Behavioral performance depends on attending to important objects in the environment rather than irrelevant objects. Regions in the right temporal–parietal junction (TPJ) are thought to be involved in redirecting attention to new objects that are behaviorally relevant. When subjects monitor a stream of distracter objects for a target, TPJ deactivates until the target is detected. We have proposed that the deactivation reflects the filtering of irrelevant inputs from TPJ, preventing unimportant objects from being attended. This hypothesis predicts that the mean deactivation to distracters should be larger when the subsequent target is detected than missed, reflecting more efficient filtering. An analysis of the blood oxygenation level–dependent (BOLD) task-evoked signals from 20 subjects during 2 monitoring tasks confirmed this prediction for regions in right supramarginal gyrus (SMG). Because the deactivation preceded the target, this mean BOLD-detection relationship did not reflect feedback from target detection or postdetection processes. The SMG regions showing this relationship overlapped or neighbored some regions associated with a “default” mode of brain function, suggesting the functional significance of deactivations in some default regions during task performance.
Because people can only respond to a small amount of the information present in the environment (Broadbent 1958), it is critical that perception and response are directed to behaviorally important objects. The attentional processes that ensure the perception of important objects and the filtering of unimportant objects have been widely studied, both behaviorally and physiologically (Treisman 1969; Eriksen and Hoffman 1973; Moran and Desimone 1985; Hillyard and Picton 1987; Rees et al. 1997; Reynolds et al. 1999; Pinsk et al. 2004; Serences et al. 2004). Sensory filtering is particularly necessary when people search for a target object in the presence of many distracting objects, such as when viewing a densely cluttered (Wolfe et al. 1989) or rapidly changing (Broadbent DE and Broadbent MHP 1987) environment. Under these conditions, top–down signals that specify the features of the target object may interact with sensory signals from distracters objects and prevent them from engaging higher level brain areas (Everling et al. 2002; Pinsk et al. 2004).
Recent studies have identified regions in right (R) temporal–parietal junction (TPJ, including the supramarginal gyrus [SMG] and superior temporal sulcus [STS] and superior temporal gyrus [STG]) that may be involved in the stimulus-driven selection of important objects in the environment (McCarthy et al. 1997; Linden et al. 1999; Downar et al. 2000; Marois et al. 2000; Kiehl et al. 2001; Corbetta and Shulman 2002). R TPJ is strongly activated by behaviorally important objects outside the current focus of attention that cause attention to be reoriented (Arrington et al. 2000; Corbetta et al. 2000; Macaluso et al. 2002; Kincade et al. 2005). When subjects monitor a central stream of objects for a target, peripheral distracters only activate R TPJ if they share a target feature, causing a shift of attention (Serences et al. 2005). The dependence of TPJ activations on behavioral relevance (Downar et al. 2001) implies that sensory signals reaching the TPJ have been filtered to remove irrelevant information, ensuring that stimulus-driven orienting does not occur to salient but unimportant objects.
Filtering may explain a surprising feature of the blood oxygenation level–dependent (BOLD) response in TPJ: TPJ is deactivated during search through displays containing only distracter objects, relative to a no-task baseline. We have (Shulman et al. 2003, 2004) proposed that this deactivation reflects the different settings of a filter for behavioral relevance. When subjects are not focused on a particular task, behavioral relevance is widely defined. Novel stimuli or stimuli with high sensory salience may be important to the individual and cause shifts of attention. Accordingly, under passive viewing conditions, TPJ is activated by any salient change in sensory stimulation (Downar et al. 2000). When a task is performed, however, behavioral relevance is mainly restricted to the small set of target objects in the task (Downar et al. 2001; Shulman et al. 2003; Serences et al. 2005). The same sensory changes that strongly activated TPJ under passive viewing now poorly activate TPJ, ensuring that unimportant stimuli do not disrupt attention to targets. This restriction in the range of stimuli that can activate TPJ may decrease the BOLD signal.
Although previous studies have measured activations/deactivations of TPJ, less is known about the relationship between these signals and behavioral performance. If large TPJ deactivations reflect efficient filtering of distracting information, then they should aid target detection. Here we tested that hypothesis by comparing the average magnitude of TPJ deactivation on trials in which a target was detected or missed. Critically, the TPJ deactivations occurred to nontarget displays (i.e., distracters) presented prior to the target, ensuring that any relationship of the mean BOLD signal to target detection did not reflect feedback from target detection or postdetection processes.
