|Home | About | Journals | Submit | Contact Us | Français|
Spatial attention influences representations in visual cortical areas as well as perception. Some models predict a contrast gain, while others a response or activity gain when attention is directed to a contrast varying stimulus. Recent evidence has indicated that microstimulating the Frontal Eye Field (FEF) can produce modulations of V4 neuronal firing rates that resemble spatial attention-like effects, and we have shown similar modulations of functional magnetic resonance imaging activity throughout the visual system. Here, we used fMRI in awake, fixating monkeys to first measure the response in twelve visual cortical areas to stimuli of varying luminance contrast. Next, we simultaneously microstimulated sub-regions of the FEF with movement fields that overlapped the stimulus locations and measured how microstimulation modulated these contrast response functions (CRFs) throughout visual cortex. In general, we found evidence for a non-proportional scaling of the CRF under these conditions, resembling a contrast gain effect. Representations of low contrast stimuli were enhanced by stimulation of the FEF below the threshold needed to evoke saccades, while high contrast stimuli were unaffected or in some areas even suppressed. Further, we measured a characteristic spatial pattern of enhancement and suppression across the cortical surface, from which we propose a simple schematic of this contrast-dependent fMRI response.
In the primate visual system, cortical sensitivity to features such as luminance contrast varies with the locus of spatial attention (Kastner and Ungerleider, 2000; Reynolds and Chelazzi, 2004). For example, attending to an isolated stimulus alters the effective luminance contrast detected by a neuron, modulating the neuron’s contrast response function (CRF). Different models have been proposed to describe these modulations: one model suggests that attention causes a non-proportional scaling across contrasts, producing a leftward shift of the CRF or effective contrast gain (Reynolds et al., 2000; Martinez-Trujillo and Treue, 2002; Reynolds and Chelazzi, 2004). Another model proposes a proportional scaling of neuronal responses with attention, which produces either a response or activity gain effect (Williford and Maunsell, 2006). Very recent work has suggested these two models may not be competitive but rather derived from a common circuit (Reynolds and Heeger, 2009), though neither offers insight into the origin of the signals causing these modulations.
Several lines of evidence suggest that the top-down sources for the presumptive feedback signals mediating spatial attention are tightly linked with the oculomotor system (Rizzolatti et al., 1987; Corbetta and Shulman, 2002; Hamker, 2005). In particular, the Frontal Eye Field (FEF) is thought to be involved in attention-dependent modulations of sensory input. FEF would be an excellent candidate for sending top-down signals to visual cortex because of its role in visual target selection (Wardak et al., 2006), defined retinotopic map of visual space (Bruce et al., 1985) and reciprocal connections with many cortical visual areas (Huerta et al., 1987; Schall et al., 1995; Stanton et al., 1995). Experiments using electrical microstimulation of FEF (FEF-EM) have shown: (i) decreased contrast detection thresholds during a spatial attention task (Moore and Fallah, 2004), and changes in both (ii) neuronal firing rates in area V4 (Moore and Armstrong, 2003) and (iii) functional magnetic resonance imaging (fMRI) activation in multiple visual areas (Ekstrom et al., 2008). Both the type of stimulus used (preferred or non-preferred) in the single unit study and the spatial pattern of modulation observed across the cortical surface in the fMRI study suggest that such modulations may depend on the strength of neuronal drive produced by a stimulus, similar to attentional effects on a CRF.
To measure the effects of luminance contrast on FEF-EM induced modulations of activity throughout visual cortex, we collected cerebral blood volume (CBV)-weighted fMRI data from two awake monkeys while simultaneously stimulating FEF. We first measured the visually-driven fMRI response to stimuli of varying contrast throughout visual cortex. With the stimuli positioned in the movement fields (MFs) corresponding to the locations of chronically implanted FEF electrodes, we then measured the fMRI response to combined visual stimulation and FEF-EM and characterized how microstimulation altered each CRF. In general, we observed the largest enhancement of activity at lowest contrast, while high contrast stimuli were unaffected or in some cases even suppressed by FEF-EM. These results support a contrast-gain model (Reynolds et al., 2000), and we generalize them to produce a spatial schematic of this contrast-dependent fMRI response across the cortical surface.
Many of the details of the primate imaging and stimulation techniques used have been described previously (Vanduffel et al., 2001; Nelissen et al., 2006; Ekstrom et al., 2008), and will only be summarized here briefly. All procedures were approved by MGH’s Subcommittee on Research Animal Care (Protocol #2003N000338) and MIT’s Committee on Animal Care, and are in accordance with NIH guidelines for the care and use of laboratory animals.
In short, we prepared two male rhesus monkeys (Macaca mulatta; MM1 and MM2, 5-7 kg, 4-5 years old) for fMRI with surgical implantation of a MR compatible head post, and then trained them for a passive fixation task. Once the subjects achieved accurate fixation performance, we chronically implanted 26 intracortical microelectrodes in the right FEF. Guided by anatomical MR images, we made a craniotomy over the right arcuate sulcus and opened the dura mater. Twenty six Teflon-coated microwires (25 μm diameter, 10% platinum / 90 % iridium; California Fine Wire Company; Grover, CA) were inserted by hand normal to the cortical surface (Mioche and Singer, 1988) along the rostral bank of the arcuate sulcus (~2-6 mm deep). Wire tips were beveled and ~40 μm of insulation stripped prior to insertion. The dural and bone flaps were then replaced. We soldered the microwires to a magnet-compatible connector (Omnetics Connector Corporation; Minneapolis, MN), which was then encased in dental acrylic as part of the subject’s head set. A ground electrode was implanted between the skull and muscle and attached to the connector. After surgery, we verified the correct anatomical placement of the electrode tips in the FEF using T2-weighted images collected at 7.0 T. We also obtained behavioral verification, in the form of successfully EM-induced saccadic eye movements, during a fixation task outside the scanner. These eye movements defined for the neurons surrounding each electrode the so-called FEF MF and the stimulation threshold to evoke a saccade [see also (Ekstrom et al., 2008)].
