The Neural Substrate of Orientation Working Memory

Transcription

1 The Neural Substrate of Orientation Working Memory L Cornette 1, P Dupont 1, E Salmon 2, and Guy A Orban 3 Abstract & We have used positron emission tomography (PET) to identify the neural substrate of two major cognitive components of working memory (WM), maintenance and manipulation of a single elementary visual attribute, ie, the orientation of a grating presented in central vision This approach allowed us to equate difficulty across tasks and prevented subjects from using verbal strategies or vestibular cues Maintenance of orientations involved a distributed fronto-parietal network, that is, left and right lateral superior frontal sulcus (SFSl), bilateral ventrolateral prefrontal cortex (VLPFC), bilateral precuneus, and right superior parietal lobe (SPL) A more medial superior frontal sulcus region (SFSm) was identified as being instrumental in the manipulative operation of updating orientations retained in the WM Functional connectivity analysis revealed that orientation WM relies on a coordinated interaction between frontal and parietal regions In general, the current findings confirm the distinction between maintenance and manipulative processes, highlight the functional heterogeneity in the prefrontal cortex (PFC), and suggest a more dynamic view of WM as a process requiring the coordinated interaction of anatomically distinct brain areas & INTRODUCTION Working memory (WM) refers to a limited-capacity system in which information can be actively maintained (ie, stored and rehearsed) for a time period of up to several seconds and manipulated A typical manipulation process is the updating of the memory content, one of the topics of the present study Many imaging studies investigating the neural correlates of WM have focused on the prefrontal cortex (PFC), a heterogeneous brain structure involved in several higher-order cognitive processes (Goldman-Rakic, 1995) While several groups have posited a dorsal/ventral segregation of PFC, depending on the modality of information (spatial vs object) held in WM (eg, Courtney, Ungerleider, Keil, & Haxby, 1996), more recent work in monkey single-cell recordings (Rao, Rainer, & Miller, 1997), human lesion (for review, see D Esposito & Postle, 1999), and neuroimaging studies (for review, see D Esposito, Ballard, Aguirre, & Zarahn, 1998) have questioned such models Rather, the dorsal/ventral subdivision may reflect the type of processing performed on the information held in WM (Smith & Jonides, 1999; D Esposito et al, 1998; D Esposito, Postle, Ballard, & Lease, 1999; Petrides, 1994), while ventrolateral prefrontal cortex (VLPFC, primarily Brodmann s areas 44, 45, and 47) is the site that maintains information stored in posterior areas in an active state to guide behavior, and dorsolateral 1 Katholieke Universiteit Leuven and Centrum voor Positron Emissie Tomografie, 2 Centre Hospitalier Universitaire de Liège, 3 Katholieke Universiteit Leuven, Belgium prefrontal cortex (DLPFC, primarily Brodmann s areas 9 and 46) is recruited only when information held in memory has to be manipulated Accordingly, we have recently (Cornette, Dupont, Bormans, Mortelmans, & Orban, 2001) attempted to chart the types of cognitive processes involved in short-term memory and WM into a 3-D space (see Figure 1) Ultra-short-term memory refers to the retention of a single item over a delay interval of 300 msec, without manipulative operations Short-term memory involves maintenance over a longer, nondistracted delay interval, generally a couple of seconds, again without manipulative operations WM refers to maintenance of multiple items with or without manipulative operations In order to better delineate the neural substrates of the cognitive processes involved in short-term memory and WM, we have begun to use orientation as a visual attribute (Cornette et al, 2001; Orban, Dupont, Vogels, Bormans, & Mortelmans, 1997), the neuronal coding of which is well understood (Hubel & Wiesel, 1968) Indeed, the use of this attribute enables us to remove several confounding cognitive processes from the WM tasks used First, its elementary nature and presentation in the central visual field obviate the need to process object identity and spatial location, respectively Indeed, some authors have suggested that visuospatial attention control and spatial WM share a number of cortical regions (Awh, Jonides, & Reuter-Lorenz, 1998; Coull & Frith, 1998) Second, the presentation of oblique orientations prevents the use of vestibular cues (Brandt, Dieterich, & Danek, 1994) or verbal codes that can be rehearsed using articulatory mechanisms, ensuring that 2001 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 13:6, pp

2 MANIPULATIVE FUNCTIONS WM-maintenance plus Update W I T H NUMBER OF ITEMS 2back 1back DURATION OF STORAGE W I T H O U T TSD Load 3 Load 6 WM-maintenance only Figure 1 Taxonomy and related tasks Tasks are plotted along three axes: (1) the x axis, indicating the duration of item storage; (2) the y axis, indicating the presence or absence of manipulative operations; and (3) the z axis, indicating the number of items to be stored WM either involves maintenance without manipulative operations (ie, retention of more than one item, eg, Load3 and Load6 tasks) or maintenance plus manipulative operations (eg, 1-back, 2-back, and Update tasks) TSD (temporal same different or successive discrimination task) exemplifies ultrashort-term memory, defined as the capacity to retain a single item over a delay interval of 300 msec, without manipulative operations Short-term memory (white box) covers a longer nondistracted delay interval, generally a couple of seconds, again without manipulative operations maintenance relies solely on visual signals Finally, and most importantly, this attribute allows performance levels to be equalized among tasks Indeed, many of the regions commonly activated by increasing task difficulty are located in PFC (Elliott, Rees, & Dolan, 1999; Sunaert, Van Hecke, Marchal, & Orban, 1999; Barch et al, 1997; Grady et al, 1996), and it has not always been clear from previous WM studies in how much recruitment of PFC regions during WM cognitive processes may have been contaminated by differences in difficulty across tasks Therefore, the first aim of the current study was to investigate the neural substrate involved in orientation maintenance only Maintenance is a general concept that necessarily involves a diverse array of