Choosing Where to Attend and the Medial Frontal Cortex: An fMRI Study

Abstract

To investigate how we orient our spatial attention, previous studies have recorded neural activity while participants are instructed where to attend. Here we contrast this classical instructed attention condition with a novel condition in which the focus of voluntary attention is not specified by the experimenter but rather is freely chosen by the participant. Central cues prompted fixating participants either to choose which of two peripheral spatial locations to covertly attend or formed an instruction. Either type of cueing initiated selective attention demonstrated behaviorally by enhanced performance at a visual detection task in comparison to a separate divided attention condition. We used functional magnetic resonance imaging to measure which areas were more active during choice than instruction. Choosing where to attend activated a large cluster of medial frontal cortical regions similar to those that have been previously implicated in the free selection of overt action. We then addressed a potential confound in contrasting choice with instruction: participants may remember their behavior more when choosing. In a separate block, and interleaved with choice trials, “memory” trials were introduced in which participants were instructed to remember where they had attended on the previous trial. The presupplementary eye fields and lateral frontal eye fields were specialized for choice-guided attentional orienting over and above any memory confound. This evidence suggests a common mechanism may underlie free selection, whether for covert attention or overt saccades.

INTRODUCTION

The neural mechanisms controlling how we orient our attention are thought to be distributed across a network of cortical areas including the frontal eye fields (FEFs) and medial frontal (MFC) and posterior parietal (PPC) cortices (Mesulam 1981), but it is not clear how the functions of these areas may differ. One popular but elusive notion is that each area may be specialized according to the extent to which the external environment is relied on in determining the direction of attention. For example, it has been suggested that ventral temporo-parietal cortex may be most important when sudden salient changes in the external environment grab attention (externally driven or “exogenous”), whereas the FEF may be needed most when the participant orients attention according to a learned rule, which might be thought of as a more internally driven or “endogenous” process (Buschman and Miller 2007; Corbetta and Shulman 2002). However, meta-reviews have reported that the FEF and PPC are activated by either type of orienting (Grosbras et al. 2005; Nobre 2001), and it is particularly uncertain where the MFC might fit in to such a classification system.

Unlike in the field of attentional orienting, there is considerable evidence from studies of motor control that some neural regions may be specialized for the control of internally driven as opposed to externally driven behavior. Specifically, the MFC has been implicated in the initiation of actions which are unconstrained by the external environment, and “free selection” is possible. MFC is preferentially activated when participants freely select between alternative finger movements (Deiber et al. 1991; Frith et al. 1991), saccades (Nachev et al. 2005), and task sets (Forstmann et al. 2006; Walton et al. 2004).

The study presented here draws on the distinction drawn between the internal and external guidance of action that has been influential in the investigation of limb movements and tests its applicability to the understanding of attentional orienting. More specifically, the study tests the hypothesis that the MFC is involved in the free selection of where to orient visual spatial attention in the absence of external instructions, here referred to as “choosing where to attend.” We present a task in which participants are either instructed where to attend, or choose, on a trial-by-trial basis.

Studies of motor-selection have previously highlighted the contribution of additional working-memory components when participants engage in free-choice tasks (Hadland et al. 2001; Lau et al. 2004). To factor out such memory components and isolate neural activity specific to choosing where to attend, an additional “memory” block was performed in which participants are instructed to attend to the same or different location as in the previous trial or are free to choose.

METHODS

Nineteen right-handed participants (aged 19–33, 10 female) all had normal or corrected-to-normal vision and were native speakers of a Latin-alphabet language. Between 1 and 7 days prior to participating in an fMRI experiment, participants were trained in the laboratory. The aim of this training was to familiarize participants with the task, ensure that they could perform the task without moving their eyes, and therefore optimize later performance of the task in the scanner. Real-time infrared eye-tracking (iScan, Burlington, MA) was used to measure the position of gaze during task performance. One participant was excluded from the experiment at this stage due to being unable to fixate during the task. The remaining 18 participants proceeded to the functional magnetic resonance imaging (fMRI) experiment.

A schematic of the task is shown in Fig. 1. Each trial in the task started with a precue, after which two streams of twenty rapidly presented letters were displayed on either side of a fixation cross, inside peripheral squares, at an eccentricity of 4° of visual angle. On most trials, a target letter ‘a’ was presented once on each side within the streams. The task was to monitor covertly the letter streams according to the precue instructions and respond with a manual button press as soon as possible after presentation of the target “a”. A minority of trials (10%) in all conditions were catch trials in which no target letter a was presented. Participants were to refrain from responding when uncertain about whether the target letter had been presented. When present, the targets in the two streams appeared ≥1 s apart so that the reaction time indicated to which side the participant had been attending. The timing of the target letters within each sequence was counterbalanced within each block with the additional constraint that a target letter was never presented within the final (20th) pair. In all blocks, errors were defined as misses (not responding within 800 ms of a target) or false alarms. The order of trial types, the order of letters in the streams, and the position of the targets were pseudo-randomized and were identical for all participants.

FIG. 1.