High resolution, T1-weighted anatomical images were collected on a 3.0 T full-body scanner (Trio, Siemens Healthcare; Erlangen, Germany) to guide electrode implantation and for the overlay of functional analyses. Under ketamine-xylazine anesthesia, a MP-RAGE sequence (178 sagittal slices, 256 × 256 in-plane matrix, TR = 2.5 s, TE = 4.35 ms, TI = 1100 ms, 0.35 × 0.35 × 0.35 mm3 isotropic voxels, flip angle = 8°) was used to obtain 9 whole-brain volumes, which were averaged together to improve the signal-to-noise ratio. A single radial transmit-receive surface coil (12.5 cm diameter) was employed.
Functional images were acquired on a 3.0 T full-body scanner (TIM Trio, Siemens Healthcare; Erlangen, Germany) with a gradient coil insert (AC88; 80 mT/m maximum gradient strength, 800 T/m/s maximum slew rate), using a gradient-echo T2*-weighted echo-planar imaging sequence (50 horizontal slices, 84 × 96 matrix, TR = 2 s, TE = 19 ms, 1 × 1 × 1 mm3 isotropic voxels). A four-channel phased array receive coil (individual coils 6 cm diameter), with GRAPPA reconstruction (Griswold et al., 2002) and an image acceleration factor of 2, and a saddle-shaped, radial transmit-only surface coil (17 cm diameter) were employed (H. Kolster et al., unpublished observations).
Prior to each session, we injected a bolus of Microcrystalline Iron Oxide Nanoparticles (MION; 6-10 mg/kg) in isotonic sodium citrate into the femoral or saphenous vein to increase the contrast-to-noise ratio compared to blood oxygen level dependent imaging (BOLD) (Vanduffel et al., 2001; Leite et al., 2002). Accordingly, we have flipped the polarity of all percent signal change values to account for the difference between MION Cerebral Blood Volume (CBV) and BOLD activation maps (increased brain activation produces a decrease in MR signal in MION CBV maps).
In many cases, we were able to generate robust and well-localized fMRI activation maps for a given stimulus location from a single session (for example, see Fig. 1D). We obtained a high MR contrast-to-noise ratio (CNR) largely due to the use of: (i) the four-channel phased array receive coil that enabled two-fold image acceleration, which increased the SNR (Griswold et al., 2002) and reduced susceptibility distortions at the air-tissue interfaces around the brain, and (ii) the gradient insert coil, which increased the strength and slew rate of the MR scanner’s gradients to permit shorter TR and TE values in our scanning sequence, so that we could collect more data per unit time. A shorter TE value also reduced the overall T2* decay caused by the MION contrast agent, yielding further increases in image SNR (Mandeville et al., 2004).
We limited our design to the five contrast values shown in Figure 1A (contrast levels of 0, 3, 6, 12 and 50%), chosen to provide a general overview of the contrast response function. The instances where multiple sessions were needed for a given stimulus were largely due to monkey behavior because of the long runs required by our block stimulus design — five different luminance contrasts × two electrical stimulation conditions (with and without EM stimulation) × two repetitions yielded individual runs of 20 epochs each 24 s long (12 TR), resulting in runs of 480 s. In total, 24 960 functional volumes were collected from MM1 (6 sessions) and 23 280 functional volumes from MM2 (6 sessions), of which 15 360 and 18 600 volumes respectively, were used to generate the data shown.
Visual stimulation was presented to the subjects at 1024 × 768 resolution and 60 Hz refresh rate with a LCD projector (DLA-SX21, JVC; Yokohama, Japan) onto a translucent screen 51 cm from the animals’ eyes. Prior to the beginning of our first experiment, we measured the luminance response of the projector’s red, green, and blue channels combined and separately with a photometer (CS-100 Chroma Meter, Minolta; Tokyo Japan). From these response curves, we selected an approximately linear operating region, and used the combined curve to calculate luminance values to generate stimuli at the desired contrasts. Compared to the stimuli in our previous study (Ekstrom et al., 2008), here we have discarded color information, and increased the size of the stimuli to reduce the possibility of any mismatch between onscreen stimulus location and the corresponding FEF MF. Stimuli were 6 visual degrees diameter, monochromatic, square-wave, contrast-varying gratings, where we defined contrast as the ratio between the difference and sum of luminance in the bright and dark regions (Michelson, 1927). The stimuli had a spatial frequency of 0.5 cycles / degree, were oriented at either 0 or 90 degrees and moved at 2 degrees / s perpendicular to the axis of orientation; mean luminance was matched to that of the uniformly gray background (76.9 cd/m2).