operations: The contents of WM can be refreshed simply by rehearsal, or scanned during retrieval, or augmented by integration with new information from the environment or from long-term memory (Haxby, Petit, Ungerleider, & Courtney, 2000) Many neuroimaging studies of the maintenance component in WM are founded on Baddeley s model, positing separate buffers for verbal and visuospatial information (Baddeley, 1986), thus, suggesting that largely different regions are involved in the maintenance of different kinds of information For instance, (passive) storage of verbal material activates the supramarginal gyrus, while (active) phonological rehearsal recruits mainly Broca s area (Salmon et al, 1996; Paulesu, Frith, & Frackowiak, 1993) Maintenance of object information involves mostly ventral regions of PFC (eg, Courtney, Ungerleider, Keil, & Haxby, 1997), while maintenance of spatial information mostly involves the right premotor cortex (eg, Courtney, Petit, Maisog, Ungerleider, & Haxby, 1998; Jonides et al, 1993) Thus, it is impossible to predict from the previous studies which regions are involved in the maintenance of orientation information Using an n-back task design, we recently reported that regions significantly involved in orientation WM included the left DLPFC, left superior frontal 814 Journal of Cognitive Neuroscience Volume 13, Number 6

3 sulcus (SFS), and left supramarginal gyrus (Cornette et al, 2001) However, we were unable to disentangle the extent to which these regions contribute to orientation maintenance as such Using a visual stimulus presentation identical to our previous study, we have now trained 9 subjects to store either 3 (Load3 task, Figure 1, see Methods) or 6 (Load6 task, Figure 1) successively presented orientations, after which a probe orientation was presented The comparison with an orientation identification control task (ID task), namely, subtractions [Load3 ID] and [Load6 ID], then revealed the rehearsal circuit involved in the active maintenance of several orientations Since memory load increased linearly from ID, through Load3 to Load6 tasks, the subtraction [Load6 ID] gave information equivalent to a parametric analysis aimed at isolating the rehearsal circuit in orientation maintenance The second aim of the current study was to investigate the neural substrate underlying one of the manipulative processes of the orientation 2-back task (Cornette et al, 2001), namely reshuffling or updating of memory content Updating is defined as the cognitive operation of entering a new item in the memory store and removing an old one upon each new stimulus presentation, while maintaining temporal order (Belleville, Rouleau, & Caza, 1998; Kiss, Pisio, Francois, & Schopflocher, 1998; Salmon et al, 1996; Morris & Jones, 1990) While the neural mechanisms involved in some of the manipulative processes of n- back tasks such as maintaining temporal order and resisting to distractors have already received some attention (Jiang, Haxby, Martin, Ungerleider, & Parasuraman, 2000; Marshuetz, Smith, Jonides, De Gutis, & Chenevert, 2000), the physiological basis of updating visual information, which is a critical component of many everyday cognitive tasks, has so far remained unresolved (but see Vanderlinden et al, 1999) To that end, we trained the same 9 subjects to continuously maintain the last 3 items in a series consisting up to 7 successively presented orientations before the presentation of a probe stimulus (Update task, Figure 1, see Methods) In order to reveal the cortical network involved in orientation updating, we compared the Update task to a maintenance-only task, in which 3 orientations needed to be retained (Load3 task) This control task involved all cognitive processes present in the Update task, except for the updating component, which is, thus, isolated by the subtraction [Update Load3] Finally, nearly all theoretical and computational neurobiology is based on the observation that the brain s implementation of cognitive processes will almost certainly show nonadditive or interactions effects (Aertsen & Preissl, 1991) One therefore expects to find interactions between cerebral regions involved in maintenance and manipulation, respectively, since the regions involved in manipulation rely on a continuous access to and use of information maintained in the WM Comodulation of the degree of activation indicates that the activities of a network s constituents are similarly affected by specific task demands (Diwadkar, Carpenter, & Just, 2000) Hence, our third aim was to investigate such cooperation between regions involved in orientation maintenance and regions involved in manipulative processes RESULTS Visual Stimulation During Positron Emission Tomography (PET) Scanning Visual input was matched over all tasks: The number of stimuli presented was equal in all tasks (20 stimuli/min), and orientation distributions were statistically undistinguishable (Kolmogorov Smirnov tests, p > 5) Subjects maintained fixation well during all tasks, as demonstrated by the electro-oculographical (EOG) recordings Task Performance During PET Scanning As revealed by post hoc questioning, subjects stored successively presented orientations with reference to each other in either clockwise or anticlockwise order, building up an abstract internal configuration (Load3, Load6, and Update tasks) No external reference such as a clock-face was used Post hoc questioning also indicated that subjects carefully scanned the memory trace of the abstract orientation configuration, as soon as a probe item was presented Performance among the 4 tasks averaged between 79% and 82% correct during PET acquisition and did not differ significantly among tasks [repeated-measures ANOVA F(3,24)=032; p > 8; Figure 2A] No additional learning occurred during PET acquisition since accuracy in all tasks was similar to that at the end of the second training session Performance levels were equated by systematically adjusting d [repeated-measures ANOVA F(3,24)= 3770; p < 10 6 ; Figure 2B], that is, the minimal orientation difference between gratings presented in each trial The d value was smaller in the Update task than in the Load6 task, indicating that during the Update task, subjects did not maintain all successively shown orientations, but, indeed, as requested, maintained and updated only the last three items of each series shown Note that differences in computational demands between tasks are reflected not only in the orientation differences (more complex tasks require larger differences) but also by the reaction times (RTs) [repeatedmeasures ANOVA F(3,24)=15749; p < 10 6 ; Figure 2C] The longer RTs during the Load6 task compared to the Load3 task (Scheffé test, p < 10 3 ) or ID task (Scheffé Cornette et al 815

