1 and 2, in which the most robust reactivation and delayed activation occurred following a long cue delay, while the weakest and most variable effects occurred when the cue duration was short and fixed in duration (the two-frame condition of experiment 1). In some regions that did not show an initial cue response, all three experiments showed a depression in the BOLD signal two frames prior to the peak of the delayed activation.
The current experiments support the existence of an endogenous brain response to the termination of a sustained state of readiness. On cue trials of experiments 1 and 2, subjects expected a test epoch involving dynamic noise, which did not occur. On delay trials of experiment 3, subjects expected a test stimulus, which was not presented. In both of these situations, reactivations were observed in regions that were activated during the task, including sensory processing regions in the occipital lobe, such as MT+, and parietal regions involved in the orienting of visual attention, such as IPs. Delayed activations were observed in other regions in prefrontal and temporal-parietal cortex, which did not show a positive BOLD signal during the task epoch. Many of these latter regions were right lateralized or showed a right hemispheric dominance (Fig. 3). Therefore, the reactivation and delayed activation had a widespread cortical distribution, indicative of a change of state across the brain when a state of readiness is terminated. In the following discussion, we review the evidence for this proposition and suggest some functional interpretations.
Reactivations are not a Passive Response to Sensory Stimulation
Several results indicated that the reactivations did not reflect a sensory response to the change in color of the fixation point. First, the reactivation in experiments 1 and 2 was dependent on the nature of the cue period. When the cue period was short and blocked, the same color change produced little or no reactivation (e.g. Fig. 1). Similarly, in experiment 3, the reactivation was more evident on delay trials than on cue trials, even though the same color change occurred in both. The weak reactivation on cue trials in experiment 3, which had a fixed duration of one frame, was consistent with the weak or absent reactivation during the short (two-frame), blocked cue trials of experiment 1. Therefore, the strength of the reactivation signal was affected by the duration of the state of readiness. Secondly, a delayed activation with identical timing was observed in all three experiments in regions that did not show an initial cue response. Therefore, these regions did not respond to either the onset of the arrow cue or the change in the color of the fixation point from red to green at the beginning of the trial, yet responded to a similar change at the end of the trial.
Reactivations do not Reflect the Expectation of the Target
An initial hypothesis was that the reactivation reflected a priming of the visual system just before the end of the cue period in anticipation of the test stimulus. Single unit studies have reported that when monkeys can anticipate the onset of a behaviorally relevant stimulus, cell activity in area 7a (Constantinidis and Steinmetz, 1996) and LIP (Colby et al., 1996) rises above spontaneous levels prior to stimulus onset. This anticipatory activity is observed even when the monkey cannot anticipate the location of the expected stimulus (Constantinidis and Steinmetz, 1996). According to the present ‘target-expectation’ hypothesis, the reactivation observed in the present experiment reflected the anticipatory activity noted in these studies.
This hypothesis, however, was not consistent with several results. First, only a weak reactivation was observed in the two-frame condition in which cue duration was fixed (experiment 1). Second, when the cue duration was mixed, subjects could not anticipate the onset of the test stimulus, preventing an anticipatory reactivation. Yet mixing the cue durations increased or maintained whatever reactivation was present in the two-frame condition. Third, the reactivation in experiment 3 was observed at the end of the test period, rather than at the end of the cue period, and therefore could not have reflected anticipatory activity related to the test period. Finally, the target-expectation hypothesis does not explain why a delayed activation was observed in regions that were not activated during the task (e.g. Fig. 6). These regions did not appear to play an important role in the task, yet were activated when it was completed.
Reactivations do not Reflect Passive Ongoing Processes during the Intertrial Interval
Since the reactivation and delayed activation occurred at the end of the trial, another hypothesis is that they reflected processes that occur during non-task resting states such as the intertrial interval (Shulman et al., 1997; Binder et al., 1999). However, the BOLD responses in the current work were transient, rather than being sustained throughout the intertrial interval. Moreover, the spatial distribution of the reactivations and delayed activations were quite different than the resting state activations, which are most prominent in midline dorsal and ventral frontal cortex and posterior cingulate. In fact, the only regions in the current study that even partly overlapped the regions which show decreases that generalize over tasks were found in the inferior parietal lobule.