Task schematic. On instructed trials, the “l” or “r” of the central precue was red, and participants were to covertly attend to the left or right stream of letters. But on choice trials, if the horizontal line was red (shown here in gray) then the participants themselves freely selected which side to attend. Participants responded when they detected the target letter “a”. This figure shows the spatial block; in the memory block the letters were “s” or “d” (instructing the participant to attend to the “same” or “different” side as the previous trial) rather than “l” or “r”.

During fMRI scanning, participants completed the three experimental blocks of the task: “divided attention”, “spatial”, and “memory”. The spatial and memory blocks were performed for the acquisition of both behavioral and functional data, constituting the experiment's factorial design. Demonstrating that participants were selectively orienting their attention on these blocks required testing for a behavioral facilitation on these trials relative to a condition that lacked selective orienting but was otherwise similar. Therefore for the collection of behavioral data, participants also performed a divided-attention block where on each trial participants attended to both sides of the screen at once and responded to the first target that appeared. fMRI volumes were still collected during the divided-attention block but only to ensure that all behavioral data were gathered under the same conditions, including scanner noise. To equate for trial numbers across conditions, the divided-attention block was half the length of the other two, which contained two conditions rather than one (see following text) and therefore lacked the power needed for an fMRI analysis. The order of the three blocks was counterbalanced across participants. Participants were given short breaks between blocks but remained in the scanner.

Divided-attention condition

The divided-attention block contained 50 trials. The precue consisted of the letters “l” and “r” separated by a horizontal line and was entirely white on every trial. On each trial, the participant was to attend to both sides at once and respond to the first “a” that was presented. Previous studies have shown that participants can divide their attention between two peripheral streams of rapidly presented letter strings (Duncan et al. 1997). Only responses to the first “a” were correct.

Spatial block

The spatial block contained 100 trials, and the trial type was indicated by a part of the precue being colored red. Half of the trials were instructed trials, according to an identical pseudorandom schedule for all participants, and were divisible into two subtypes. When the “l” was red (25 trials), the participant was instructed to attend covertly to the left stream and respond as quickly as possible when they saw an “a” on the precued side. Similarly, if the “r” was red (25 trials), the participant was to detect the “a” on the right. These two trial types formed the instructed condition. Reaction times (RTs) were only included in the median RT measure if the responses were correct and if the side that the participant was instructed to attend to was also the side on which the first of the two targets was presented. This was so that the data were comparable with the divided-attention condition, where only responses to the first target were considered. This precaution ensured that any differences in RT between conditions were not confounded by differences in the temporal interval between trial onset and response.

In the other 50 trials of the spatial block, the horizontal line between the letters in the precue was red, instructing the participant to choose, arbitrarily and at will, which side to attend. This is the choice condition. During training, the participants were instructed that they should choose randomly. No other guidelines were provided. Before each block, instructions appeared on the monitor screen reminding participants of the meaning of the cues: “red L = attend left, red R = attend right, red line = choose left or right side”. Again, RTs were only included in the median RT for correct responses where the chosen target was the first of the two targets in the trial to be comparable with instructed and divided-attention trials. The participants' performance determined the classification of all error and choice trials: the chosen side was determined from the behavioral reaction time, assumed to fall within 800 ms of chosen-target presentation.

Memory block

The memory block was similar to the spatial block except that instruction cues required maintenance of working memory for the previously attended side. The default precue consisted of the letters “s” and “d” with a line between them. Before the experiment, the participant was instructed to attend left on the first trial of the memory block. This trial was later discarded from the analysis. From then on, if the “s” was red (25 trials), the participant was to attend to the same side as on the previous trial. If the “d” was red (25 trials), the participant was to attend to the opposite side as on the previous trial. These two trial types formed the instructed condition of the memory block and differed from the instructed condition of the spatial block in that participants were explicitly asked to remember where they attended and to then use that memory as the basis on which to orient their attention on the next trial.

If the line was red (50 trials), the participant was to choose where to attend. This was similar to the choice condition of the spatial block, although unlike the spatial block, participants were explicitly asked to remember where they chose to attend.

fMRI acquisition

Participants responded using a custom-made fMRI-compatible response box. Stimuli were presented on the display screens of a binocular head-mounted fMRI-compatible eye-tracking system using combined infra-red illumination and recording from the right eye (Avotec), fiber-optic transmission of the data out of the magnet room (Silent Vision, SMI, Berlin, Germany) and on-line preprocessing of the data (iViewX, SMI, Berlin, Germany). The stimuli were focused by the participant until a crisp, fused image was reported. Eye position was recorded concurrently with stimulus presentation. Eye-tracker data were analyzed off-line with iLab software (Gitelman 2002). Trials were eliminated from the analysis if there were eye gaze deviations of >2° of visual angle from central fixation during the trial: these trials were excluded before errors and reaction times were calculated. Participants responding on more than half of catch trials or committing >50% errors in any block were rejected from all analysis. Seven participants were rejected according to these criteria. Data from the remaining 11 participants were used for behavioral and fMRI analysis.