4 Figure 2 Psychophysical data Bar charts showing for each task: performance (A); mean minimal orientation difference d between gratings within one trial of Load and Update tasks and mean d as used in the ID task (B); and mean reaction times (C) Error bars indicate standard error of the mean (SEM) A Performance (% correct) ID Lo ad3 Load6 Up date B 16 Orientation difference (degree) ID Lo a d 3 L o a d 6 Up d a te C Reaction time (msec) ID L o a d3 L o ad 6 Up d a t e test, p < 10 3 ) most likely reflect longer memory scanning processes (Sternberg, 1966) Maintenance of Orientation The subtraction [Load3 ID] enabled us to investigate the neural correlate of orientation maintenance No region reached the stringent criterion of p corr < 05 The subtraction [Load6 ID], which entails a higher orientation maintenance load, yielded four significant regions ( p corr < 05, Table 1): (1) a region located in the left inferior frontal gyrus (VLPFC, BA 47), below the orbito-frontal sulcus, (2) a region located on the lateral bank of the left superior frontal sulcus (SFSl), (3) left precuneus (BA 7), and (4) right superior parietal lobe (SPL, BA 7) In Figure 3 (upper panel), this fronto-parietal activity pattern is superimposed on a mean magnetic resonance (MR) image of all participating subjects In addition, Figure 3A indicates the locations of the left DLPFC, significantly activated in the orientation Journal of Cognitive Neuroscience Volume 13, Number 6

5 Table 1 [Load6 ID] Coordinates Brain Region x y z Z Score Frontal L inferior frontal gyrus BA 47 (VLPFC) R inferior frontal gyrus BA 47 (VLPFC) L superior frontal sulcus BA 6/8Ad (SFSl) R superior frontal sulcus BA 6/8Ad (SFSl) L precentral sulcus BA 6/ Cingulate R anterior cingulate BA Parietal L precuneus BA R superior parietal lobe BA 7 (SPL) L intraparietal sulcus R inferior parietal lobe BA Cerebellum L cerebellum x, y, and z are the Talairach coordinates of the local maxima (in mm); x = 0 at the midline (+/ = right/left-sided); y = 0 at the anterior commissure (+/ = anterior/posterior); z = 0 at the anterior posterior commisure level (+/ = superior/inferior) L = Left; R = Right; BA = Brodmann s area Z score designates the level of significance: bold = activation significant at p corr < 05 (Z score 440) for peak height; regular = activation significant at p uncorr < 001 (Z score 309) for peak height back task, compared to the 0-back task (Cornette et al, 2001) and of the left frontal eye fields (FEF), ie, the area of activation common to eye saccades, fixation and pursuit tasks, using the Talairach coordinates as reported by Petit, Clark, Ingeholm, and Haxby (1997) The left SFSl region revealed by the current study is situated more than 20 mm more anterior and superior than the FEF and closely neighbors the left SFS region active in the 2-back orientation WM task (Cornette et al, 2001) It is located anterior to the junction between the superior frontal and the precentral sulcus, corresponding to the most cranial part of the middle frontal gyrus In addition, the activity profiles, namely, the adjusted rcbf plotted for all tasks in the most significant voxel, are indicated for the left SFSl, left VLPFC, left precuneus, and right SPL The profiles demonstrate a clear memory load effect, that is, a monotonic increase of rcbf going from ID through Load3 to Load6 tasks, suggesting that these regions mainly index orientation maintenance It is therefore worth mentioning that the subtraction [Load3 ID] yielded several regions reaching the p uncorr < 001 criterion, that is, bilateral SFSl ([ 30, 6, 56], Z score = 348 and [24, 8, 54], Z score = 338) and right precuneus ([4, 12, 70], Z score = 395) Other frontal regions significant at p uncorr < 001 in the subtraction [Load6 ID] were the right homologues of SFSl and inferior frontal gyrus, left precentral sulcus (BA 6/44), and right anterior cingulate (BA 32) In posterior cortex, orientation maintenance involved the right inferior parietal lobe (BA 40), the left intraparietal sulcus, and a region in the left lateral cerebellum Updating of Orientation We used the subtraction [Update Load3] to isolate the neural correlate of orientation updating Both tasks were equated for visual input, performance level, and memory load As indicated in Figure 3 (lower panel), the left SFS was significantly ( p corr < 05) involved, Cornette et al 817

6 Figure 3 Anatomical localization of regions significant in [Load6 ID] and [Update Load3] with corresponding functional profiles Upper panel: SPMs showing 4 regions differentially active (p corr < 05) in the subtraction [Load6 ID], superimposed on sagittal sections through the averaged MR image of 9 subjects (red crosses indicate local maximum on MR): left SFSl (A), left inferior frontal gyrus (B), left precuneus (C), and right SPL (D) are indicated The x Talairach coordinate for each sagittal section and color scale of significance are indicated Top figure also indicates locations of the left FEF (red symbols, [ 30, 19, 47] and [ 39, 15, 38], that is, coordinates from Petit et al (1997)) and left DLPFC (orange symbol, [ 32, 24, 30], that is, coordinate from Cornette et al, 2001) The adjusted rcbf ( y axis, arbitrary units) is plotted for the four conditions (x axis, from left to right: ID task, Load3 task, Load6 task, Update task) in these 4 regions Bars with white outlines refer to the subtraction generating the significant differential activation Horizontal bars indicate standard error of the mean (SEM) Lower panel: SPM showing the regions differentially active in the subtraction [Update Load3]: left SFSm ( p corr < 05, E) and DLPFC ( p uncorr < 001, F) Conventions as in upper panel Inset: Axial section through the averaged MR image of 9 subjects (Talairach z level = +58 mm), indicating regions involved in maintenance ([Load6 ID], Z score >4, green) and regions involved in updating ([Update ID], Z score >4, red) Common activation is indicated in blue with its main center of activity located more medially (SFSm) than during orientation maintenance (see inset, Figure 3) The activity profile clearly demonstrates that this part of the SFSm region, in addition to its sensitivity to an increasing memory load, indexes the updating operation (compare profiles A and E in Figure 3) Note that when comparing the Update to the Load3 task, small increases in activity are also observed in the left VLPFC, left precuneus, and right SPL (Figure 3), possibly reflecting a more intense storage and/or refreshing of information in the situation of a higher computational load However, the increase became significant only in the left SFSm region Other prefrontal regions involved in orientation updating (p uncorr < 001) were the right SFSm and right DLPFC (BA 9/46) Its functional profile indicates that activity in the DLPFC reflects primarily the updating component, rather than the load (Figure 3) Parietal regions involved were the bilateral precuneus, bilateral SPL, and left supramarginal gyrus Finally, 3 regions in 818 Journal of Cognitive Neuroscience Volume 13, Number 6