Reactivations Reflect the Termination of Task States
Task states change constantly with changes in the environment and behavioral goals. One interpretation of the reactivation/ delayed activation is that it was caused by the cancellation of a state of readiness. On both cue trials and delay trials, subjects waited in readiness for an event that did not occur (i.e. on cue trials the test period was omitted, while on delay trials the target stimulus was omitted). Terminating this state may have produced reactivations and delayed activations for three general reasons. First, termination may have evoked transient processes related to the new ‘task’ state (i.e. the state appropriate for the intertrial interval), such as a reorientation of attention to the display or a change in arousal. Secondly, BOLD responses may have been produced by terminating processes related to the old state of readiness. For example, studies of the countermanding of eye movement commands have reported signals related to successful cancellation in frontal and supplementary eye fields (Hanes et al., 1998; Stuphorn et al., 2000).
Finally, some BOLD responses may have reflected the analysis of the stimulus signaling the end of the trial (i.e. the change in color of the fixation point) and the generation of appropriate commands for terminating the state of readiness. These commands may have been initiated by some of the regions that showed a delayed activation, with no positive response during the task epoch. These regions included a set of right hemisphere regions (e.g. SMg, STg, precuneus, lateral precentral) that have also been observed when subjects respond to invalidly cued targets (Corbetta et al., 2000). Some of these regions may have been responsible for initiating a cancellation signal that was sent to task-relevant regions or regions involved in maintaining or initializing the task set. The reception of the cancellation signal in these regions triggered a reactivation.
The end-of-trial signal — during cue and delay trials, the change in the color of the fixation point, during cue + noise trials, the offset of the dynamic noise — was a behaviorally task-relevant event. However, it is important to emphasize that other behaviorally relevant events did not activate these regions. These events included the onset of the arrow cue/change in fixation color at the start of all trials, the onset of the dynamic noise at the start of the test period on cue + noise and cue + noise/motion trials (experiments 1 and 2) and the offset of the arrow cue at the start of the test period on delay, valid and invalid trials (experiment 3). Each of these sensory events indicated an important transition point in the task but did not result in a BOLD response. These results emphasize that the reactivations and delayed activations were specific to events reflecting the termination of an active task state.
Earlier workers (Konishi et al., 2001) have observed transient signals at the beginning and end of blocks in block-design paradigms, in which an active task period is followed by a passive control period. Of the 34 foci listed in Konishi et al., the five foci with the largest z-scores corresponded to five regions in the present study that showed a delayed activation in the absence of an initial cue response (R lat PrCs, R STg, R precuneus, R ant Insula, anterior cingulate/preSMA). Konishi et al. interpreted their results in terms of regions involved in establishing an appropriate task set or shifting between task sets. This interpretation does not, however, apply to the present results. The delayed activations in the above regions were not observed during the transition from the intertrial interval to the cue period, or the transition from the cue period to the test period, but only during the transition from an active task period to the intertrial interval. It is possible that this discrepancy between Konishi et al. and the present study reflects a difference between blocked and event-related paradigms. The initialization of task states may not be as evident for individual trials within a series.
Reactivations Involve Modulations of Task-relevant Networks
Reactivations were observed in virtually all areas initially activated by the cue, even those that only responded transiently to the onset of the cue and had returned to baseline by the end of the trial. These latter areas, while no longer active, nevertheless remained linked as part of a task network that was modulated as a whole by the termination of the trial. This linkage may be part of a mechanism by which task states, involving the coordination of activity in widely disparate areas, are maintained. Synchronization of the neural activity in separate areas, modulated by task demands, has been reported in several studies (Varela et al., 2001).
The task dependency of the reactivation was suggested by the fact that reactivations were observed in almost all regions initially activated by the cue. This task dependency can be further examined in experiments that directly compare preparatory task sets that activate different regions.
Reactivations do not Reflect Low Probability Events
An alternate hypothesis to that presented above is that the reactivations and delayed activations reflected the unexpected presentation of the end-of-trial signal (the change in the color of the fixation point). For example, since cue trials occurred on only 25% of the total trials, the trial termination stimulus on these trials was relatively unexpected. However, this hypothesis is inconsistent with the results. In experiment 1, the probability of a cue trial in the two-frame condition was the same as the probability of a cue trial in the four-frame condition. Yet the reactivation in experiment 1 was almost absent for the two-frame cue trials (e.g. Fig. 2).
