The “Parahippocampal Place Area” Responds Preferentially to High Spatial Frequencies in Humans and Monkeys

1There is lively interest in defining “category-specific” areas in the object processing regions of inferior temporal (IT) cortex, in both humans and monkeys. In such higher-order areas of the ventral visual pathway, discrete clusters of neurons reportedly respond selectively to specific categories of complex images such as faces –, places,, body parts,, tools, animals, word forms, and perhaps even chairs. However, in the campaign to map the category-specific regions of visual cortex, it is sometimes overlooked that stimuli of a common category often also share lower-level visual cues. Correspondingly, many cells in IT cortex are selective for specific lower-level properties, including surface curvature,, Fourier descriptor shapes, simple geometry,, non-accidental features (geons), diagnostic features, color, and/or retinotopic location. Thus, a hypothetical group of IT cells that responds robustly to a specific lower-level visual feature may also be interpreted as having a category-selective response—to the extent that these lower-level features are common to images from that category. For instance, functional magnetic resonance imaging (fMRI) studies have shown large-scale shape, and eccentricity, maps in human and macaque temporal cortex, which coincide with category-selective areas (e.g., face-selective patches). In fact, a recent theory suggests that overlapping continuous maps of simple features give rise to discrete modules that are selective for complex stimuli.

2Here we demonstrate such a coincidence between a lower-level feature-selective map and category selectivity, in a prominent region of IT cortex known as the parahippocampal place area (PPA). Previously it has been reported that PPA responds more to images of “places” (or “scenes”) than to images from certain other object categories,. However, place images encompass a virtually infinite range of possible visual stimuli: how could such a wide range of stimuli be coded in visual cortex? Similarly, how can “placeness” be quantified, to enable experimental study? Here we found that a simpler, lower-level feature selectivity contributes strongly to the higher-level place-selective response in PPA. This feature selectivity enhances the visual borders and edges of surrounding landmarks, which is useful in brain processing of places, scenes, and higher-order processes such as navigation.

3The first sign of this lower-level selectivity arose serendipitously, when we were testing fMRI responses to very simple geometrical shapes including cubes and spheres. Figure 1 shows the cortical response to a cube relative to a sphere, when both stimuli were closely matched along lower-level dimensions other than shape. Surprisingly, we found that the cube activated PPA robustly and selectively—unlike the response in other visual areas (e.g., the adjacent, face-selective fusiform face area [FFA]).

4This result has several implications. First, it revealed a vigorous and selective response to one computer-generated shape relative to another in PPA—though neither shape was a “place.” Second, this activity distinction occurred despite numerous similarities of the cube and the sphere: many other lower-level cues (e.g., surface reflectance, light source location, 3-D volume, and the visual field position of the shape center) were equated across the two stimuli.

5What could explain this unexpected result? Among the remaining properties that distinguished these two shapes, the cube had distinct spatial discontinuities (edges and 3-D corners) in the foreground, whereas the sphere did not. These feature differences were reflected as a difference in fast Fourier transform (FFT), a quantitative global measurement of image properties: a complex pattern of high spatial frequency (SF) components was present in the cube, but not in the sphere (see Figure 1A). From this, it appears that PPA responses are modulated by spatial discontinuities in some form, perhaps reflected in the form of higher SFs.

6To further explore the effect of spatial discontinuities, we added either a white-noise texture (“textured”) or surface irregularities (“bumpy”) to the cube and sphere, and measured the fMRI responses independently to these stimuli, in addition to the original “smooth” sphere and cube, in common scan sessions. Again, we found that the responses in PPA were always higher for cubes than spheres, irrespective of the surface properties of these two shapes (Figure 2).