Stimulus location was matched to the MF of the single electrode stimulated in that particular session (Moore and Armstrong, 2003; Ekstrom et al., 2008). Note that this single stimulus method differs from the multiple sequential stimuli and FEF sites approach used in our previous study, which produced fMRI activation maps averaged across all stimuli presented and electrodes used within a TR. While such a sequential presentation is arguably closer to the physiological operating state of the visual system and produces very robust EM-only effects, the results when only one stimulus is used may be easier to interpret. Here, we used only a single stimulus-FEF electrode pairing in each session and concatenated the data after GLM analysis to calculate CRFs averaged across the stimulus-electrode pairings in both subjects.
We used a pseudo-randomized block design with multiple stimulus orders. In the epochs with FEF-EM, we stimulated the selected electrode twice within a TR (ISI = 1 s). The visual stimulus was presented for 133 ms, followed by 250 ms of combined visual stimulation and FEF-EM and then 617 ms with neither; this sequence was repeated twice within the 2 s TR. A central fixation point was continuously visible and the monkeys performed a passive fixation task throughout each run. Eye position was monitored at 120 Hz using an infrared pupil / corneal reflection tracking system (Iscan Inc.; Burlington, MA). The EM signal was generated by an eight-channel Digital Stimulator (DS8000, World Precision Instruments; Sarasota, FL), controlled by custom software that also generated the visual stimulation. Stimulation trains consisted of biphasic square-wave pulses with a frequency of 335 Hz for 250 ms, and were delivered in a monopolar configuration. Each pulse consisted of 190 μs of positive and 190 μs of negative voltage, separated by 100 μs of zero voltage. The sum of the cables’ stray capacitance and the monolithic capacitor array EMI filters in the scanner penetration panel ranged from 5.1 to 5.3 nF.
Prior to each fMRI experiment, we determined the threshold to elicit a saccade with each electrode by varying the EM amplitude until ~70% of stimulation trains induced a saccade from central fixation. A saccade vector was determined to identify the MF of the FEF site of interest for that experiment (see Fig. 1B for MM1’s MFs and Fig. 1C for MM2’s MFs used in the current study); a stimulation level of ~50% of this behaviorally-defined threshold was then used for the actual imaging experiment (Moore and Armstrong, 2003). At the beginning and end of each fMRI session, we measured the impedance of the stimulated channel, as well as the impedance of the three channels not used, with a 1V, 100 Hz reference signal to estimate the injected current. For MM1, impedance was 32 ± 2 kΩ, 35 ± 1 kΩ, 48 ± 2 kΩ, and 55 ± 3 kΩ for the 4 electrodes used; for MM2, impedance was 24 ± 1 kΩ, 27 ± 1 kΩ, 36 ± 1 kΩ, and 42 ± 11 kΩ for the 4 electrodes used (mean ± standard deviation across all sessions). Estimated stimulation amplitudes for the four FEF electrodes used in MM1 were 26 μA, 36 μA, 55 μA and 72 μA and in MM2 37 μA, 38 μA, 49 μA and 58 μA. We note that the current necessary to evoke saccades through chronic electrodes may be larger than with higher impedance acute electrodes. One reason for this is the possible growth of a fibrous barrier between the electrodes and neural tissue over the 2-3 years between implantation and collection of the present data set (Bartlett et al., 2005). Another reason may be the use here of short pulse durations (0.19 ms) compared to those used (0.20 – 0.25 ms) in previous studies [e.g., (Bruce et al., 1985)].
We performed a voxel-based analysis with SPM99, following previously described procedures to fit a general linear model (GLM) (Friston et al., 1995; Vanduffel et al., 2001; Leite et al., 2002; Vanduffel et al., 2002). Images were motion-corrected within session and non-rigidly co-registered to each subject’s own anatomical template using Match software (Chef d’hotel et al., 2002). Images were then smoothed (Gaussian kernel, σ = 0.67 mm) and registered using Match to MM1’s T1-weighted anatomical volume. Global scaling, and high- and low-pass filtering were employed prior to fitting the GLM.
We used the main visual effect at 50% contrast (that is, 50% VEM + 50% V versus 0% VEM + 0% V, Fig. 1A) as a localizer to identify visually driven voxels (hereafter referred to as visual voxels) for analysis across all contrast levels. t-score maps of this contrast from each session in each monkey were thresholded [P < 0.05, uncorrected; for split population analysis, low t voxel distribution is 0.05, uncorrected > P > 0.05 corrected, high t voxel distribution is P < 0.05, corrected; correction for multiple comparisons used the Family-Wise Error procedures in SPM99 (Worsley et al., 1996)] and overlaid on the common anatomical space defined by MM1’s T1-weighed anatomical volume as reconstructed by FreeSurfer (Dale et al., 1999). We created flattened cortical representations with Caret using the F99 atlas (http://www.nitrc.org/projects/caret/, http://sumsdb.wustl.edu/sums/macaquemore.do) (Van Essen et al., 2001) (see Fig. 1D for an example of one session). The borders of twelve visual areas were identified on this flattened cortical representation using retinotopic mapping data previously collected in three animals (Fize et al., 2003) and a Caret Atlas based on two previous studies (Ungerleider and Desimone, 1986b; Felleman and Van Essen, 1991). We extracted the population of visual voxels inside each area satisfying the above contrast and sampled the percent change in MR signal for each condition with SPM99. The first 3 TRs in each epoch were excluded and an additional TR appended at the end to compensate for hemodynamic delay. Standard error of the mean was calculated over the condition epochs.