Explanations of the reactivation in terms of the momentary probability of the end-of-trial signal also do not explain the results of experiment 3. In experiment 3, cue trials occurred on 20% of the trials, delay trials occurred on 20% of the trials and target trials occurred on the remaining 60%. However, the momentary probability of the end-of-trial signal on delay trials was actually higher than on cue trials. Once the test period started, the probability that the trial would be a delay trial was actually 25% (i.e. 20/80). Moreover, on target trials, the target was randomly presented 1500–3000 ms into the test period, which lasted 4720 ms. No targets were ever presented in the final 1.7 s prior to the onset of the end-of-trial signal on delay trials. Since subjects received a separate behavioral practice session prior to the scanning session, these contingencies had been repeatedly experienced. As a result, by the end of the delay trial, the end-of-trial signal was not particularly unexpected. Yet the observed reactivations occurred at the end of the trial, one frame later than the activation on target trials (Fig. 7). Therefore, while the momentary probability of the end-of-trial signal on delay trials was certainly higher than the momentary probability of the end-of-trial signal on cue trials, reactivations/delayed activations were more robust on delay trials than cue trials, consistent with the difference between two-frame and four-frame cue trials in experiments 1.
Finally, several papers have examined the activations produced by low probability stimuli. The principal effects of these frequency manipulations were generally observed in prefrontal, IPL and temporal-parietal regions, the regions that showed delayed activations in the present work, rather than regions in occipital cortex, posterior SPL and posterior ventral and dorsal IPs, which showed reactivations (McCarthy et al., 1997; Linden et al., 1999; Braver et al., 2001).
However, the similarity of the regions showing delayed activations with those regions affected by target frequency is interesting and consistent with the basic argument of the paper. When a target stimulus is unexpected, subjects need to shift from their current task set or stimulus–response mapping to the task set or mapping appropriate to the unexpected target. Changes in task set, whether induced by low probability events or by trial termination signals, involve a similar set of regions.
The present analyses have uncovered an endogenous reactivation signal that is generated at the end of a state of readiness, particularly when the trial is extended. These signals reflected the shift from an active to an inactive task state. The most striking aspect of the reactivation signal was its widespread distribution, with components in frontal, parietal and occipital cortex. Task-relevant areas may maintain a linkage that allows them to be modulated as a whole.
As imaging techniques improve in spatial and temporal resolution, the observed complexity of the brain's response to a cognitive task will no doubt increase. This complexity was revealed in the variety of time-courses that were observed in the present analyses: transient activations, transient activations coupled with reactivations and signal decreases followed by delayed activations. Other types of time-courses not discussed in the present report, including decreases and biphasic responses, were also observed. This variety emphasizes the need for analytic techniques that do not make assumptions about the shape of the hemodynamic response (Josephs et al., 1997; Shulman et al., 1999; Corbetta et al., 2000; Ollinger et al., 2001a,b). In the current work, both the models for estimating time-courses and the statistical techniques for assigning significance to those time-courses, made no shape assumptions.
|Reactivation regions||Delayed activation reegions|
|L IPs||−35||−55||44||6.1||R SMg||59||−41||28||6.5|
|R IPs||33||−59||50||6.5||R STg||53||−47||18||7.1|
|R MT+||41||−65||−6||6.2||L ant insula||−35||13||2||5.5|
|L MT+||−43||−63||−8||6||R ant lnsula||37||17||2||5.9|
|R lat occ||33||−83||−4||6.3||33||21||8||5.5|
|R lat PrCs||43||−3||42||6.9|
|Reactivation regions||Delayed activation reegions|
|L IPs||−35||−55||44||6.1||R SMg||59||−41||28||6.5|
|R IPs||33||−59||50||6.5||R STg||53||−47||18||7.1|
|R MT+||41||−65||−6||6.2||L ant insula||−35||13||2||5.5|
|L MT+||−43||−63||−8||6||R ant lnsula||37||17||2||5.9|
|R lat occ||33||−83||−4||6.3||33||21||8||5.5|
|R lat PrCs||43||−3||42||6.9|
This work was supported by NIH grants EY12148 and EY00379. We thank John Ollinger, Tom Conturo, Erbil Akbudak, Abraham Snyder and Fran Miezin for software and hardware development.