BOLD deactivations during task performance have been observed in many brain regions, including a subset that deactivate across a wide range of cognitive tasks relative to a low-level baseline condition (Shulman et al. 1997; Mazoyer et al. 2001) and have been associated with a “default” mode of brain function (Gusnard and Raichle 2001). The current hypothesis relating filtering to deactivations in the TPJ is based not simply on the fact that the TPJ deactivates but on the evidence, briefly summarized above, that the TPJ is activated by stimulus-driven reorienting to behaviorally relevant stimuli outside the focus of attention. Therefore, filtering is necessary to prevent stimulus-driven reorienting to inappropriate stimuli. Because deactivations per se could reflect a variety of processes and default regions include many whose function is unclear and do not have a documented relationship with stimulus-driven reorienting, the current hypothesis does not concern default regions in general. However, because regions in the TPJ overlap or neighbor some default regions, a secondary question was the anatomical relationship between regions that show a mean BOLD-detection relationship and default regions.
The current analysis was based on data from a previously published experiment (Shulman et al. 2003). Because a full description of the method for this experiment has been reported, an abbreviated description is given below.
Data were analyzed from 20 right-handed subjects, who gave informed consent in accordance with guidelines set by the Human Studies Committee of Washington University.
Stimuli and Procedure
Each subject performed 2 tasks. One involved the detection of a 300-ms period of coherent dot motion embedded in a sequence of dynamic noise, with the stimuli confined to a 3.25° circular aperture. Subjects were told the direction of coherent dot motion prior to each scan (defined by 45° increments), and this direction was changed over scans. The other task involved the detection of a digit among letters during a rapid serial visual presentation (RSVP) sequence. On each display frame of the RSVP sequence, 4 alphanumeric characters were positioned at the endpoints (eccentricity = 1.6°) of a virtual plus sign centered on fixation (Fig. 1). Each display frame was presented for 45 ms. The interstimulus interval separating display frames (i.e., the blank period between frames) was adjusted for each subject as described below. Subjects were told the target digit (from the digits 2 to 9) prior to each scan, and this digit was changed over scans. All distracter characters were letters. Both the noise sequence in the motion task and the RSVP sequence in the digit task lasted for 7.1 s, corresponding to three 2.36-s repetition time (TR) or magnetic resonance (MR) frames, on each of which a whole-brain image was collected. On 75% of the trials, a target was randomly presented between 580 and 1780 ms from the onset of MR frames 1 (“early-target” trial), 2 (“middle-target” trial), or 3 (“late-target” trial), corresponding to 1.18, 3.54, or 5.54 s on average from the start of the trial. All stimuli presented prior to the target were distracters. No target was presented on the remaining 25% of the trials. Because at most one target was presented on a trial, target detection corresponded to the functional end of the trial. Subjects pressed a button with their right hand as quickly as possible upon target detection and otherwise withheld a response. For each subject, the coherence level (percentage of dots in coherent motion) in the motion task and the interstimulus interval between successive display frames of the RSVP sequence in the digit task were determined in a behavioral presession so that the hit rate in both tasks was 70–80%. These values were occasionally adjusted between scans to maintain performance at a desired level. A scanning session included 16 scans, segmented into 4 groups of 4 scans. Each group involved a single task (digit, motion) that changed across successive groups. Task order was counterbalanced between subjects.
Magnetic resonance imaging scans were collected on a Siemens 1.5-Tesla Vision system, using an asymmetric spin-echo echo planar imaging sequence sensitive to BOLD contrast (T2*)(time repetition = 2360 ms, T2* evolution time = 50 ms, flip angle = 90°). One hundred and twenty-eight 2.36-s MR frames were acquired on each scan, where each frame included 16 contiguous 8-mm axial slices (3.75 × 3.75 mm in-plane). Structural images were collected with a sagittal magnetization-prepared rapid acquisition gradient echo 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°).