To account for any head and eye movement-related fMRI activity, covariates of no-interest from six motion realignment parameters and the x and y eye trace components were generated and included in the GLM analysis. Eye traces were thresholded at approximately 2 degrees, convolved with the MION hemodynamic response function, and sub-sampled to the TR. We performed several analyses to assess the potential effects of eye movements on fMRI activation. First, for each run collected, we aligned the stimulation events across all contrast levels to EM onset and the visual-only conditions to the matching time. We then performed a t-test at each time point to identify any significant differences in eye position (P < 0.05, corrected for multiple comparisons across time points; two sample, two-tailed t-test across all trials in that run). Any run with a significant difference at any time point was excluded from further analysis. The mean x and y eye position aligned to EM onset across all sessions is illustrated in Figure 2. Second, for each subject’s aggregate data we utilized a one-way ANOVA to test for any significant differences in percent fixation, saccade rate, and eye position along the x and y axes between the VEM and V conditions across all contrast levels (Table 1).
where c is the fractional contrast level, c50 the contrast at which the response curve attains half of its maximum, Rmax the maximum response of the region and b the baseline response value. When fitting the V conditions, we omitted the baseline response parameter because the 0% V condition was used as reference. This equation is a simplified form of the more general one used to fit contrast response data from single cell recordings (Albrecht and Hamilton, 1982). As an estimate of how well eq.  described the fMRI response in each region, we calculated an adjusted R2 value (Radj2); this quantity describes how much of the variance in the data is captured by the model, with a penalty for the use of additional model parameters. To estimate 95% confidence intervals for all parameters we implemented a non-parametric bootstrap algorithm (Efron and Tibshirani, 1993). In a particular region, the visual voxel population was randomly sampled with replacement and the best fitting parameters calculated for that particular sample. We repeated this sampling 1000 times and determined the 2.5% and 97.5% quantile values from the resulting parameter distribution to indicate the confidence interval.
As an additional estimate of the difference between the distributions of V and VEM percent change in MR signal values at a particular contrast level for the voxels in a given functional brain region, we calculated the area under a receiver-operator-characteristic (ROC) curve (McNicol, 1972). A ROC curve is determined by plotting the percentage of one distribution less than a given criterion value as a function of the percentage of the second distribution less than that same criterion, for all possible criterion values in the distributions. The area under the ROC curve then reflects how separate or distinguishable the two distributions are. An area of 0.5 means the V and VEM distributions could not be discriminated at that contrast level in that region; deviations from 0.5 indicate increasing discriminability, with maximum possible values of 0 or 1. To include the 0% contrast values in this analysis, we recalculated all our signal change values with respect to the mean fMRI signal value across all 10 stimulus conditions rather than the 0% V baseline condition prior to computing ROC curves. To estimate 95% confidence intervals on each of the ROC areas, we again used the non-parametric bootstrap algorithm (n = 1000) (Efron and Tibshirani, 1993) described above.
In both monkeys, we collected data associated with four stimulus locations (Figs. 1B and 1C), corresponding to four implanted FEF electrodes and associated MFs, over six fMRI sessions. Stimuli were spread across the left visual field, contralateral to the side of the brain in which we implanted the electrodes, with examples in the upper and lower visual field and along the horizontal meridian. Eccentricities ranged from 1.8 visual degrees to 15.0 visual degrees in MM1, and from 3.5 visual degrees to 7.3 visual degrees in MM2. Saccade vectors were for the most part larger in MM1 than in MM2, suggesting that we stimulated slightly different parts of the FEF in the two subjects. Because of this small number of samples we did not attempt to distinguish between eccentricities and polar angles, and instead included all stimulus locations in one analysis (Figs. 1B and 1C).
For MM1, data from three identical sessions were combined for one target location, and data for the remaining three stimuli were each collected in a single session. For MM2, data for two stimulus locations were collected in individual sessions and the remaining two locations were each collected over two sessions. On average, we collected 17.3 and 16.2 runs per session in MM1 and MM2, of which 10.7 and 12.9 runs per session were included in the analysis presented here. Inclusion criteria were based on a run by run analysis of the eye monitoring data recorded during each run. Runs in which the monkey’s eye position at any time point in the stimulation sequence (1 s, sampled at 120 Hz) was significantly different between V and VEM epochs were excluded from further analysis (see Materials and Methods, Fig. 2).
For each stimulus location, we identified a population of visual voxels for further analysis from the fMRI contrast defined by the main visual effect at the highest luminance contrast (that is, 50% VEM + 50% V versus 0% VEM + 0% V). Using this population, we extracted the percent change in MR signal with respect to the fixation only or 0% V condition for each voxel at each of the other luminance contrast levels presented to the subjects. For each cortical area of interest, we then pooled the population of visual voxels across all sessions from both subjects and computed the mean response as a function of contrast to generate a V CRF.
Figure 3 shows the average CRF for each of the twelve cortical areas identified, grouped according to approximate cortical hierarchy (Felleman and Van Essen, 1991) and plotted on the same scale for comparative purposes. The smooth curves were calculated from the best fitting model to eq.  with the baseline parameter b omitted (see Materials and Methods). Figure 3A shows the response for early visual areas, including areas V1, V2, V3 and V3A. The CRFs for V2 and V3 were relatively similar to each other, as were the CRFs for V1 and V3A. Figure 3B shows the response functions for visual areas in the ventral stream leading into inferotemporal (IT) cortex, including areas V4, TEO and TE. The V4 CRF was qualitatively similar to the response seen for area V3. The two areas located more rostrally in IT cortex showed a decreased response to the contrast stimuli, with area TE showing the lowest contrast response. This low level of fMRI activity was likely because our gratings were non-optimal stimuli for driving these more object sensitive areas (Vogels and Orban, 1994).