Participants were scanned in a 3-Tesla human MRI system comprising a Siemens AS22 body gradient coil, a Magnex head-dedicated gradient insert coil, and a birdcage head radio-frequency coil tuned to 127.4 MHz driven by a Varian Unity Inova console. A gradient-echo EPI sequence was used for image collection [measurement repetition time (TR) of 3.0062s, echo time (TE) of 30 ms, matrix resolution of 64 × 64 voxels, voxel size of 3 × 4 × 5 mm, field of view of 220 × 220 mm]. Each volume was constructed from 25 axial slices. In each of the spatial and memory blocks, 328 volumes were collected (∼16 min). In the divided-attention block, 114 volumes were collected (∼8 min). These totals included four “dummy volumes” taken before each block started, to allow the signal to saturate before data acquisition.

fMRI analysis

Data were processed and analyzed using FEAT v5.1, (from FMRIB's Software Library, www.fmrib.ox.ac.uk/fsl). Slice-timing correction used Fourier-space time-series phase-shifting. The MCFLIRT algorithm was used for motion correction (Jenkinson et al. 2002), producing six motion-correction parameters that were then used as regressors in the general linear model (GLM). Brain extraction tool (BET) was used for brain extraction (Smith 2002). Spatial smoothing used a Gaussian kernel of FWHM 5 mm. Grand mean scaling was applied across all volumes by the same factor. High-pass temporal filtering [Gaussian-weighted line spread function (LSF)] used straight line fitting, with sigma = 64 s. All registration to high resolution space defined by the participant's structural scan or standard Montreal Neurological Institute (MNI) space used FLIRT (Jenkinson and Smith 2001; Jenkinson et al. 2002). Analysis of the time-series data used a general-linear-model approach, implemented by FILM (fMRIB's improved linear model) with local autocorrelation correction (Woolrich et al. 2001). The hemodynamic response function (HRF) was modeled as a gamma function, with SD of 3 s and mean lag 6 s. The events of interest were each modeled with their temporal derivatives, and the motion covariates were modeled without any temporal derivative.

The main purpose of the fMRI analysis was to reveal brain areas involved in choosing where to attend within each of the spatial and memory conditions. The analysis therefore concentrated on isolating and differentiating cue-related activity for the different conditions of interest. Initially, parallel first-level analyses were carried out separately for the spatial and memory conditions. There were seven events of interest in each primary analysis. Cues were modeled as events of 1 s duration, separately for the different experimental conditions in the spatial and memory blocks. There were six cue conditions in each block because on either instructed or choice trials, the responses could either be to a leftward target, a rightward target, or an error. Catch trials that were correctly performed (i.e., not responded to) were incorporated into correct instructed or choice events. The letter series from all trials in each block were modeled together as an event of 3 s duration.

The data were subsequently analyzed in two ways. First the activity for each contrast (e.g., choice-guided > instructed) was compared across the whole brain across participants. Second, a region-of-interest (ROI) analysis explored how the neural activity measured by the blood-oxygen-level-dependent (BOLD) signal varied within the networks previously implicated in attentional orienting and voluntary action control.

Whole-brain analysis

At the second level (combining blocks across participants), analysis used FLAME stage 1 [FMRIB's local analysis of mixed effects, without the final MCMC-based stage (Beckmann et al. 2003; Behrens et al. 2003; Woolrich et al. 2004)]. Then a third-level analysis (combining participants) used the “full” version of FLAME. Z-statistic images were thresholded using clusters determined by z >2.3 and a (corrected) cluster significance threshold of P < 0.01 (Forman et al. 1995; Worsley et al. 1992).

ROI analyses

The ROI analyses tested for any differences between areas within the cluster of activated medial frontal cortical regions reported in the whole-brain analysis (see results), and carried out similar analyses on a group of FEF areas and a group of parietal cortical regions. ROIs were centered on the peak coordinates reported in previous fMRI studies investigating either attentional orienting (FEF and PPC) or voluntary action (medial areas). The source articles and coordinates of these ROIs are shown in Table 2. The medial ROIs included the SEF, pre-SEF, presupplementary motor area (pre-SMA), and anterior cingulate cortex (ACC). The PPC ROIs included dorsal and ventral intraparietal sulcus (IPS), and angular gyrus. FEF ROIs included medial and lateral FEF. All ROIs were positioned bilaterally over symmetrical positions except for the angular gyrus. The ROIs were defined using voxels in structural space in which each voxel was isotropic and had a 2 mm side (volume = 8 mm³). ROIs were centered on the coordinates given in Table 2. The FEF and PPC ROIs were spheres with a diameter of 18 mm (volume = 257 voxels 2,056 mm³). For example, the left lateral FEF ROI extended at its outermost points from −35 ≤ x ≤ −53, −3 ≤ y ≤15, 33 ≤ z ≤51. To ensure that the ROIs in different hemispheres did not overlap, the ROIs for medial regions were rectangular cuboids of identical orientation and volume (volume = 4 × 8 × 8 = 256 voxels 2,048 mm³). For example, the left pre-SEF ROI extended from −7 ≤ x ≤ −15, −1 ≤ y ≤15, 51 ≤ z ≤67. The center of the left medial FEF ROI was displaced laterally so that it did not overlap with ROIs on the medial surface. To compare the functional modulation of activation in each ROI, the percentage change in the parameter estimate was extracted for each event of interest and for each participant. Note that there was no overlap between anatomical ROIs in individual participants because the parameter estimates were extracted only after transformation of each participant's data in the standard space in which the ROIs were built. Subsequent analysis used a repeated-measures ANOVA, with Huynh-Feldt sphericity correction, to test the effects of volition (choice or instructed trials), memory (spatial or memory blocks), switch (same or different), hemisphere (left or right), and area (the 4 medial ROIs, 3 PPC ROIs, or 2 FEF ROIs). A subsequent analysis tested for the effect of switch, independent of memory or volition: all correct instructed or choice trials were recoded according to whether the relevant or chosen target appeared on the same side (stay) or on the other side (switch), relative to the most recent correct trial (usually the previous trial). Both spatial and memory blocks were re-coded and re-analyzed.