Preprocessing of BOLD Signals
Data were realigned within and across scans to correct for head movement. A whole-brain normalization, applied to each scan, corrected for changes in signal intensity across scans. Differences in the acquisition time of each slice in an MR frame were compensated by sinc interpolation so that all slices were aligned to the start of the frame.
The BOLD signal in each subject was analyzed with 2 types of regression models (general linear models [GLMs]). The first model made no assumptions about the shape of the hemodynamic response (HR) function or underlying neural activity and estimated a separate time course for each of the 8 trial types in a task: hit trials (early, middle, and late targets that were detected), miss trials (early, middle, and late targets that were missed), correct-rejection trials, and false alarm trials. This model (hereafter called the “no-assumptions” model, although like all GLMs it does assume linearity) has been extensively described and validated (Shulman et al. 1999; Ollinger, Corbetta, and Shulman 2001; Ollinger, Shulman, and Corbetta 2001) and yields an unbiased estimate of the time course for each trial type. Here we use the same GLM as in Shulman et al. (2003).
The time courses obtained from the no-assumptions GLM mix the effects of different neural processes because the BOLD signal in a region may be affected by the stimulus display, by searching for a target, by detecting a target, or by some combination of these processes. Shulman et al. (2003) computed a second 5-parameter regression model that specified the effects of stimulus, search, and 3 detection-related processes on the observed BOLD time course. This “process” model assumed that the observed BOLD response on each trial was the sum of component HRs that were generated by these processes and accurately accounted for the time courses from the no-assumptions model. In the present paper, the 5-parameter model of Shulman et al. (2003) was augmented to 7 parameters. This 7-parameter model was used to identify the voxels deactivated by search and compare the average search-related signal at those voxels on hit and miss trials.
The 7 processes are described below. A “stimulus” HR was generated by the 7.1-s stimulus display and was therefore the same on all trials. A “search” HR was generated for the duration for which the display was searched. The 5-parameter model of Shulman et al. (2003) contained a single search parameter that applied to all trial types, whereas the augmented 7-parameter model included 3 separate search parameters. A “search-hit” parameter was estimated for trials in which a target was presented and detected, a “search-miss” parameter for trials in which a target was presented and missed, and a “search-correct-rejection” (search-cr) parameter for trials in which no target was presented and no response was made. On hit trials, search was terminated when the target was presented/detected while on miss and correct-rejection trials, search was terminated at the end of the trial. A “hit” HR was generated by targets that were detected, a “miss” HR was generated by targets that were missed, and an “end-of-trial” HR (which was called a “correct-rejection” HR in Shulman et al. 2003) was generated by the functional end of a trial. Because only one target was ever presented on a trial, on hit trials the hit HR and end-of-trial HR were the same.
Figure 1B shows the temporal waveforms, that is, the time of onset and duration, for the 7 processes on an early-target hit trial (left panel) and an early-target miss trial (right panel). The stimulus waveform on both early-target hit and early-target miss trials corresponded to the 7.1-s duration of the display. The search-hit waveform on an early-target hit trial had a short duration because search was ended by target detection, whereas the search-miss waveform on an early-target miss trial corresponded to the duration of the trial. On a hit trial, a transient hit waveform occurred at the time the target was detected, whereas on a miss trial a transient miss waveform occurred at the time the target was presented and missed and a transient end-of-trial waveform occurred at the end of the trial.
The temporal waveform for each process, replicated over the entire set of trials in a scan, was convolved with a hemodynamic impulse function (Boynton et al. 1996) to produce the assumed HR for that process (e.g., the stimulus, search, or hit HR) over the scan. This assumed HR was entered into the design matrix of the GLM. Least-squares multiple linear regression was then conducted to calculate at each voxel the regression coefficient for that HR, indicating the degree to which the assumed HR for each process contributed to the observed signal at that voxel.
For each subject, a voxel-level map was formed from a contrast that combined the search-hit and search-miss parameters (i.e., the regression coefficients for the assumed search-hit and search-miss HRs) from the 7-parameter model. This map was put into atlas space (Talairach and Tournoux 1988) and smoothed by a filter with a full-width-at-half-maximum of 4 mm. A voxel-level one-sample t-test was conducted in which subject was treated as a random effect to determine if the contrast differed from zero over the group. This statistical map was corrected over the brain for multiple comparisons using a cluster size/z-threshold algorithm (Forman et al. 1995).