7 Table 2 [Update Load3] Coordinates Brain Region x y z Z Score Frontal L superior frontal sulcus BA 6/8Ad (SFSm) R middle frontal gyrus BA 9/46 (DLPFC) R superior frontal sulcus BA 6/8Ad (SFSm) Parietal L precuneus BA R precuneus BA R superior parietal lobe BA 7 (SPL) L superior parietal lobe BA 7 (SPL) L supramarginal gyrus BA Cerebellum L cerebellum See Table 1 for abbreviations and conventions the left lateral cerebellum were involved in orientation updating Functional Connectivity In order to assess how the left SFSm region, involved in orientation updating ([ 16, 2, 56], Table 2), is functionally connected with distant regions involved in orientation maintenance, we examined the correlation between the activity in this SFSm region and that in the load-sensitive left precuneus, right SPL, and left VLPFC In addition, we studied the correlations among these latter three regions, resulting in a total of 6 correlational analyses (Figure 4) For each analysis, we plotted the mean adjusted rcbf in the two corresponding most significant voxels (9 subjects and 4 tasks, yielding 36 data points) Significant correlations (all p corr µ 05) were observed among SFSm, precuneus, and SPL (r 07) Correlations were weak ( p uncorr µ 001) between each of the parietal regions and VLPFC, and nonsignificant ( p uncorr > 01) between the SFSm and VLPFC The importance of the functional link between the left precuneus and SFSm was further underscored by the results of a psychophysiological interaction analysis, investigating context-specific modulation of effective connectivity between remote brain areas (Friston, 1997) The correlation between the left SFSm and precuneus activities was tighter within the experimental context of the Update task (r = 7) compared to the context of the Load3 task (r = 3) Conversely, the correlation between the activity of the left SFSl region involved in orientation maintenance ([ 28, 2, 58], Table 1) and precuneus activity was closer within the experimental context of the Load3 task (r = 8) compared to the context of the Update task (r = 4) Finally, we tested for functional connectivity between the right DLPFC ([34, 36, 42], Table 2) and left SFSl (Table 1), left SFSm (Table 2), left VLPFC, left precuneus, and right SPL (all Table 1) None of the activities were significantly correlated, although DLPFC activity was weakly correlated ( p uncorr µ 001) with that of the left SFSm ([ 16, 2, 58], Table 2) and right SPL Nonsignificant ( p uncorr > 01) correlation was observed between the DLPFC activity and that of the left SFSl ([ 28, 2, 58], Table 1), left VLPFC, and left precuneus DISCUSSION In the current PET study, we have characterized two major cognitive components involved in orientation Cornette et al 819

8 SFS [ ] A Precuneus [ ] C A Adjusted rcbf in [ ] r = Adjusted rcbf in [ ] D Adjusted rcbf in [ ] r = Adjusted rcbf in [ ] F D B E SPL [ ] B Adjusted rcbf in [ ] r = Adjusted rcbf in [ ] E Adjusted rcbf in [ ] r = Adjusted rcbf in [ ] VLPFC [ ] C Adjusted rcbf in [ ] r = Adjusted rcbf in [ ] F Adjusted rcbf in [ ] r = Adjusted rcbf in [ ] Figure 4 Correlational analysis Correlation between rcbf in the left SFSm and left precuneus (A), SFSm and right SPL (B), left precuneus and right SPL (C), left VLPFC and left precuneus (D), left VLPFC and right SPL (E), and left VLPFC and left SFSm (F) Left SFSm results from [Update Load3], left precuneus, right SPL, and left VLPFC from [Load6 ID] (left panel) Projection onto a lateral view of the left hemisphere: solid arrows indicate robust correlation (A, B, C, all p corr µ 05), dotted arrows indicate weak correlation (D, E, all p uncorr µ 001) and broken arrow refers to a nonsignificant correlation (F, p uncorr > 01) (right panel) Scatter plots: the regression lines and regression coefficients (r) are indicated Similar correlation coefficients were obtained when the correlation was calculated between the average activity in a spherical region (5-mm radius) centered on the most significant voxel rather than between activity in the most significant voxel itself The r values were 71, 71, 78, 59, 55, and 38 for pairs of regions shown in A F, respectively Note the relative independence of the SFSm and VLPFC WM We have delineated a fronto-parietal network underlying the maintenance of several orientation items, namely, bilateral SFSl, bilateral VLPFC, bilateral precuneus, and right SPL In addition, we have identified bilateral SFSm and, possibly, right DLPFC as the frontal regions underlying orientation updating Functional coupling between superior frontal and posterior parietal regions underlies the cooperation between maintenance and updating processes in orientation WM Design Issues Investigating the neural substrate involved in WM processes such as maintenance and updating requires careful matching of experimental and control conditions for difficulty, typically indexed by performance This prerequisite was met in the current study through the use of orientation as a visual attribute We carefully adapted the orientation difference d between successive stimuli, this for each subject and for each task Although some of the manipulations of difficulty level, such as image degradation, both decrease accuracy and increase RTs (Grady et al, 1996), manipulating the orientation difference within each of the present tasks equated performance levels, while RTs increased only slightly (Cornette et al, 1999) Hence, psychometric matching of tasks ensures that differences in rcbf between tasks are not confounded by differences in task difficulty Increases in RTs between tasks largely reflect increased computational demands Indeed, the prefrontal regions commonly engaged by difficulty (Elliott et al, 1999; Sunaert et al, 1999; Barch et al, 1997; Grady et al, 1996) are located anteriorly and ventrally to those observed in the present study The use of multiple, closely spaced orientations, which also changed each 150-sec run, made it impossible for subjects to use verbal labels, as confirmed by debriefing the subjects The use of oblique orientations and phase randomization excluded the contribution of vestibular, or visual positional or rotational cues Experimental (Load6 and Update) and control (ID and Load3) tasks were carefully matched for visual processing, attentional demands, number and type of 820 Journal of Cognitive Neuroscience Volume 13, Number 6