7Additional experiments revealed that even simpler, 2-D stimuli (classical flickering checkerboards) could selectively activate PPA, when those stimuli included high spatial frequencies. Cortical maps (Figures 3, S1, and S2) and region-of-interest analysis (Figure S3) showed a robust bias for high spatial frequencies in PPA, in response to checkerboard patterns—though such stimuli were neither places, nor objects, nor even 3-D shapes. The PPA response to high frequency checkerboards was even higher than that to normal, unfiltered checkerboards (see Figure S2)—suggesting that the presence of lower spatial frequencies in the normal image can effectively reduce the PPA response to high spatial frequencies, perhaps due to nonlinear inhibitory interactions (see Discussion).

8These results raise the question of whether at least part of the PPA response to “places” in previous reports may reflect the effect of uncontrolled spatial factors, including high spatial frequencies. To re-test this idea in the originally reported context, we first presented a range of naturalistic place and face images, because such place versus face comparisons have been used as a basic “localizer” for PPA (e.g., ). Intermixed with these normal stimuli, we presented the same set of images, after spatially filtering those images for low, middle, or high SFs. The initial SF-filtered images were generated based on arbitrarily chosen cut-off frequencies of 1 and 5 cycles/degree (c/deg) for low-pass and high-pass filtering, respectively (Figure 4). Since normal face and place images had different SF distributions (Figure S4A), the cut-off frequencies were shifted for place images, so that the resultant SF-filtered places had the same fraction of total power as the SF-filtered faces (see Figure S4 for further details). Outside the scanner, the effect of SF filtering on image perception was behaviorally assessed in each subject. All subjects could easily distinguish all SF-filtered images as either faces or places.

9Again, we found that PPA was especially responsive to “high SF” stimuli, compared to “middle SF” and “low SF” stimuli. This preferential response to high SFs was present for both face (Figures 5 and S5) and place (Figures 6 and S6) stimuli. In fact, the PPA response to faces (“nonoptimal” stimuli in PPA) almost doubled relative to that for normal images, when low and middle SF components were removed (see Figure S5). Maps of this high SF bias even reflected the same idiosyncratic topography as classically defined PPA (localized using normal, naturalistic images of places versus faces) in each subject (see Figure 5)—thus emphasizing the underlying biological link between PPA and its high SF bias. Interestingly, the high SF face images were as effective as the place images in activating PPA (Figure S7).

10As an expected control result in the lower-tier (occipital) visual cortex (e.g., in areas V1, V2, and V3), the SF sensitivity co-varied systematically with the retinotopic representation of visual field eccentricity; a preference for higher SFs was clearly shown nearer the foveal cortex (Figure S8; see also,). The foveal representation in area V8/VO, may also have a preference for high SFs; however, our maps confirmed that the high SF activity in PPA was located anterior and ventral to V8/VO (Figure S9).

11In the experiments above, the SF content of the images was manipulated by frequency filtering. However, natural scenes can also vary dramatically in their SF distributions. Does PPA activity reflect this spectral variation in normal (unfiltered) natural scenes? To answer this, we conducted an additional experiment comparing the fMRI response to normal scenes that were relatively dominated by either high or low SFs (Figures 7 and S10). Although both types of images were from the same semantic category (scenes), PPA (as defined by a scene/object localizer) responded significantly more to the scenes that had more power at high SFs (see Figure 7). This result indicates that intrinsic SF components in natural scenes can strongly modulate the PPA activity, even during natural scene viewing per se.

12If this spectral sensitivity reflects a fundamental mechanism for neural scene coding in PPA, presumably it should also exist in a region of macaque visual cortex that corresponds to human PPA. To our knowledge, no such mechanism of high SF sensitivity has been reported previously in monkey. In fact, a monkey homolog of the human “place” area (PPA) has not yet been demonstrated. To test for a homolog of PPA in monkeys and for high SF sensitivity in this homolog area, we presented the same set of normal and SF-filtered stimuli to awake fixating monkeys, using fMRI procedures very similar to those used in the human scanning (see Materials and Methods).