Regions of interest (ROIs) were then created for search-related regions in the left (L) and R TPJ. A peak-search algorithm identified local extrema in the uncorrected z-map and then consolidated foci closer than 12 mm by coordinate averaging. This consolidation replaced the cluster of points due to noise in spatially extended responses with a single locus at the center of mass. An ROI was created by defining a spherical region of radius 12 mm centered on a consolidated focus and excluding voxels not contained in the multiple-comparison corrected z-map (i.e., only voxels passing the multiple-comparison correction were included in the ROI). A voxel that met these criteria for 2 different ROIs was assigned to the ROI with the closest peak. This procedure has the effect of partitioning thresholded voxel-wise maps into disjoint regions centered on z-score peaks (Kerr et al. 2004). Regional group statistical analyses were conducted on these ROIs to determine if the search-hit parameter differed significantly from the search-miss parameter. In these regional analyses, a significance level of P < 0.01 was used, subject was treated as a random effect, and a correction was included for nonsphericity.
Subjects detected targets on 70.8% of the trials for the motion task and 79.6% for the digit task, a significant difference (F1,19 = 25.9, P < 0.001), but this task effect was most pronounced for early targets (i.e., targets presented near the start of the trial), as evidenced by a significant Task by Target Frame interaction (F2,38 = 4.03, P < 0.05). On late-target trials (i.e., a target was presented near the end of the trial), the hit rates for the motion and digit tasks were 75.9% and 81.4%, respectively, a significant difference (F1,19 = 6.37, P < 0.05). False alarm rates were low (2.7% for motion and 4.0% for digit) and did not differ between tasks (F1,19 = 1.92, P > 0.1). The low number of false alarms precluded a separate analysis of the BOLD data from those trials.
Time course of Bold Signal for Regions Showing a Search-Related Deactivation
The present analysis is confined to voxels that showed significant search-related deactivations, as determined by a one-sample t-test on a contrast that combined the search-hit and search-miss parameters from the 7-parameter model. These voxels are shown in Figure 2 by the cool colors. Although the quantitative analysis of search-related deactivations was based on the 7-parameter model, the time courses computed from the no-assumptions model provided a straightforward description of the deactivation that qualitatively supported the quantitative conclusions from the model. Therefore, we first describe the basic properties of the time course of the BOLD signal in regions that showed search-related deactivations and discuss qualitatively how these time courses related to the model parameters. The solid lines in the graph in Figure 1C show the time course from a region in R SMG for trials in which the target was presented and detected early, middle, or late in the trial (i.e., hit trials). On late-target hit trials, search proceeded for most of the 7.1-s display duration. Correspondingly, a large deactivation was evident in this region that was terminated when the target was detected, near the end of the trial.
The time course on early-target hit trials supported the conclusion that the deactivation on late-target hit trials was related to the duration of search rather than to the duration of sensory stimulation (see Shulman et al. 2003 for regions in which deactivations were maintained for the duration of sensory stimulation). On early-target hit trials, the search duration was brief because the trial was functionally over once the target was detected. Correspondingly, no overall deactivation was evident on these trials even though the display continued for the full 7.1 s. Following target detection on early-target hit trials, the signal returned to baseline rather than going below baseline while the stimulus display remained on the screen. Therefore, the deactivation was limited to the duration of search, rather than to the duration of the stimulus display, and was most evident in the time course on late-target trials.
The dotted lines in Figure 1C show the fit of the 7-parameter model directly to the observed time courses and confirm that the time courses were reasonably accounted for by the model (for more detail, see Shulman et al. 2003). The graphs in Figure 1D show the time courses of the 3 component processes (stimulus, search-hit, and hit-detection) that according to the model were separately present on early-target hit trials (left graph) and late-target hit trials (right graph). A component time course (e.g., search-hit) was computed by convolving the hemodynamic impulse function with the duration of the corresponding cognitive process shown in Figure 1B and scaling the result with the appropriate coefficient from the model fit (e.g., the value of the search-hit parameter). These component time courses summed to produce the overall predicted time course, the solid squares in the graphs of Figure 1D (for numerically exact results, the very small constant term from the model fit for this region, −0.013, also needs to be added to the sum). The overall predicted time courses in Figure 1D are the same as the time courses shown with the dotted lines in Figure 1C.