Figure 3C shows the response functions for visual areas in the STS more commonly identified with the dorsal stream, such as areas MT and MST, as well as area FST. Area MT showed the highest contrast sensitivity and the largest overall fMRI activation of all the areas sampled (mean response of 0.578 ± 0.011 percent change in MR signal across all contrasts). MST and FST had lower contrast responses, with area FST being the least responsive of this group, though qualitatively quite similar to area MT in terms of change in response between contrasts. Figure 3D shows the response functions for the two highest-order visual areas sampled, area LIP in the lateral bank of the intraparietal sulcus and area STP in the upper bank of the STS. These two areas were the least responsive of all to luminance contrast, perhaps reflecting their role in integrating information from multiple sensory modalities (mostly STP) or their large receptive field sizes compared to the small stimuli size we used.
Despite considerable differences in methodology and the stimuli used, we attempted to compare the CRFs obtained here with CRFs collected in previous human (Tootell et al., 1995; Buracas and Boynton, 2007; Li et al., 2008) and non-human primate studies (Albrecht and Hamilton, 1982; Sclar et al., 1990; Williford and Maunsell, 2006; Palmer et al., 2007) (see Figure S1 and discussion in the accompanying Supplemental Material). In general, we found good agreement with past human fMRI studies in early visual areas for which human fMRI data are available. In area MT, our monkey fMRI data matched more closely the single unit results than the human fMRI data.
Table 2 gives a summary of the best fitting parameters for each area from the Naka-Rushton model, with 95% confidence intervals estimated from a non-parameteric bootstrap algorithm (n = 1000). Several of the qualitative trends described above are apparent in these modeling results. Area MT had the sharpest rising CRF, and accordingly the lowest c50 value, indicating that this area is the most contrast sensitive of all visual areas measured. Extrastriate areas V2, V3 and V4 all had similar CRFs and thus relatively similar model parameters. Some of the higher order areas that showed the least responsivity to contrast (LIP, TE and STP in particular) had the highest c50 and lowest Rmax values.
From the same population of visual voxels, we also extracted the percent change in MR signal for the epochs that combined a visual stimulus with simultaneous electrical stimulation of the FEF location with a matching movement field. We again calculated a mean response at each contrast level, across all stimulus locations in both subjects, to generate a VEM CRF. Figure 4A shows for area V1 the resulting VEM CRF, along with the V CRF from Figure 3A for comparison. Note that in this and following figure, we plot only the measured mean VEM and V responses and their derivatives, rather than any fitted functions, to give the most accurate indication of how FEF-EM modulated the cortical contrast response. For area V1, the FEF-EM dependent modulation of visually driven activity varied as function of the underlying luminance contrast of the visual stimulus. For the low contrast stimuli, we measured an increase in mean fMRI activation (from 0.094 ± 0.010 to 0.221 ± 0.011 percent change in MR signal at 3% luminance contrast, without and with FEF-EM, respectively), whereas for the highest contrast stimuli we measured a decrease in mean fMRI activation (from 0.816 ± 0.011 to 0.695 ± 0.010 percent change in MR signal at 50% luminance contrast). The crossover point, when activation in the visual-only condition would presumably exceed activation in the combined visual and FEF-EM condition, was somewhere between 6 and 12% luminance contrast.
Next, we computed the difference between these two CRFs, shown by the black trace in Figure 4B, and used a two-way ANOVA to assess whether there was a significant interaction between FEF-EM and visual stimulation, relative to the 0% V and 0% VEM conditions (P < 0.05; factors EM and visual stimulation). A positive interaction, here indicating a larger response in the combined VEM condition at a given contrast than the sum of the responses in the V only condition at that same contrast and the 0% VEM condition, is noted by the presence of a black filled square at each contrast level. A negative interaction, here indicating the opposite effect, is noted by an open black symbol. In area V1, we found that there was a significant interaction at the 3% (positive), 12% (negative) and 50% (negative) luminance contrast levels. In agreement with our previous work (Ekstrom et al., 2008), no significant difference was found in V1 at the 0% contrast level, which indicates no EM-only effect or so-called baseline shift in this region (P > 0.05; one sample, two tailed t-test).
After comparing the difference between the VEM and V epochs, we also computed the fractional change this difference represents in terms of the underlying visual-only activation. For a particular area, we reasoned that a large change to a small initial activation would convey more information to any region reading out these signals than a change of equal magnitude on top of an already large activation. Figure 4C shows this fractional change as a function of contrast for area V1. The use of the 0% V condition as our reference point precluded us from calculating this value at 0% contrast (though, as shown in Fig. 4B, the difference between V and VEM at this contrast is clearly quite small), but for all other contrast levels we calculated a fractional change response function. Using this measure, we found that the modulations due to FEF-EM at the lowest contrast became even more apparent: the change in activation caused by FEF-EM in V1 represented a 1.4 ± 0.3 fold increase in fMRI activity. Conversely, the response at high contrast decreased by 0.15 ± 0.02 times the initial visual-only activation, which is a much smaller effect than that observed at low contrast.