9 decisions, and number of motor responses, in addition to difficulty Maintenance in Orientation WM The current study confirms the concerted activity of multiple and widely distributed regions (eg, Postle, Stern, Rosen, & Corkin, 2000) necessary for keeping several items online, that is, bilateral SFSl, bilateral VLPFC, bilateral precuneus, and right SPL A high load, consisting of 6 items, involved a region located anterior to the junction between the superior frontal and the precentral sulcus, corresponding to the most cranial part of the middle frontal gyrus, straddling Brodmann s area 6 and dorsal area 8A (BA 6/8Ad) (Petrides & Pandya, 1999) The role of this SFSl region in orientation maintenance is underscored not only by its functional profile, but also by its functional link with parietal regions and the task dependence of this connectivity As indicated in Figure 3, and in agreement with the findings of Leonards, Sunaert, Van Hecke, and Orban (2000) and Courtney et al (1998) (but see Postle, Berger, Taich, & D Esposito, 2000), the SFSl location, as most of the frontal regions activated in our orientation WM study, is clearly situated anterior to the FEF landmark regions Review of the WM literature reveals that spatial maintenance tasks, in which spatial locations must be maintained in memory for short periods, routinely activate a region located immediately anterior to the FEF (Rowe, Toni, Josephs, Frackowiak, & Passingham 2000; Corbetta et al, 1998; Courtney et al, 1998; Owen, Evans, & Petrides, 1996; Petit, Courtney, Ungerleider, & Haxby, 1998; Jonides et al, 1993) While it is unlikely that the SFSl activation in the present study reflects allocation of visuo-spatial attention, it is not clear to what extent it reflects spatial WM Indeed, the internal buildup of an abstract orientation configuration (see debriefing by subjects) could be regarded as some higher-order form of spatial WM However, spatial WM is classically investigated using attributes presented at different locations in the visual field Recent conventional (Postle, Stern, et al, 2000) and event-related (Haxby et al, 2000; Postle & D Esposito, 1999) fmri studies have indicated spatial, as well as object- and face-delay-specific, activity in this frontal area SFSl Hence, these studies concur with our findings, indicating that the SFSl region is involved in orientation maintenance, although it is not clear to what extent this function is related to the modality of information used In addition, part (BA 47) of VLPFC was recruited during the retention of several orientation items General agreement exists concerning the involvement of VLPFC in tasks requiring active maintenance in the absence of manipulative operations (Haxby et al, 2000; Stern et al, 2000; D Esposito et al, 1998; Courtney et al, 1997; Owen, 1997) We believe that load sensitivity in this region is primarily due to the need to bridge long time intervals Indeed, Elliott and Dolan (1999) observed greater activation in VLPFC and anterior temporal regions during a 15-sec delay compared to a 5-sec delay in a delayed-(non)match-to-sample task Accordingly, Rypma and D Esposito (1999) failed to observe any correlation between the number of suprathreshold voxels in VLPFC and the slope of subjects RTs as a function of increasing load On the other hand, VLPFC has often been implicated in governing inhibitory operations, although such operations generally involve a more posterior part of the inferior frontal gyrus (D Esposito et al, 1998; Jonides et al, 1998; Konishi, Nakajima, Uchida, Sekihara, & Miyashita, 1998) The involvement of superior parietal and precuneus regions in orientation maintenance is clear from their functional profiles, indicative of a clear load-dependent effect Upon debriefing, subjects indicated that the 6 items were maintained by building up an abstract internal configuration in which the stored orientations were continuously compared to each other Involvement of SPL in maintaining internal representations has been indicated by functional imaging (Postle, Stern, et al, 2000; Smith & Jonides, 1999) and human lesion studies (Wolpert, Goodbody, & Husain, 1998) The recruitment of precuneus has been indicated whenever the generation of a mental image relies on the reactivation of a memorized concept in long-term memory (Mellet et al, 2000; Fletcher et al, 1995; Kosslyn et al, 1993) It is unclear, so far, whether such reactivation is functionally similar to keeping abstract visual information online in WM Updating in Orientation WM Several imaging studies using the n-back WM paradigm, involving updating among other manipulative functions, have reported activation of a region located in the caudal SFS, the precise function of which has remained largely unspecified (Cornette et al, 2001; Nystrom et al, 2000; Braver et al, 1997; Cohen et al, 1997; Jonides et al, 1997; Schumacher et al, 1996; Awh, Smith, & Jonides, 1995) This is the first study to indicate that the frontal SFSm region is not only involved in orientation maintenance, but that updating of orientation information is one of the computational processes residing in the SFS We cannot completely exclude the possibility that the differential activity in the subtraction [Update Load3] reflects the maintaining of temporal order rather than the updating of orientation itself Although any across study comparison of coordinates requires caution, the study of Marshuetz et al (2000) suggests that the SFS region involved in temporal order is the same as that involved in orientation maintenance, namely, SFSl Such functional heterogeneity in the SFS is in keeping with the demonstration that the SFS region can exhibit sustained (indexing maintenance), as well as transient Cornette et al 821