TABLE 2.

MNI coordinates of the regions of interest and the peak activations reported in previous studies on which they are based

Area	ROI Centre			Cited Activation			Cited Paper
	X	Y	Z	X	Y	Z
Left FEF lat	44	6	42	44	4	46	Grosbras et al. 2005
Left FEF med	−22	0	42	−20	11	56	Grosbras et al. 2005
Left IPS d	−28	−54	54	−28	−54	−54	Grosbras et al. 2005
Left IPS v	−26	−72	32	−26	−72	32	Grosbras et al. 2005
Left SEF	−3	−1	69	−8	−4	64	Grosbras et al. 2001
Left preSEF	−11	7	59	−12	8	60	Grosbras et al. 2001
Left ACC	−3	17	37	0	16	36	Walton et al. 2004
Left preSMA	−3	9	53	−10	9	53	Rushworth et al. 2002
Left ANG	−42	−74	30	−46	−76	32	Mort et al. 2003
Right ANG	42	−66	30	42	−66	30	Mort et al. 2003

Regions of interest in the right hemisphere are the mirror image of those in the left hemisphere except for the angular gyrus. ANG, angular gyrus; FEF, frontal eye field; IPS, intraparietal sulcus; ACC, anterior cingulate cortex; preSMA, presupplementary motor area; SEF, supplementary eye field; preSEF, presupplementary eye field; lat, lateral; med, medial; d, dorsal; v, ventral.

RESULTS

Behavioral data

Participants' performance at target detection depended on how they had been precued (Fig. 2). The main effect of block on error rates was apparent not only when the spatial, memory and divided-attention blocks were compared against each other [F(2,20) = 17.5, P < 0.001] but also for the direct comparisons of spatial versus divided attention [F(1,10) = 26.1, P < 0.001], and memory versus divided attention [F(1,10) = 14.6, P < 0.01] but not for spatial versus memory [F(1,10) = 1.1, P = 0.3]. The main effect of precue type followed a similar pattern with differences among instructed, choice, and divided-attention trials being evident both when the data from spatial and memory blocks were collapsed together [F(2,20) = 17.31, P < 0.0001 (Fig. 2)] and when each was compared separately to the divided-attention block {spatial block: [F(2,20) = 18.62, P < 0.0001 (divided attention, 27.9%; instructed, 7.4%; choice, 15.5%)]; memory block: [F(2,20) = 11.78, P < 0.001 (divided attention, 27.9%; instructed, 10.6%; choice, 15.7%)]} but not when all instructed trials were compared with all choice trials [F(1,10) = 1.1, P = 0.3]. Compared with when their attention was divided, participants were more accurate both after being instructed where to attend and when they themselves chose where to attend [spatial block: instructed vs. divided attention, t(10) = 5.8, P < 0.001; choice vs. divided attention, t(10) = 3.4, P < 0.01; memory block: instructed vs. divided attention, t(10) = 4.1, P < 0.01; choice vs. divided attention, t(10) = 3.1, P < 0.05]. Error rates were slightly lower on instructed than on choice trials in the spatial block [t(10) = 2.7, P < 0.05], and marginally so in the memory block [t(10) = 1.9, P = 0.08], meaning that although choice and instructed conditions both confer a behavioral advantage over divided attention, they may do so to different extents.

FIG. 2.

Behavioral data. Participants covertly oriented their selective attention on both choice and instructed trials of the spatial block (A and B) and the memory block (C and D), shown as improved accuracy compared with divided attention trials (A and C) without any reaction time trade-off (B and D). Bars show SE.

Effects on reaction time were in general less apparent than those on error rates described in the preceding text. Main effects of block were present when including data from all three block types [F(2,20) = 4.5, P < 0.05], when comparing memory with divided attention [F(2,20) = 6.1, P < 0.05] or spatial with memory [F(1,10) = 12.6, P < 0.01], but not with the spatial versus divided-attention comparison (P = 0.4). The effect of precue type was apparent when combining data across all blocks [F(2,20) = 4.9, P < 0.05] or when comparing memory with divided-attention blocks [F(2,20) = 7.89, P < 0.01 (divided attention, 489.3 ms; instructed, 449.2; choice, 483.0 ms)]. When the spatial and memory blocks were contrasted, there were main effects of block [F(1,10) = 12.6, P < 0.01], precue type [F(1,10) = 10.0, P < 0.05] and a marginally significant interaction between block and precue [F(1,10) = 4.1, P = 0.07]. This was driven by an effect that was specific to the memory block, namely that there was faster performance on the instructed condition than on either the divided-attention block [t(10) = 3.9, P < 0.01] or on choice trials [t(10) = 3.1, P < 0.05] with a marginally significant improvement on choice relative to divided-attention trials [t(10) = 0.55, P = 0.06]. No other main effects or interactions were observed.