In addition to modulations of the mean response, another way to characterize how well an ideal observer could read out changes due to FEF-EM is to determine how separate or distinguishable the distributions of V and VEM signal change values are for the population of visual voxels. We estimated this distinguishability by implementing a ROC (receiver-operator-characteristic) analysis (see Materials and Methods). Figure S2 shows a series of ROC curves for each contrast level and Figure S3A (black trace) the resulting area under each ROC curve as a function of luminance contrast for area V1. In general, we observed in area V1 a very similar trend with this measure as with the arithmetic difference measure. The VEM and V signal change distributions were most distinguishable for the lowest contrast stimuli, with VEM greater than V, and for the highest contrast stimuli, with V greater than VEM. The largest suppressive effect of FEF-EM at high contrast (i.e. V greater than VEM) was observed in area V1.
We extended the above analysis to all visual cortical areas which were driven by the 50% contrast gratings. Figure 5 shows difference CRF and fractional change plots for the twelve cortical areas first identified in Figure 3, with the corresponding plots for V1 repeated from Figure 4 for comparative purposes. We have grouped areas together that are close to each other in terms of cortical hierarchy and that show similar qualitative features in their difference CRFs. In all areas except STP, we observed the largest increase in fMRI activity in the VEM epochs at the 3% luminance contrast level, while the highest contrast stimuli produced no or even negative interactions, which indicate a reduction in fMRI activity due to EM. We think these high contrast observations are not likely attributable to a ceiling effect of our fMRI measurements, given that we found a decrease in activation in some regions and were able to measure much larger percent signal change values in some areas in our previous study [e.g., Fig. 1C in (Ekstrom et al., 2008)]. In several areas, including V3, V4, MT, MST, FST, LIP and STP, we measured a significant baseline shift due to FEF-EM (P < 0.05; one sample, two tailed t-test). This group of areas is mono-synaptically connected with the FEF (Huerta et al., 1987; Schall et al., 1995; Stanton et al., 1995) and so a robust, fMRI-detectable, EM-only effect is expected here (Tolias et al., 2005; Ekstrom et al., 2008; Moeller et al., 2008). In four areas, including V2, MT, MST and FST, we observed a significant negative interaction at the 6% contrast level that does not seem to match the trend established by the next lower and higher contrasts. Although we have no obvious explanation for this effect observed in both monkeys, we note that this feature is common to mainly dorsal stream areas which are strongly interconnected anatomically (Boussaoud et al., 1990; Ungerleider and Desimone, 1986a; Ungerleider and Desimone, 1986b).
As we did for the V only conditions, we fit the Naka-Rushton model (eq. ) to the mean VEM response in each area (see Table S1 in Supplemental Material). In this instance we included the baseline parameter b in the model fit. In general, eq.  provided a less satisfactory description of the data in the VEM conditions than in the V conditions. This deficiency is indicated by the general decrease in Radj2 values for the VEM model fits compared to the V fits. Figure 4A illustrates well the features of the data responsible for this difference in fit quality. The Naka-Rushton model (Naka and Rushton, 1966) and more general hyperbolic ratio models (Albrecht and Hamilton, 1982) produce a smooth sigmoid function that seems to characterize well the visual only CRF in cortical visual areas. With the addition of FEF-EM, however, the model does not well capture both the pronounced rise in activity that we observed at 3% contrast coupled with the decrease in activity at high contrast. As a result, the best-fitting baseline parameter b in V1 (Table S1) is found to be non-zero, which disagrees with our earlier statistical analysis and our previous findings (Ekstrom et al., 2008). Further, many of the c50 values (Table S1) do not reflect the downward shift expected by the increase in activity observed at 3% contrast. Thus, although eq.  appears to adequately model the V data, this does not hold not true for the VEM data in all visual areas.
As before, we computed the fractional change response functions for all twelve areas, by dividing the difference between VEM and V epochs by the underlying visual-only activation. These response functions are presented in Figures 5D to F, following the same grouping of areas used above. The response function for V1 is repeated from Figure 4C for comparative purposes. These curves further accentuate the pattern seen in the preceding graphs. For most areas, the largest enhancement of fMRI activity by FEF-EM above the underlying visual response was seen for the lowest contrast stimuli used. In Figure S4, we show this same analysis for MM1 and MM2 individually. Despite the relatively different distribution of saccade amplitudes between the subjects, we observed roughly the same modulation pattern in both. At the highest contrast though, we do see more examples of suppression of fMRI activity in MM2 than in MM1.
Finally, for each area we computed a ROC curve using the VEM and V signal change distributions at each contrast level, and associated areas under the ROC curve (Fig. S3). This analysis produced largely the same results as the arithmetic difference measure outlined above.
By pooling all voxels from the identified cortical areas, we calculated an overall CRF, difference CRF and fractional change function for all visually driven voxels in occipital cortex. The overall CRF (black circles in Fig. 6A) gives an accurate estimate of the overall response of the visually driven areas in cortex, although it is biased towards the larger visual areas. The black trace in Figure 6B shows the overall difference CRF; black squares indicate the contrast levels with a significant interaction between EM and the visual response (P < 0.05; two-way ANOVA). We observed a significant positive interaction at 3% luminance contrast and a significant negative interaction at 50% luminance contrast. In addition, we measured a significant overall baseline shift at 0% contrast (P < 0.05; one sample, two tailed t-test). Figure 6C shows the corresponding fractional change response function.