10 activity at higher loads (indexing manipulative operations) (Cohen et al, 1997) Since orientation updating requires continuous access to stored information, the SFSm update region should functionally connect with the posterior parietal areas that are involved in orientation maintenance That is exactly what we observed and, moreover, the functional link depended on the task context With respect to the widely accepted dorsal/ventral subdivision in PFC, many previous WM studies have focused predominantly on the DLPFC (BA 9/46), a region recruited when information held in memory has to be manipulated (Smith & Jonides, 1999; D Esposito et al, 1998, 1999; Petrides, 1994) However, there is increasing evidence that while the DLPFC is a key component in executive control, it is involved in other executive functions than manipulation per se DLPFC recruitment under conditions requiring executive control has been demonstrated in both human lesion (eg, Baddeley, Della Sala, Papagano, & Spinnler, 1997) and imaging studies (eg, dual-task coordination in D Esposito et al, 1995; task switching in Collette et al, 1999 and Evans et al, 1996; response sequencing and monitoring in Owen et al, 1996; coordination of memory buffers in Rypma & D Esposito, 1999; attention switching in Garavan, Ross, Li, & Stein, 2000) Using an orientation 2-back WM task, we previously observed significant activation in the left DLPFC (Cornette et al, 2001) Based on this a priori knowledge, the right DLPFC involvement in the current Update task might be considered significant Indeed, although the hemispheric dominance was different in the two studies, these interhemispheric asymmetries were not significant The present results, however, demonstrate that the DLPFC neither have the same tight nor task-dependent coupling with the posterior parietal regions as the SFSm In addition, the more significant activation of the DLPFC in the 2-back task compared to the Update task may reflect a greater need for task coordination in the former task, in which an intense and continuous switching between the processes of updating and matching orientations is required upon probe presentation Hence, we suggest that the involvement of the DLPFC in the 2-back and Update tasks reflects executive control processes different from manipulative WM processes Functional Connectivity Just as in the study of long-term memory, it has been a long-standing issue whether WM can be localized to a specific cerebral region or whether it represents a distributed property of the brain Lashley s (1950) neuropsychological experiments in rats and monkeys provided the first indication of the distributed character of WM More recently, Goldman-Rakic (1987) provided evidence that WM processing relies heavily on the cooperation between anterior (prefrontal) and posterior (postcentral) regions Multiple reciprocal anatomical connections between unique sets of frontal areas and several subdivisions of parietal cortex of the monkey brain support this view of WM as a distributed function (Wise, Boussaoud, Johnson, & Caminiti, 1997; Cavada & Goldman-Rakic, 1989; Petrides & Pandya, 1984) In this perspective, functional imaging is extremely powerful since it stresses the network level of brain organization Several neuroimaging studies investigating WM (for review, see Smith & Jonides, 1999) suggest that the parietal cortex houses the storage buffers, while frontal regions are involved in rehearsal for refreshing the contents of the buffers, in addition to their manipulative and executive functions We therefore examined functional connectivity between the distributed regions involved in orientation WM Correlational analysis was indicative of a clear cooperation between SFSm and the precuneus plus the superior parietal regions Thus, the frontal regions involved in manipulating orientation information held in memory are indeed functionally connected to the regions putatively storing that information Furthermore, the functional link between frontal and posterior parietal regions switches to SFSl during orientation maintenance Human scalp EEG studies have indicated that synchronized neuronal activity between prefrontal and posterior regions occurs in the 4- to 7-Hz frequency range during WM maintenance (Sarnthein, Petsche, Rappelsberger, Shaw, & von Stein, 1998) We did not observe any significant correlation between the VLPFC and SFSm plus posterior parietal regions The latter observation makes sense, if one accepts that the VLPFC is mainly recruited to bridge long delays, independent of the load (Rypma & D Esposito, 1999) Hence, the functional connectivity reinforces the findings of the subtraction analysis since the frontoparietal regions involved in orientation WM have the links required by their function In summary, the current results agree with two major concepts in our understanding of WM functional organization First, a major organizational principle within the PFC is the distinction between maintenance and manipulative processes, which are subserved to different degrees by several major prefrontal WM regions (SFSl/ m, DLPFC, and VLPFC) Second, both processes involve a concerted action between distributed PFC and posterior regions In addition, several new insights emerge from the current data Most WM studies assume that the SFSl/m region is instrumental in maintenance The current study expands this view, indicating that this region can likewise accommodate manipulative processes, such as orientation updating In addition, the SFSl/m region is tightly coupled to the posterior storage regions, more so than the DLPFC and VLPFC Further work, preferably using high-resolution event-related fmri, is required to clarify the heterogeneous distribution of functions within SFSl/m and to specify the exact 822 Journal of Cognitive Neuroscience Volume 13, Number 6