To compute how often participants voluntarily switched their attention, we divided up the choice trials on both spatial and memory blocks into choose-same or choose-different trials. There were slightly more choose-different than choose-same trials (spatial block: choose same, 45.1%; choose different, 54.9%. memory block: choose same, 42.5%, choose different, 57.5%), but this difference did not reach statistical significance whether data from both spatial and memory blocks were collapsed together or treated separately [paired t-test, all t's (10) < 2.2, P's > 0.05]. There was also no interaction between the type of block (spatial or memory) and whether participants chose to repeat or switch [F(1,10) = 0.4, P = 0.5].

fMRI analysis

A whole-brain analysis incorporating the entire dataset from both spatial and memory blocks, and time-locked to the precue, revealed that a single cluster of areas in medial frontal cortex was more active when participants chose where to attend than when they were instructed (Fig. 3). This cluster, the only activation surviving the choice > instructed contrast, consisted of six local maxima in the SEF, pre-SMA, ACC, and a region of dorsal medial frontal cortex that has previously been described as pre-SEF (Table 1). Neither the reverse contrast (instructed > choice) nor any other main effects or interactions yielded any activations that survived correction for multiple comparisons across the whole brain.

FIG. 3.

A single contiguous cluster of medial frontal cortical regions were more active during choosing where to attend than being instructed [choice > instructed, across both spatial and memory blocks, time-locked to cue onset, overlaid on the average structural magnetic resonance image (MRI) from all participants, corrected for multiple comparisons (thresholded at z = 2.3, cluster P = 0.01)]. Left: cross-hairs on MNI (x,y,z) coordinates 14, 10, 62 [right presupplementary eye field (pre-SEF)]. Right: cross-hairs on 4, 18, 36 [right (ACC)].

TABLE 1.

Local maxima within cluster of activations surviving contrast of choose > instructed

Area	Z-value	X	Y	Z
Right preSEF	3.32	14	10	62
Left preSMA	3.26	−6	16	46
Right ACC	3.12	4	18	36
Right preSMA	3.12	2	14	46
Left preSEF	3.01	−10	6	62
Left preSMA	2.99	−4	6	52

MNI coordinates of all the local maxima within cluster shown in Fig. 3.

ROI analysis

The ROI analyses tested for any functional differences between the areas within the large medial cluster as such heterogeneity would be one possible explanation for the pattern of results found in the whole-brain analyses. For completeness, we also performed similar ROI analyses within a group of FEF areas, and again in a group of PPC areas all of which show functional specialization during attentional selection (Table 2, Fig. 4). Testing for differences between the areas in each of these three separate groups entails three separate ANOVAs at the opening stage of the analysis and so, to correct for multiple comparisons, an adjusted P of 0.05/3 = 0.017 was used throughout as the threshold for significance.

FIG. 4.

Location of each ROI. ROIs in the left hemisphere are numbered in the order presented in Table 2, i.e., 1, left lateral frontal eye field; 2, left medial frontal eye field; 3, left dorsal intraparietal sulcus; 4, left ventral intraparietal sulcus; 5, left supplementary eye field; 6, left presupplementary eye field; 7, left anterior cingulate cortex; 8, left presupplementary motor area; 9, left angular gyrus. To show all ROIs, cross-sections are taken with the cross-hairs within the left presupplementary motor area (left) and the left anterior cingulate cortex (right). Note that all ROIs are of equal volume but the medial ROIs are rectangular cuboids to prevent overlap.

In the medial regions, there were main effects of volition [F(1,10) = 24.2, P < 0.01], consistent with the areas' greater activation in the choice condition, hemisphere [F(1,10) = 9.3, P < 0.017], consistent with left medial frontal areas being more activated than right medial frontal areas, and an interaction between volition and memory [F(1,10) = 8.1, P = 0.017]. Functional differences between the areas in the medial group were apparent from a four-way interaction among the factors area, hemisphere, volition, and switch [F(1,10) = 8.2, P < 0.01]. This interaction justified exploring these effects further by running separate ANOVAs on each ROI. Looking at each medial ROI separately, there was a main effect of volition in the pre-SEF [F(1,10) = 23.9, P < 0.01], pre-SMA [F(1,10) = 28.2, P < 0.001], and ACC [F(1,10) = 10.4, P < 0.01] and a marginally significant effect in the SEF [F(1,10) = 6.5, P = 0.028]. In the pre-SEF and SEF, the effects of volition did not interact with memory, suggesting that their role in volition is independent of the need to remember past actions. Activity in the pre-SMA and ACC showed interactions between the factors memory and volition [F's(1,10) ≥ 13.2 , P's ≤ 0.01], meaning that their preferential activity for choice may be interrelated to a role in remembering previous actions during choosing rather than choosing per se. The within-ROI analysis showed main effects of hemisphere in the pre-SMA [F(1,10) = 11.2, P < 0.01] and reaching marginal significance in the pre-SEF [F(1,10) = 8.07, P = 0.018]. Paired t-tests on each area in each hemisphere in each block showed that all effects of volition were due to there being more activity on choice than on instructed trials (all P's ≤ 0.05) and that the interaction with memory was always due to choice and instructed trials being more similar in the memory block.