In an additional analysis, we split our population of visual voxels into highly and weakly visually responsive groups, on the basis of each voxel’s t-score in the localizer contrast (that is, 50% VEM + 50% V versus 0% VEM + 0% V). All voxels that exceeded a P value of 0.05 at the corrected level were included in the high population (in other words, visually very well driven voxels). The remaining visual voxels (that is: 0.05, corrected < P < 0.05, uncorrected) were included in the low population. This split resulted in an approximately equal-sized pool of voxels in each population. In Figure 6A we also present the CRFs for these two subdivisions of our larger population along with the best fitting model for each group (red and blue curves). Figure 6B also shows the difference CRFs for these additional populations, and Figure 6C the fractional change functions for each group. Table 3 gives the best fitting model parameters to the V only conditions and Table S2 the fit parameters for the VEM responses. Both the weakly and strongly driven voxels exhibited a baseline shift, or an increase in activity due to FEF-EM in the absence of visual stimulation (P < 0.05; one sample, two tailed t-test). Further, we observed a remarkable difference in the effects of FEF-EM on the CRF of the low and high population of voxels. The low population displayed a significant positive interaction at 3% and 12%, and a significant negative interaction at 50% (P < 0.05; two-way ANOVA). For the high population, we also found a positive interaction at 3%; however, we found a negative interaction at all three higher contrast levels (P < 0.05; two-way ANOVA).
The results of the split-population analysis are summarized schematically in Figure 7. The colored circles represent the two populations of voxels on the cortical surface, with the highly responsive voxels in the center and the weakly responsive voxels in the margin of the schematic. The lower panels of Figure 7 show a projection of the activity profile onto a single spatial dimension. For the lowest contrast stimuli, FEF-EM amplified the activity of both weakly and strongly driven voxels (compare solid, V-response curve and dotted, VEM-response curve, Fig. 7B). For the highest contrast stimuli (Fig. 7D), the macroscopic circuit switched to a discriminatory mode; the activity of all voxels was slightly suppressed, though weakly driven voxels were suppressed more. For intermediate contrasts (Fig. 7C), the results show a mixture of effects: the activity of weakly responsive voxels was still amplified, while strongly responsive voxels were slightly suppressed.
Using CBV-weighted fMRI, we have measured response functions to visual stimuli at five luminance contrast levels in twelve defined functional areas of the macaque visual cortex. Further, we observed how this contrast response was modulated by the simultaneous stimulation of FEF sites with movement fields at the location of the visual stimuli. We observed a significant baseline shift (i.e. at 0% contrast) due to FEF-EM in a subset of areas known to be well connected to the FEF, and in general we found the largest increases of fMRI activity for the lowest contrast stimuli that produce smallest amounts of visual drive in cortical areas. Activity for high contrast stimuli, which produce the largest drive, were unaffected or reduced by FEF-EM.
With respect to commonly-used attention models, such as contrast and response gain (Reynolds et al., 2000; Williford and Maunsell, 2006; Li et al., 2008), we note several key distinctions between such attention studies and our own. Our subjects performed a passive fixation task as we stimulated the FEF; hence, attention was not modulated here. While FEF stimulation has been shown to produce results that physiologically resemble spatial attention effects in visual cortex (Moore et al., 2003; Armstrong et al., 2006; Armstrong and Moore, 2007), exact correspondence with an attention task has not been established (that is, the behavioral and physiological results after FEF-EM were obtained in separate experiments in these studies). Second, the comparison most often made during attention tasks is to direct attention towards and away from (typically to the contralateral visual field) the stimulus of interest. This approach holds the task difficulty constant and permits a comparison between neuronal responses to attended and ignored stimuli (Martinez-Trujillo and Treue, 2002). Here, we compared visual responses during the presence and absence of FEF stimulation, which would more closely parallel an attend-toward / no-attend design. Thus, our baseline condition at each contrast level was a visual-only response and presumably lacks the suppressive interactions between simultaneously presented stimuli (Luck et al., 1997; Kastner et al., 1998) as well as putative top-down inhibitory effects present during an attend-away condition.
Previous combined fMRI-microstimulation experiments in monkeys (Tolias et al., 2005), revealed increased activity only in relatively proximal areas mono-synaptically connected with the site of stimulation. In this and our previous study (Ekstrom et al., 2008), we observed FEF-EM dependent modulations of visually-driven activity in areas distant from the site of stimulation and connected with FEF through multiple synapses. Importantly though, we only observed EM-only activity [i.e., in the absence of visual stimuli, similar to (Tolias et al., 2005)] in areas directly connected with FEF. Moreover, there are several methodological differences between our studies and that of Tolias and colleagues, including the use of awake versus anesthetized subjects, and the use of relatively small versus large currents. Though speculative, an important factor may be that the microstimulation of presumptive feedback pathways allows the investigation of multi-synaptic effects more easily than that of feedforward pathways, as the former may be more modulatory in nature and easier to capture with fMRI (Logothetis, 2008).
The models that have been proposed to describe the effects of attention on neural activity in the visual cortex (Williford and Maunsell, 2006) provide a useful framework for describing the effects of FEF stimulation. As mentioned above, one model that has been proposed to explain the effects of attention on CRFs is the response gain model, which suggests that attention produces a proportional increase in response at all contrasts so that the effect is most pronounced at the highest contrasts. Under the conditions of our study, we did not observe evidence for a response gain model; instead, in many areas and in the overall response across visual cortex we saw a reduction in activity at high contrast. An alternative is the contrast gain model, which proposes that attention produces an increase in the apparent contrast of a stimulus, causing a leftward shift of the entire CRF (Reynolds et al., 2000). The principal characteristic of a contrast gain effect is an increase in activity for stimuli at intermediate contrast, with little effect at baseline and saturation activity levels. We do observe common features with this model, specifically at lower and intermediate contrasts. The correspondence is not perfect, however, particularly at high contrast. As we noted above, many of the previous studies that investigated the effects of attention on CRFs compared attend-toward and attend-away conditions in the presence of multiple stimuli rather than an analogue of the stimulation / no-stimulation design with a single stimulus employed here. However, we cannot exclude the alternative explanation that FEF-EM does not reproduce all the effects of selective attention on neuronal activity in the visual cortex.