13First, using the original set of normal, naturalistic stimuli, we demonstrated a place-selective region in macaque IT cortex (Figure 8) that was located in a consistent location, immediately ventral to the largest IT face patch (Figures 9B and S11). Thus, this place-selective region in macaques is located exactly where it is predicted to lie, based on (1) the maps of face- and place-selective areas (FFA and PPA) in humans and (2) the previously reported location of the monkey homolog of FFA (mFFA) in macaques,,. This monkey homolog of PPA (mPPA) was confirmed in an averaged map from three monkeys tested with additional control stimuli (Figure S12). All this evidence demonstrates that macaque monkeys do have a homolog of cortical area PPA, as defined originally in humans.

14Importantly, subsequent tests showed that the mPPA was strongly selective for high SFs—again like PPA in humans (Figures 9C–9F and S13). In fact, the SF bias here was strong enough to reveal a striking double dissociation: relative fMRI activation in mPPA/mFFA could be effectively reversed by changing either object category (place versus face) or SF (high versus low) (see Figure 9). This finding also reemphasizes the lower-level nature of the trigger features in this “place processing” region of IT cortex: it is unlikely that monkeys have the same higher-order place associations as humans, in response to the same naturalistic images.

15Why would this high SF bias exist in PPA? From a “top-down” perspective (e.g., a category-based model), this kind of information might be especially useful in defining “places” because high SFs, which are emphasized in edges, borders, and small spatial details, are crucial for survival in primate arboreal navigation (in reaching and grasping during brachiation [Figure 10], avoiding obstacles, etc.) or when detecting predators in visually complex surroundings. By this idea, a sensitivity to spatial discontinuities (as reflected in high SFs) may have been enhanced in PPA during evolution, in the service of spatial perception and navigation. This view integrates the data here with the “place-selective” model of PPA function.

16However, our data also raise an alternative, “bottom-up” possibility (e.g., a lower-order interpretation). If a sensitivity to spatial discontinuities (e.g., a high SF bias) explains some of the PPA response, can an improved understanding of this lower-level spatial selectivity eventually explain all of that response? Our evidence clearly indicates a spatial selectivity in PPA, but this initial spatial characterization based on linear SF analysis is likely imperfect. Even in primary visual cortex, many cells show nonlinear inhibitory interactions in response to two SFs,. Most commonly, such single-unit interactions consisted of lower SFs inhibiting higher frequencies. Our fMRI data showed analogous interactions in PPA (e.g., Figure S2). Thus, a more refined spatial characterization (beyond simple SF filtering) may ultimately explain even more of the variance in the PPA response. For instance, measurements of SF per se did not capture all of the variation in the PPA response in the cubes-and-spheres experiment. Ultimately, it may turn out that PPA responds selectively to a specific type of spatial discontinuity (e.g., corners, or Y-junctions), but not to others (e.g., linear edges)—though all these features add power at higher SFs in the FFTs from the entire image.

17How can the evidence for a high SF bias be reconciled with the previously reported response to conventional place/scene images in PPA? In fact, there may be little or no discrepancy. As one might suspect, an FFT analysis (Figure S14) confirms that conventional (unfiltered) place images (e.g., images of buildings) have substantially more energy at high SFs than images of faces. Thus, both “high SF” and “place” hypotheses predict a robust activation in PPA in response to conventional place images.

18According to previous reports (e.g., ), “intact” (normal) place images evoke a larger response in PPA than “scrambled” object images. In fact, this comparison is sometimes used as a conventional localizer for PPA. Intuitively, one might expect that the scrambled images would have more high SF power, because additional edges are added to the image during scrambling. Does this finding violate our conclusion that PPA responds best to higher SFs? To resolve this, we measured the normalized high SF power (the high SF power divided by the total power) for place and scrambled images used recently by others. Counterintuitively, we found that the place images had more high SF power than the scrambled images (p<0.00001)—presumably reflecting the lower spatial density in the latter.