There were also main effects of volition in the FEF [F(1,10) = 17.6, P < 0.01] and a trend toward an effect of hemisphere [F(1,10) = 6.9, P = 0.025] with activations being more pronounced in the left hemisphere. As with the medial ROIs, there were functional differences between the areas within the FEF group, shown as a marginally significant interaction between area and volition [F(1,10) = 7.2, P = 0.02]. Looking at each FEF ROI separately, only the lateral FEF showed a main effect of volition [F(1,10) = 15.9, P < 0.01]. Like the pre-SEF and the SEF, the lateral FEF also showed no interaction with memory and so may be attributed a role in volition that is partly independent of memory. Again, this volition effect was driven by there being more activity on choice than on instructed trials (all P's ≤ 0.05). The medial FEF showed a hemisphere effect with greater activity in the left hemisphere [F(1,10) = 8.95, P < 0.017].

The group of PPC areas was not sensitive to whether participants were choosing or being instructed where to attend [no effect of volition, F(1,10) = 2.3, P = 0.16]. Although there was variation in the general level of activation within the PPC group {main effects of area [F(1,10) = 6.6, P < 0.01] and hemisphere [F(1,10) = 13.1, P < 0.01]}, there were no interactions with any other factors and so no further analyses were conducted.

The histograms in Fig. 5 display the pattern of mean activity within each of the ROIs that showed main effects of volition, and this is plotted separately for the choice and instructed conditions from each of the spatial and memory blocks. Areas are plotted separately for each hemisphere if that area showed an effect of hemisphere. The left and right pre-SEF, and the lateral FEF all showed more activity on choice than instructed trials in both spatial and memory blocks and so were preferentially active for choosing regardless of memory. By contrast, the SEF, pre-SMA, ACC, and the medial FEF only showed more activity on choice trials in the spatial block and not in the memory block. For example, for the SEF the choice effect is clearly carried by the spatial block, whereas on the memory block the mean signal change for both instructed and choice conditions was approximately zero. The dorsal and ventral IPS regions are active throughout but are not sensitive to whether the participants were choosing where to attend or were instructed.

FIG. 5.

Histograms showing mean % blood-oxygen-level-dependent (BOLD) change from baseline signal within each region of interest (ROI). ROIs that showed a main effect of hemisphere are shown for each hemisphere. A: medial areas. B: parietal areas. C: frontal areas (spatial, spatial block; memory, memory block, choice, choice trials; instr, instructed trials. Only left pre-SEF and lateral FEF showed more activity on choice than instructed trials in both the spatial and memory blocks.

Some of the differences between conditions were due to relative deactivations from the implicit baseline used in fMRI analysis. Apart from the right medial FEF, all areas preferentially activated by choice showed some deactivation in the instructed condition. It is also apparent from the histograms that the hemispheric differences (medial FEF, pre-SEF, pre-SMA) were due to differences in the extent of deactivation with the right hemisphere being more deactivated rather than any changes in the pattern.

DISCUSSION

A group of medial frontal areas, including the pre-SEF, pre-SMA, and the ACC were more active when participants themselves chose where to attend compared with when they were instructed. Further analysis revealed some evidence for a similar activity pattern in the lateral part of the FEF. It is possible that when participants are asked to make free choices about how to act or to attend that they do so by reference to previous choices they have made (Hadland et al. 2001; Lau et al. 2004). In the case of the pre-SEF and lateral FEF, however, activation in the choice conditions was not simply attributable to such a concomitant increase in memory load. This is the first report of choice-guided covert visual spatial attentional orienting and its neural correlate.

There is currently a controversy over the extent to which any of the medial-fronto-parietal attentional control areas might be specialized for voluntary as opposed to involuntary attentional orienting, and here we explored whether using choice-guided attentional orienting as an experimental paradigm could contribute to this debate. The task demonstrated that people can freely choose where to attend: participants were more accurate in the choice condition than in the divided-attention condition in which their attention was equally distributed across both peripheral letter-streams. The behavioral improvement relative to this divided-attention condition shows that participants were selectively orienting their attention, and the absence of any location being specified by an external stimulus indicates that participants were choosing where to attend. The behavioral benefits of selective attention after choice and instruction differed slightly, suggesting possible differences in the underlying neural mechanism. One potential behavioral advantage of choosing where to attend covertly might be the ability to attend selectively to a region of space without indicating this to a nearby agent: the direction of gaze is an attentional precue (Friesen et al. 2005). Additionally, learning by trial and error requires generating an initial exploratory action (the “trial”) on insufficient evidence, so that the outcome (the potential “error”) can be interpreted (Sutton and Barto 1998). When action selection is exploratory rather than exploitative of known reward contingencies (Daw et al. 2006) or constrained by explicit instruction (Yoshida and Ishii 2006), then it is associated with increased activation of medial frontal cortex. Although the strength of the task in this experiment does not depend on its ecological validity, random behavior can also be useful when interacting with other agents (Barraclough et al. 2004), and the activity of neurons in medial frontal areas is modulated during the performance of such tasks (Seo and Lee 2007).