Recently, two models were proposed in an attempt to unify the apparently contradictory contrast gain and activity or response gain modulation of single unit responses using divisive normalization of stimulus driven activity (Reynolds and Heeger, 2009; Lee and Maunsell, 2009). In the Reynolds-Heeger model, excitatory and suppressive stimulus drive and the spatial spread of the attentional feedback all contribute to the observed response. For example, if an attended stimulus matches the receptive field (RF) size in a given area, yet the attentional field is smaller than the stimulus, the model predicts a strong response gain. If, however, the stimulus is small relative to both the RF and attentional field, the model predicts a contrast gain effect. If one assumes that FEF-EM mimics spatial attention processes, the Reynolds-Heeger model would predict very different effects in different visual areas, depending on respective RF sizes relative to our 3 degree radius stimuli. We observed, however, very similar FEF-EM effects in all areas, irrespective of neuronal RF size (Fig. S5). Importantly, the size of the presumptive attentional field is a variable in the Reynolds-Heeger model that is difficult to assess in our experiments. Although this size was fixed for a given electrode, it may be larger than the largest RFs in any of the areas investigated (e.g., FST, MST and STP). Future experiments in which both stimulus size and attentional field size are manipulated (e.g., by using variable currents or stimulating multiple electrodes simultaneously) would allow a more direct test of the Reynolds-Heeger model.
Our results, and specifically the schematic of Figure 7, bear resemblance to previous frameworks for center-surround interactions in rat somatosensory cortex (Moore et al., 1999; Simons, 1985; Armstrong-James and Fox, 1987; Ghazanfar and Nicolelis, 1997), models for contrast-dependent interactions in area V1 (Somers et al., 1998; Schwabe et al., 2006) and the above Reynolds-Heeger model (Reynolds and Heeger, 2009). We suggest that the presumptive feedback connections activated from area FEF: (i) amplify all the input at low levels of visual drive, (ii) sharpen salient responses at high levels of activity (i.e., increase discriminability), and (iii) cause a mixture of these effects at intermediate activity levels. Such a mixture is consistent with our previous findings (Ekstrom et al., 2008), where we used chromatic stimuli (~31% luminance contrast) and observed a slight suppression of the fMRI response at the center of the representation of these stimuli and an enhancement of activity in the surrounding less well activated voxels. These findings may appear to be at odds with a previous report of FEF-EM causing the largest enhancement of V4 neuronal firing rates for optimally oriented stimuli (Moore and Armstrong, 2003). We note, however, that the use of fMRI precludes the stimulation of all neurons within a voxel with optimal stimulus orientation, so that the apparent discrepancy may be due to the combination of neuronal responses with a range of stimulus preferences.
Several predictions relating to the modulation of a stimulus representation and perception due to increased FEF activity, either through EM in monkeys or transcranial magnetic stimulation (TMS) in humans (Grosbras and Paus, 2002; Ruff et al., 2006; Silvanto et al., 2006; Taylor et al., 2007), are possible from this operating schematic (Fig. 7). First, low contrast stimuli in isolation should be perceived more readily, similar to the facilitated detection of luminance changes following FEF-EM (Moore and Fallah, 2004) and as seen in many attention experiments (Bashinski and Bacharach, 1980; Hawkins et al., 1990; Muller and Humphreys, 1991; Handy et al., 1996). A second prediction arises from the comparison of the mid and high contrast regimes. Specifically, a discrimination task using stimuli comprised of a central region and annulus, with the discrimination made in the annular compartment, should be differentially impaired or facilitated by FEF output depending on the contrast used. Parallels of this prediction can be seen in recent work combining fMRI and TMS of human FEF (Ruff et al., 2006).
Our results thus strengthen theories proposing that increased output from areas responsible for oculomotor control modulates incoming visual activity (Rizzolatti et al., 1987; Corbetta et al., 1998; Moore and Armstrong, 2003; Cavanaugh and Wurtz, 2004; Ruff et al., 2006). Moreover, this top-down modulation depends on the underlying sensory drive so that neuronal activity in visual cortex is amplified at low stimulus contrast, but sharpened at high contrast levels.
We thank H. Deng for animal training and care, G. Madan, T. van Kerkoerle, G. Bonmassar and L. Wald for technical assistance, B. Rosen for advice and support, and C. Moore and R. Vogels for valuable comments. This work received support from a Natural Sciences and Engineering Research Council of Canada Postgraduate Scholarship, Human Frontier Science Program Organization, Geneeskundige Stichting Koningin Elisabeth, Inter-University Attraction Pole 5/04, Excellence Financing/05/014, Geconcerteerde Onderzoeksacties 2005/18, Fonds Wetenschappelijk Onderzoek-Vlaanderen G.0.622.08, European Union grant FP7/2007-2013 # F2-2008-200728, NIH grant R01-EB000790 and NSF grant BCS-0745436. The Martinos Center is supported by N.C.R.R. grant P41RR14075 and the MIND Institute.