11 function of DLPFC in Update and n-back tasks On a more speculative note, our findings may indicate that perhaps the true functional homologue of DLPFC in monkeys includes both DLPFC and SFSl/m regions of the human species, underlying our impressive capacity for abstract and flexible associative reasoning METHODS Subjects We studied 9 male volunteers (mean age 234 years, range 21 26) All subjects were right-handed as judged by the Edinburgh inventory They all had normal vision and a normal brain structure as visualized with MR imaging There was no history of neurological/psychiatric complaints, pathology, or drug abuse Prior to the PET session, subjects thoroughly practiced all tasks in two 15-hr sessions The study was approved by the Ethical Committee of the Medical School, Katholieke Universiteit Leuven, and written informed consent was obtained from all subjects in accordance with the Declaration of Human Rights, Helsinki, 1974 Stimulus Characteristics Stimulus characteristics were chosen in accordance with our previous study (Cornette et al, 2001) and were identical among all tasks: a static square wave grating (48 diameter, mean luminance 231 cd/m 2, contrast 90%, cycle width 138) presented in the central visual field Subjects were instructed to fixate a red central fixation point in all conditions Accuracy of fixation was monitored with EOG recordings Only orientations between and were used, that is, no vertical or horizontal orientations were shown to avoid verbal and semantic encoding Phase was randomized between trials and noise was superimposed on the edges of the bars to prevent subjects from using any cues other than orientation Duration of stimulus presentation was 500 msec in all conditions Stimuli were presented at a rate of 20 stimuli/min (interstimulus interval = 2500 msec) and were displayed on a high-resolution color screen (Philips Brilliance 2120, horizontal width 380 mm, vertical height 285 mm, resolution pixels, refresh rate 78 Hz noninterlaced), hosted by a 486 TIGA workstation The monitor was mounted above the scanner bed at an angle of 528 relative to the horizontal Subjects viewed the stimuli binocularly in a dimly lit room (007 cd/m 2 ) from a fixed distance of 114 cm Task Configuration We designed four tasks, namely, the Update task, Load3 task, Load6 task, and Identification task (ID) (Figure 5) During acquisition, each task lasted 150 sec and was performed three times All 12 conditions were presented in random order using a Latin square design Within each trial of the Load6 task, subjects were requested to retain 6 successively presented orientations (ie, one grating every 3 sec, giving a total maintenance period of 18 sec per trial) Debriefing of subjects showed that all 6 items were maintained by building up an abstract internal configuration in which the stored orientations were continuously compared to each other When an instructional cue was presented at the end of each series of 6 gratings (ie, a green central fixation point for 200 msec), subjects had to judge whether the orientation of the probe stimulus was a member of their current memory set (press right-hand key) or not (press left-hand key) The Load6 task corresponds to the maximal orientation memory set size, as psychophysically determined in pilot testing The Load3 task was a similar serial probe recognition task, but the memory set size was reduced to 3 items (ie, a 9-sec maintenance period) No monitoring of temporal order or any other manipulation of memoranda was required in either load task The length of the delay (>15 sec) between the presentation of each series and the probe prevented the intrusion of any serial position effect, that is, a faster response to a probe that corresponds to an item late in the series (Jensen & Lisman, 1998) In the Update task, adapted from Vanderlinden et al (1999), Salmon et al (1996), and Morris and Jones (1990), the number of orientations presented in each series pseudorandomly varied among trials from 3 to 7 Upon cue presentation, subjects judged whether or not the probe item was present in the last three orientations of the series just shown Since subjects were not informed of the series length before the start of a trial, the 3 items in store (= set size in Load3 task) had to be updated continuously as soon as the series length exceeded 3 To ensure maximal capture of the updating component, only series lengths 5 were presented during the 60-sec PET acquisition phase Within each trial of the Load3, Load6, and Update tasks, orientations shown differed by an angle d or a multiple of d, with each orientation shown only once Motor responses were allowed at any point during the interval between presentation of the probe stimulus and the first stimulus of a new trial, with the instruction being to press as soon as a decision was made On average, the number of motor responses during the Update task was programmed to match that of the Load3 task, that is, 5/ min The ID control task was adapted from the study of Orban et al (1997) In each trial of a given ID condition, only two oblique orientations were presented, tilted either d /2 (press right) or +d /2 (press left) degrees from a single internal reference orientation A different reference orientation was used in each of the 3 ID conditions Subjects needed to identify the orientation of each single grating presented, but were allowed to make a response only when instructed by a cue similar to that used in the other tasks The cue was shown Cornette et al 823

12 A/ ID B/ Load3 C/ Load6 D/ Update 0 n = 1 3,000 6,000 9,000 12,000 15,000 n 7 18,000 21,000 Time (msec) Figure 5 Schematic representation of stimulus timing in the four tasks ID task (A), Load3 task (B), Load6 task (C), and Update task (D) Duration of stimulus presentation is 500 msec in all conditions There are 20 stimuli/min in all tasks Orientations are shown within both the and ranges Small triangles indicate cue stimuli Each white rectangle outlining a grating in B D refers to a probe stimulus, while in A, it indicates the grating to be responded to Gratings shown for the ID task are 258, 198, 198, 198, 258, 258, 198 [d = 68] Gratings shown for the Load3 task are 508, 428, 268, probe 348, 1328, 1488, 1568 [d = 88] Gratings shown for the Load6 task are 1568, 1208, 608, 1328, 248, 728, probe 248 [d = 128] Gratings shown for the Update task are 188, 488,, 1488, 1188, 288, probe 688 [d = 108] Series length varies between 3 and 7 in each trial of the Update task within a 25-sec random interval following a grating presentation The number of motor responses (or presentations of instructional cues) during the three ID conditions matched once the Load3 task (ie, once every 9 sec), once the Load6 task (ie, once every 18 sec), and once the Update task (ie, on average once every 16 sec) Note that, to ensure the identification of each orientation shown, subjects were asked to press the keys within 600 msec after the cue presentation, according to the orientation just preceding the cue In order to equate the visual input among all tasks and subjects, the stimulus presentations during the Update and Load tasks were