On early-target hit trials, the summation of a small BOLD deactivation due to the brief search duration with a larger BOLD activation when the target was detected (the stimulus component in this region was negligible and slightly negative) resulted overall in the small, predicted activation on those trials rather than a deactivation. On late-target hit trials, the longer duration of search resulted in a more prominent search-related deactivation that added with the delayed activation component due to target detection to yield in the overall time course a deactivation followed by a small activation.
Regions in R SMG Showed a Larger Average Deactivation Preceding Detected than Missed Targets
We tested the hypothesis that in TPJ regions that showed search-related deactivations, the average magnitude of the deactivation was greater when the subsequent target was detected than missed. Figure 2 presents a z-map of voxels showing significant search-related deactivations, as determined by a contrast that combined the search-hit and search-miss parameters from the 7-parameter model. The map was very similar to that described previously using the search parameter from the 5-parameter model (Shulman et al. 2003). Search-related deactivations were observed bilaterally in SMG and STS, with more extensive deactivations in the right hemisphere. Deactivations were also noted in prefrontal cortex, including middle frontal gyrus (MFG) and inferior frontal gyrus, again more extensive in the right hemisphere. Search-related deactivations were also observed (not shown) on the medial surface in anterior portions of early visual cortex and small regions in the precuneus and posterior cingulate, and ventrally in posterior fusiform cortex. All these regions also showed significant activations related to detecting a target and/or the functional end of the trial (as noted above, because only one target was ever presented on a trial, target detection functionally ended the trial).
The graphs in Figure 2 show the time courses of the BOLD signal in several regions of R SMG, STS, and postcentral gyrus on late-target trials in which the target was detected (black solid lines) or missed (black dotted lines). In several regions, the average deactivation on these trials appeared larger when the subsequent target was detected, suggesting that search-related deactivations were larger on hit than on miss trials.
A random effects 2-factor ANOVA with Search Parameter (hit and miss) and Task (motion task and digit task) was conducted to determine if the search-hit parameter from the 7-parameter model was significantly different from the search-miss parameter in the L and R TPJ regions shown in Figure 2. The search parameter was significantly more negative (P < 0.01) on hit than miss trials in 3 of the 10 ROIs comprising L and R TPJ (including SMG, STS, and STG). All were in R SMG and are outlined in red and indicated by an asterisk in Figure 2 (52, −49, 26: F1,19 = 8.90, P = 0.0077; 37, −49, 27: F1,19 = 9.49, P = 0.0062; 45, −49, 46: F1,19 = 8.65, P = 0.0086). The interaction of this effect with task did not reach significance at the 0.01 level, but there was a trend for a larger effect with the digit task than with the motion task. Similar hit-miss differences were observed in a more posterior region of R SMG, extending into the angular gyrus (AG), but did not reach the 0.01 significance level (43, −58, 37; F1,19 = 5.92, P = 0.025, black outline in Fig. 2). Regions anterior to R SMG in the ventral part of the postcentral sulcus/gyrus (e.g., 55, −34, 35, black outline in Fig. 2) showed no difference in the magnitude of the deactivation preceding hit and miss trials (F1,19 = 1.05). No regions in L TPJ were significant, although a region in L SMG approached the 0.01 significance level (−51, −43, 39: F1,19 = 6.85, P = 0.017). These results did not depend on the specific process model used to define ROIs. ROIs for search-related deactivations were also defined using the voxel-level z-map computed from the search parameter of the 5-parameter model. When the search-hit and search-miss parameters from the 7-parameter model were compared in these ROIs, significant differences in the search-hit and search-miss parameters were observed in 3 similar foci in R SMG (51, −50, 27: F1,19 = 8.93, P = 0.0076; 37, −44, 25: F1,19 = 11.87, P = 0.0027; 48, −49, 43: F1,19 = 8.29, P = 0.0098).
Consistent results were also observed in a voxel-level analysis of the mean BOLD-detection relationship, in which the search-hit and search-miss parameters from the 7-parameter model were compared in a random effects voxel-wise paired t-test. The resulting z-map was multiple-comparison corrected and masked so that only voxels showing significant search-related deactivations were included. This analysis yielded significant foci in R SMG (53, −45, and 28) and R AG (37, −69, and 36).