Repetitive random free selection might bear an additional memory component when compared with an instructed condition. The memorization of previous choices may be intrinsic to the performance of free-choice paradigms. To test whether choice-related activity was greater than that expected simply as a consequence of remembering the last choice, another instructed cueing condition was used in an additional memory block in which participants were explicitly told to remember each response and were instructed on each trial how to respond in the context of that previous response. In this memory block, the memory demands were equated for choice and instructed attentional shifts. Therefore if an area was preferentially active on the choice-guided condition relative to the instructed condition and if this preferential activity remained even when the instructed condition also involved a memory component, then this would be good evidence that the free selection of attentional orienting was driving activity in that area. Accordingly the pre-SEF and lateral FEF showed increased activity for choice versus instructed trials on both spatial and memory blocks. Preferential activation for choice was not eliminated by controlling for the possible memory demands when participants chose in the context of previous trials. This is therefore strong evidence that these areas contribute to the control of choice-guided attentional orienting.

The term “pre-SEF” was coined in a PET study by Petit and colleagues (1996) to describe a region within the pre-SMA, and rostral to the SEF, that was active while participants performed newly learned saccade sequences. Subsequently an fMRI study described a similar area as pre-SEF that was more active during newly learned than familiar saccade sequences (Grosbras et al. 2001). This region [MNI (x, y, z) coordinates −11, 7, 59] was the only medial area that showed more activity during choice-guided than instructed orienting on both spatial and memory blocks. The performance of newly learned sequences is, like free selection of covert visual attention, an example of visuospatial control in the absence of explicit instruction by external stimuli. The task may contain a free-selection component because participants are less practiced and therefore have to choose what to do. A nearby region [MNI (x, y, z) coordinates −4, 8, 54] has been associated with the free selection of eye movements (Nachev et al. 2005). The similarity between that peak and the pre-SEF peak in the current study is consistent with previous reports of commonalities between the neural systems mediating covert attention and overt eye movements (Grosbras et al. 2005; Nobre 2001). The current data contribute to the lively ongoing debate over whether eye movement plans and attentional shifts are neurally dissociable (Rizzolatti and Craighero 2004) and show that an eye movement is not necessary for the pre-SEF to be active: a choice-guided attentional shift is sufficient. It should be cautioned, however, that the peak activations reported in the current study and by Nachev et al. (2005) were not identical although lying within 8 mm of each other. This distance is at the limit of the spatial resolution offered by fMRI and by comparisons across experiments, and so it is possible that there may be differences between the neural areas responsible for the activations found in the Nachev et al. (2005) study and the current experiment. The findings therefore do not rule out the possibility of differences between the neural systems governing the control of overt and covert visual orienting but rather show that the pre-SEF does not require an overt eye movement to show activation and has a role in free selection that is both covert and attentional.

A degree of dissociation between pre-SMA and pre-SEF in the current study is consistent with the view that there may be differences between the roles of tissue in the vicinity of the pre-SEF and pre-SMA even if both are involved with the free selection of action. The term pre-SEF has previously been used to refer to a similar region to the one we label pre-SEF in tasks requiring visual selection, but it may also play a role during manual selection. It is certainly the case that in the macaque, only a single pre-SMA region has been identified, and it is largely amodal, containing cells that are driven by both hand and eye movements (Fujii et al. 2002). Although the activations reported in previous studies of freely selected finger movements are more posterior than the pre-SEF region identified here (e.g., Jenkins et al. 2000; Jueptner et al. 1997), pre-SEF effector specificity may be unlikely given the lack of evidence for an oculomotor- or visual-specific neural subpopulation within the macaque pre-SMA. Future work may help determine whether the pre-SEF region is differentially sensitive to the type of orienting (covert or overt), the style of orienting (choice, instructed, reflexive) and the domain (visuospatial/oculomotor or manual).

As with all the hemispheric interactions reported here, the relative pattern of activity did not differ across the left and right pre-SEF. Instead the hemisphere effect found in the analyses was the consequence of lower overall levels of activation and even deactivation in the right hemisphere. On the basis of the current evidence, the left pre-SEF is more easily and unambiguously associated with choice-guided attentional orienting than is the right pre-SEF. Although it is not currently clear how to interpret deactivations in fMRI data, one way to interpret the deactivations in the present study (Fig. 5) may be in relation to the possible occurrence of some spontaneous, freely chosen, but covert, orienting in the baseline inter-trial period. The period in between trials is important in fMRI analysis because it contributes to an implicit baseline against which the BOLD signal change is scaled. Freely chosen overt or covert orienting may also have occurred in the inter-trials intervals in previous fMRI studies. One way to address this possibility in the future might be to vary parametrically the rate or number of freely selected versus instructed covert attentional shifts.