programmed to include as many orientations as possible within both ranges and to distribute them equally across all subjects (Kolmogorov Smirnov test for distribution matching) In addition, all reference orientations used in the ID tasks were chosen to cover the two orientation ranges as homogeneously as possible among subjects, taking into account each subject s orientation threshold Training Sessions Subjects were admitted for training only if their performance during a relatively easy Load6 task (d = 208) exceeded 80% correct During two 15-hr practice sessions, d was gradually decreased for each task, as soon as performance reached a steady level of 80 85% correct The d for which each subject reached 82% correct at the end of the second training session was then used in the corresponding condition of the PET study No auditory feedback was provided during training sessions or PET acquisition Statistical Analysis of Behavioral Data Task performance during PET sessions, expressed as percent correct responses, was normalized to Z scores (CSS software) and analyzed with analysis of variance (ANOVA) and post hoc comparisons for significance (Scheffé test) The number of memory tasks was in- 824 Journal of Cognitive Neuroscience Volume 13, Number 6

13 cluded as a repetitive factor within each subject The mean RTs were calculated as the average of all latencies corresponding to trials in which subjects responded within the response window Data Acquisition Brain activity was monitored as the relative change in regional cerebral blood flow (rcbf) using the H 2 15 O method (Fox et al, 1986) All measurements were performed in 3D mode with a Siemens-CTI Ecat Exact HR+ (Brix et al, 1997) Subjects were not allowed to speak during the procedure and had been instructed not to think of anything in particular, apart from concentrating on the stimulus and the task The room was kept as quiet as possible The head was immobilized with a foam headholder (Smither Medical Products, Akron, OH) Each subject had a catheter inserted into the left brachial vein for tracer administration Accuracy of fixation was monitored with EOG recordings, using contact electrodes placed on the outer ocular canthi and a reference electrode placed between the eyes Before each experiment, this EOG was calibrated for fixation and for horizontally visually guided saccades of 28 and 48 amplitude A transmission scan was taken ( 68 Ge rod sources) to correct for attenuation The start of each task coincided with the intravenously injection of 300 MBq H 2 15 O (half-life = 123 sec) over 12 sec Each subject performed all four tasks three times, with a 10-min interval between two successive injections Different orientations were used in the replications This yielded 12 emission scans per subject and 27 emission scans per task The order in which tasks were presented was randomized both within and between subjects Data acquisition (60 sec) began as soon as the intracranial radioactivity count rate rose sharply, that is, usually about 40 sec after injection Total duration of each task performed was 150 s The attenuation corrected data were reconstructed using the reprojection algorithm (Kinahan & Rogers, 1989) resulting in 63 planes (plane separation 2425 mm) The integrated radioactivity counts were used as a measure of rcbf Data Analysis Analysis was done on Sun SPARC computers (Sun Microsystems, Mountain View, CA) with the Statistical Parametric Mapping software (Wellcome Department of Cognitive Neurology, London, UK), version SPM96, implemented in MATLAB (Mathworks, Sherborn, MA) Realignment and Spatial Normalization The scans from each subject were realigned using the first scan as a reference The 6 parameters of this rigid body transformation were estimated using a least-square approach Images were subsequently stereotactically transformed to a standard template in the Talairach space (Talairach & Tournoux, 1988) This normalizing spatial transformation matches each scan (in a leastsquare sense) to a reference or template image that already conforms to the standard space The procedure involves a 12-parameter affine (linear) and quadratic (nonlinear) 3-D transformation, followed by a 2-D piecewise (transverse slices) nonlinear matching, using a set of smooth basic functions that allows for normalization at a finer anatomical scale (Friston et al, 1995) Finally, images were smoothed with an isotropic Gaussian kernel of 16 mm full width at half maximum (FWHM) The final image smoothness estimates (FWHM) were x = 140 mm, y = 168 mm, and z = 187 mm Statistical Analysis Statistical parametric maps (SPMs) are spatially extended statistical processes used to characterize regionally specific effects in imaging data, combining the general linear model (to create the statistical map of SPM) and the theory of Gaussian fields (to make statistical inferences about regional effects) (Worsley, Evans, Marrett, & Neelin, 1992; Friston et al, 1991; Friston, Worsley, Frackowiak, Mazziotta, & Evans, 1994) In order to take into account the subject-to-subject variability in response, the data were analyzed using the random-effects model as provided with SPM96 (RandFx tool kit) Implementation of this model required the construction of adjusted means for each condition, resulting in one condition per task for each subject The condition, subject and covariate effects are estimated according to the general linear model at each voxel To test hypotheses about regionally specific condition effects, the estimates were compared using linear contrasts The resulting set of voxel values for each contrast constitutes a SPM of the t statistic SPM{t} These SPM{t} values were then transformed to the unit normal distribution (SPM{Z}) Activations reaching p corr < 05 (corrected for multiple comparisons) for peak height (Z 440) were considered significant Regions significant at p uncorr < 001 (uncorrected for multiple comparisons) for height (Z 309) were considered equally significant only if based upon an a priori hypothesis Other activation sites significant at p uncorr < 001 for height were added for descriptive purpose MR Imaging Template Each subject also underwent a high-resolution MR imaging scan of the brain, using a 3-D Magnetization Prepared Rapid Gradient Echo (MPRAGE) sequence (Mugler & Brookeman, 1990) Acquisition parameters were: repetition time 10 msec, echo time 4 msec, flip angle 88, field of view 256 mm, acquisition matrix The 3-D volume had a thickness of 160 mm, Cornette et al 825