The mean BOLD-detection relationship was demonstrated within subjects, rather than across subjects (i.e., in the latter case subjects who showed larger mean deactivations would show better performance). We believe that a within-subject analysis provides a stronger confirmation of the hypothesis because it controls for extraneous variations in the BOLD signal across subjects. Moreover, task parameters were adjusted for each subject to ensure that performance across subjects and tasks was roughly equivalent, making across-subject correlations uninformative in the present case.
These results indicate that the average magnitude of the search-related deactivation in R SMG was larger preceding a detected than missed target, consistent with the filtering hypothesis. Critically, because the deactivation preceded target onset and signals resulting from target detection or trial offset activate rather than deactivate regions in SMG, this effect cannot be explained by feedback from target-related processes.
The Mean Bold-Detection Relationship Was Not Generally Observed in Other Regions
We primarily analyzed the search-related deactivations in regions within the TPJ because, a priori, these were the critical regions for the filtering hypothesis. However, in post hoc analyses, we also examined whether similar effects occurred in any other region that showed a search-related deactivation. Significantly larger deactivations preceding hits than misses were found in a region in R MFG (39, 34, 31: F1,19 = 8.69, P = 0.0084), which is outlined in red in Figure 2, with a marginal effect in a second R MFG ROI (40, 9, 49: F1,19 = 7.94, P = 0.011). Therefore, although a significant mean BOLD-detection relationship was not present in most regions that showed search-related deactivations, it was not completely restricted to regions that have been previously associated with the filtering hypothesis. The relationship in additional ROIs may reflect their involvement in a filtering process or perhaps some other function.
We also analyzed regions showing search-related activations in order to determine whether the mean search-related activation was greater when the subsequent target was detected than missed. No significant effects were observed.
Relationship of Regions Showing the Mean Bold-Detection Relationship to Default Regions
The regions in R SMG that showed a significant relationship between the average BOLD deactivation and target detection were adjacent to or overlapped default regions (Gusnard and Raichle 2001; Raichle et al. 2001).
Figure 3 superimposes the significant regions in R SMG on a flat map of right hemisphere default regions (the envelope of the borders of the significant ROIs is outlined in red, and the black outlines indicate the outlined ROIs from Figure 2 that did not show a significant relationship). The default regions were taken from the meta-analysis of positron emission tomography studies by Shulman et al. (1997), which identified regions that deactivated consistently across cognitive tasks with respect to a passive viewing baseline that held sensory stimulation constant. The magnitude difference image for the active task conditions minus passive viewing conditions from that study (n = 132 subjects) has recently been recomputed (Buckner et al. 2005). Although the degree of overlap between regions showing a significant mean BOLD-detection relationship and default regions depended on the magnitude threshold used for the default region image, parts of ventral SMG showed both a significant BOLD-detection relationship and a strong deactivation with respect to a passive viewing baseline.
Default regions showed a much more widespread anatomical distribution than those showing the mean BOLD-detection relationship. Some default regions that showed search-related deactivations, including regions in L SMG, did not show a significant BOLD-detection relationship. Other default regions did not show search-related deactivations, which was the criterion for inclusion in the present analysis, but exhibited other task-evoked temporal profiles that likely reflected a different functionality.
For example, we showed in prior work that regions in inferior parietal lobule (IPL) that were mostly posterior (e.g., AG extending to superior occipital gyrus) to the search-related ROIs showed stimulus-related deactivations (Shulman et al. 2003), with the deactivation extending for the entire 7.1 s of stimulus presentation rather than being terminated by target detection. Regions in the AG/superior occipital gyrus also showed small or absent end-of-trial- or detection-related signals, again distinct from SMG regions showing search-related deactivations. A recent paper by Golland et al. (forthcoming) reported a related result. Golland et al. (forthcoming) defined an “intrinsic system” that consisted of regions that showed low temporal correlations between repeated viewings of the same movie and which largely mirrored default regions. This intrinsic system was compared with a set of regions that deactivated during a relatively undemanding block visual localizer task that was used to identify object-selective regions in occipital cortex. The robust stimulation from the localizer task did not deactivate anterior regions within the IPL segment of the intrinsic system but only more posterior regions. These results suggest functional heterogeneity of default or intrinsic regions within IPL.