The lateral FEF was the only other area that responded to the task design by showing the characteristic functional profile of being more active during choosing where to attend compared with during instruction on both spatial and memory blocks. This activation was not contiguous with the medial cluster (Fig. 3) and so did not appear in the whole-brain analyses, which are designed to emphasize large clusters of activity. Rather the lateral FEF was found to be dissociable from the rest of the fronto-parietal attention network. This is important because it has proven difficult to separate out exactly how the various nodes of this network differ from each other (Grosbras et al. 2005; Nobre 2001). Some recent studies have reported neurons across the FEF to be driven more by instructed attentional orienting then a reflexive condition (Buschman and Miller 2007; Corbetta and Shulman 2002). Here we find that an even more extreme style of attentional control characterizes the lateral FEF and even dissociates it from the medial FEF (as well as the PPC). There is little other extant human functional data distinguishing between lateral and medial FEF, but there is some evidence for heterogeny within the monkey FEF. Compared with medial FEF, the monkey lateral FEF has larger layer-III pyramidal cells and more diffuse patterns of myelination (Preuss and Goldman-Rakic 1991). Lateral FEF microstimulation produces saccades of smaller amplitude than in medial FEF (Preuss and Goldman-Rakic 1991; Robinson and Fuchs 1969). Lateral FEF also has small receptive fields and strong connections to foveal representations in the ventral visual stream in contrast to the large receptive fields in medial FEF and strong connections to peripheral representations in the dorsal stream (Bullier et al. 1996; Schall 1995). Future work is necessary to explore how such differences might relate to choosing where to attend or being instructed. Furthermore, in the human, this medial-lateral pattern may be reversed because the lateral FEF may be specialized for smaller-amplitude saccades than the medial FEF (Paus 1996).

A large medial cluster of areas has been previously implicated in the free selection of motor responses (Deiber et al. 1991; Frith et al. 1991; Jenkins et al. 2000; Lau et al. 2004). These same areas were more activated during choosing where to attend than in the instructed condition, suggesting that there may be a general mechanism involved in internally guided control, whether for the visuospatial/oculomotor domain or for the manual domain. For example, the SMA and the SEF proper may be important for inhibiting potential automatic actions initiated in response to environmental affordances in order for alternative actions to be made (Sumner et al. 2007). In some cases, however, medial frontal activity may be driven partly by the fact that choices are often made by reference to previous trials rather than simply to the action generation process itself. For example, the pre-SMA has previously been implicated in switching between tasks involving finger movements (Brass and von Cramon 2002; Dosenbach et al. 2006; Rushworth et al. 2002), suggesting that it may have a role in determining current response sets partly on the basis of previous response history. The ACC has been implicated in the guidance of action selection on the basis of previous reward history (Behrens et al. 2007; Kennerley et al. 2006; Matsumoto et al. 2007; Walton et al. 2004). The pre-SEF region may, however, have a more direct role in the generation of an uninstructed action. In the macaque, activity in the adjacent SEF region has been recorded in a high proportion of neurons at an early time point when free saccade choices are made (Coe et al. 2002). The current study suggests that the pre-SEF region may play a particularly critical role in the generation of voluntary behavior. The SEF ROI in the current study was based on an ROI in a previous study that was not more active during newly learned than familiar saccade sequences but which, instead, was part of a common fronto-parietal network (Grosbras et al. 2001).

Free selection can also be considered within the context of “conflict”. On choice trials (and not on instructed trials), there is conflict between the two potential locations to be attended. Choosing resolves this conflict. An equivalent description of the current results would, then, be to ascribe the pre-SEF a role in resolving or processing the conflict that occurs during choosing where to attend, over and above any considerations of response history. The other MFC areas, which had activity patterns reflecting the need to remember choices made on the previous trial, have previously been implicated in detecting or resolving conflicting stimulus-response associations in the context of recent task history (Botvinick 2007; Botvinick et al. 2004; Rushworth 2008). Future work might explore whether the correlation between pre-SEF region activity and choosing where to attend reflects a causal role for the area in actively resolving conflict; analogous tests have already been conducted in other parts of dorsal MFC (Isoda and Hikosaka 2007; Stuphorn and Schall 2006; Taylor et al. 2007).

All areas selectively activated by choice were in the frontal lobe. However the current task did not ostensibly tax dorsolateral prefrontal cortical (DLPFC) function. One proposed role of the DLPFC is in the active maintenance in working memory of response options relevant for future behavior (Passingham and Sakai 2004) which may not have been required in the current task because the possible target locations were displayed on the screen. The parietal cortex was well activated on choice trials but also on the instructed condition and so seems to be driven by attentional control independently of whether the orienting is initiated by a rule or by choice.

It is unclear why participants made slightly more errors on the choice than the instructed condition. One possibility is that it may take longer to generate a covert attentional shift when the recent response history is considered or when there is less information available to guide the decision. Although tentative, this speculation would predict that on a small minority of choice trials, participants may not have yet successfully oriented their attention selectively to one side or the other by the time the relevant target appeared, leading to an increase in incorrect responses. Future work using modified tasks or techniques with higher temporal resolution (e.g., electroencephalography) might help explore this intriguing aspect of the behavioral data.

GRANTS

Funded by the Wellcome Trust, Royal Society, the James S McDonnell Foundation, and the Medical Research Council.