The mean magnitude of the deactivation in R SMG was significantly larger on trials in which the subsequent target was detected than missed. This result is consistent with the hypothesis that the deactivation indexes the degree to which irrelevant stimuli are filtered from this region (Shulman et al. 2003).
Mechanisms Relating R SMG Activity to Subsequent Target Performance
Increased filtering of the sensory input to the R SMG may improve the accuracy of target detection by ensuring that people attend to objects that are more likely to be targets rather than distracters (Shulman et al. 2003, 2004). Corbetta and Shulman (2002) proposed that R TPJ is involved in stimulus-driven orienting, signaling dorsal regions involved in orienting attention that a behaviorally important object has occurred. Many studies have shown that orienting to an unexpected but behaviorally important stimulus strongly drives regions within R TPJ (McCarthy et al. 1997; Linden et al. 1999; Arrington et al. 2000; Corbetta et al. 2000; Downar et al. 2000; Marois et al. 2000; Kiehl et al. 2001; Macaluso et al. 2002). Similarly, R SMG is only activated by peripheral distracter stimuli if they share a feature with a foveal target (e.g., color) and cause an inappropriate shift of attention (Serences et al. 2005). Critically, regions that are activated during stimulus-driven orienting overlap extensively with those that are deactivated during search (Shulman et al. 2003). This overlap supports the idea that deactivation/filtering is particularly important for regions involved in stimulus-driven orienting because it prevents inappropriate shifts of attention to distracters.
Todd et al. (2005) reported that increases in visual short-term memory (VSTM) load are correlated with greater mean deactivations in R SMG and suggested that increased VSTM load suppressed SMG activity in order to prevent orienting to irrelevant stimuli. The observation of TPJ deactivations during both monitoring of an RSVP stream (Shulman et al. 2003) and maintenance of information in VSTM (Todd et al. 2005) likely reflects the fact that orienting to or detecting irrelevant stimuli can impair performance on a variety of tasks. It is also interesting that TPJ deactivations occurred in the presence (Shulman et al. 2003) and absence (Todd et al. 2005) of visual stimulation, although in both experiments, ambient stimulation (e.g., scanner noise) was present throughout the scanning session.
In a separate behavioral experiment, Todd et al. (2005) showed that larger memory loads produced greater interference with detection of a target stimulus. Therefore, changing the task induced a relationship between mean BOLD activity and performance. Here, we showed a relationship between mean TPJ activity and subsequent behavioral performance even when the task was held constant.
Given the large expanse of cortex over which search-related deactivations have been observed, from STS to dorsal IPL, stimulus-driven selection is probably only one of many processes that are engaged by deactivated regions. Significant mean BOLD-detection effects were concentrated in R SMG. The significant SMG focus from the voxel-level analysis (53, −45, and 28) was located in ventral SMG and corresponded closely to the ventral SMG ROI at 52, −49, and 26 (4.6 mm distant from the voxel-level focus). Interestingly, regions of ventral R SMG are also activated by peripheral distracters that match the color of a foveal target (Serences et al. , 55, −44, 25, indicated by the red disk in Fig. 3 and 5.9 mm distant from this same ventral SMG ROI focus) and are deactivated by increases in VSTM load (Todd et al. , 59, −47, 24, indicated by the black disk in Fig. 3, which overlaps the red disk and is 7.5 mm distant from the same SMG ROI focus). Moreover, in the work of Corbetta et al. (2000), greater activations to invalid than valid targets were most consistently localized across experiments in ventral rather than dorsal SMG (Corbetta et al. , 53, −49, 30, 4.1 mm distant from the same R SMG focus and Kincade et al. , 51, −51, 26, 2.2 mm distant from the same focus). A regional analysis in which the SMG ROI centered at 52, −49, and 26 was applied to both of those data sets confirmed that this ROI showed a highly significant effect of validity in both experiments (Corbetta et al. 2000, P < 0.0001; Kincade et al. 2005, P = 0.0005). This convergence of results suggests that stimulus-driven orienting principally engages the ventral part of R SMG.
The filtering hypothesis maintains that deactivations occur during search because the input to R SMG is regulated but does not specify how this filtering is achieved. If filtering originates in early visual areas (Rees et al. 1997; Tootell et al. 1998; Pinsk et al. 2004) (for filtering in primary somatosensory cortex, see Drevets et al. 1995), a mean BOLD-behavior relationship should also be evident in middle- or upper-tier visual processing regions. Although the post hoc regional analysis yielded no significant effects at the 0.01 level in occipital cortex, many of the occipital search-related deactivations may have occurred in regions that were not critical to the task. Within medial occipital cortex, for example, search-related deactivations occurred anteriorly (Shulman et al. 2003), corresponding to parts of retinotopic cortex that coded the unstimulated periphery (Tootell et al. 1998).
Role of Filtering in Producing Deactivations from a Resting State
As noted earlier, a set of regions deactivate across a wide range of cognitive tasks relative to a resting or passive viewing baseline (Shulman et al. 1997; Mazoyer et al. 2001). One possibility is that these deactivations reflect ongoing processes during rest (e.g., unconstrained verbal thought processes, monitoring of the external environment, body image, and emotional state) that are “suspended during active tasks” (Shulman et al. 1997). Many related ideas and other candidates for ongoing processes have been proposed (e.g., Andreasen et al. 1995; Binder et al. 1999; Gusnard et al. 2001; Mazoyer et al. 2001; McKiernan et al. 2003, 2006). Raichle et al. (2001) demonstrated that during resting conditions the brain is metabolically in an equilibrium state in which the oxygen extraction fraction is relatively uniform and argued for a “default state” of the brain during rest (Greicius et al. 2003) characterized by ongoing processes (Gusnard and Raichle 2001).
The filtering hypothesis adds another perspective on ongoing processes. Broad monitoring of the environment is incompatible with monitoring of one or more foveal streams for a particular target. Some deactivations during active tasks may reflect changes in the selectivity or state of a mechanism that operates during both active tasks and resting states (e.g., high selectivity during the active task, low selectivity during rest), rather than the addition of processes during the rest that are suspended by the task. The relatively greater activity in R SMG during rest than during the monitoring phase of a task may maintain receptivity during rest to a broad class of events.
The current results provide evidence for the functional significance of the deactivation of some default regions during an active task (i.e., a relationship between the mean BOLD deactivation and target detection when the task is held constant). The ventral R SMG regions that showed a significant mean BOLD-behavior relationship overlapped with default regions as defined from a previous meta-analysis (Fig. 3). Hester et al. (2004) measured the responses to cues that indicated whether a “lure” requiring the withholding of a response would be subsequently presented in a go–nogo task, using assumed response functions within a mixed block and event-related design. They reported that deactivations in the right posterior cingulate/precuneus, left insula, and left cingulate/medial frontal gyrus were significantly larger to cues that preceded successful inhibition of motor responses to the lures than errors of commission. Noting the similarity between their regions and default regions, they argued for the functional significance of default region deactivations (Hester et al. 2004). Weissman et al. (2006) also noted that several regions in which larger deactivations over trials were associated with faster target reaction times appeared similar in location to default regions, although the deactivation–reaction time relationship in this study reflected a BOLD deactivation that occurred at/after target presentation rather than before the target.
It is interesting that the SMG regions from the present study that appeared to overlap default regions differed completely from those of Hester et al. (2004). The current monitoring task emphasized processes related to stimulus selection, whereas that of Hester et al. (2004) emphasized response inhibition, likely accounting for the different regional distributions in the 2 experiments.
When subjects monitor a stream of objects for a target, the mean BOLD deactivation preceding the target in R SMG is significantly greater when the target is detected than missed. We suggest that this effect arises because the deactivation reflects the degree to which the input to R SMG is restricted to potential targets, and this restriction ensures that attention is oriented to these candidates rather than to unimportant objects. This relationship is also consistent with a role for R SMG during nondirected or resting states in maintaining receptivity to a broad class of events.
This work was supported by NIH grants MH 71920-06 and NS048013. We thank T. Conturo, E. Akbudak, A. Snyder, and F. Miezin for software and hardware development. Conflict of Interest: None declared.