19Previous human fMRI studies by Malach and colleagues, showed that PPA responds better to peripheral stimuli than to foveal stimuli, relative to the neighboring FFA. It is also known that psychophysical sensitivity to higher SFs decreases at increasing eccentricities,. Thus, our finding of a high SF bias in PPA appears to conflict with the peripheral bias in PPA. However, this apparent contradiction does not challenge our present results when all details are considered. In the experiments here, the full range of SFs was readily visible, at all eccentricities stimulated. In the fovea, human observers can resolve SFs up to ∼50 c/deg; this is a log unit higher than the maximum cut-off frequency used in our experiments (i.e., >5 c/deg). Even at the peripheral edge of our stimuli (10° eccentricity), human observers can resolve SFs more than twice as high as the 5-c/deg cut-off frequency. Thus, even the “high” SFs presented here were all “low” enough to be clearly visible (and thus accessible to PPA), even in the periphery of our stimuli. Nonetheless, given the previous reports of retinotopic biases in PPA,, it is conceivable that the strength of high SF bias in PPA varies with corresponding variations in retinotopic location and size of the SF-filtered stimuli.

20In some (but not all) maps (e.g., Figures 3, 6, and 9), there was a partial overlap between place selectivity and high SF selectivity in PPA; the high SF effect was stronger in the posterior-lateral part of PPA. It is possible that PPA contains two functionally distinct sub-regions: (1) a “posterior-lateral” sub-region more influenced by low-level features and (2) an “anterior-medial” sub-region more influenced by cognitive factors. A similar subdivision of parahippocampal cortex has been reported previously,.

21To resolve whether PPA responds to places versus lower-level spatial parameters such as SF, it would be ideal to quantitatively measure the relative strength of these two parameters in PPA. We did this test in monkeys (Figure 9) and in humans (Figure S7), and we obtained the qualitatively expected result. However, ultimately, it is not clear how to resolve this question quantitatively. How does one quantify “placeness”? What common metric can be used for both dimensions (i.e., what SF distribution equals how much placeness)? How does one deal with the known variation in acuity (and SF sensitivity) with retinal eccentricity? Further research is required to resolve these issues. Irrespective of the outcome, it should be simpler to interpret the functional processing in this intriguing higher-tier visual area, now that a robust and quantifiable lower-level stimulus selectivity has been identified.

22Following the demonstration of face-selective regions in human visual cortex (e.g., ), it was shown that homologous face-selective regions also exist in visual cortex of macaque monkeys,. This demonstration of homologous cortical regions across ∼50 million years of divergent primate evolution suggested that such neural mechanisms are fundamental to face processing, rather than some peculiarity of the human brain. More practically, the demonstration of face patches in monkeys made it possible to test the neural connections and single-neuron response properties, in the face patches, in ways that are impossible in human subjects. Here the demonstration of homologous “place areas” enables similar experimental opportunities for understanding the neural mechanisms underlying this intriguing system.

23Seven human subjects (with normal or corrected-to-normal vision) and three juvenile (5–7 kg) male rhesus monkeys (Macaca mulatta) were tested, in several experimental sessions each. Written informed consent was obtained from all human subjects prior to the scanning session. Surgical details and the training procedure for monkeys have been described elsewhere,. All experimental procedures conformed to US National Institutes of Health guidelines and were approved by Massachusetts General Hospital protocols (#2000P-001155 for the human subjects, and #2005N-000201 for the monkeys). This research was conducted in accordance with the Weatherall report.

24All subjects were scanned in a horizontal 3T scanner (Siemens Tim Trio). Gradient echo EPI sequences were used for functional imaging (human: TR = 2,000 ms, TE = 30 ms, flip angle = 90°, 3.0-mm isotropic voxels, and 33 axial slices; monkey: TR = 2,000 ms, TE = 19 ms, flip angle = 90°, 1.0-mm isotropic voxels, and 50 axial slices). To increase functional magnetic resonance sensitivity in the monkey scans (to compensate for smaller voxels), we used a gradient insert coil (Siemens AC88), parallel imaging with a four-channel phased array coil, and an exogenous contrast agent (MION; 8–10 mg/kg IV). Throughout the functional scans, all subjects continuously fixated a small fixation spot at the center of the visual display, and eye position was monitored in monkeys using an infrared pupil tracking system (Iscan). A 3-D MP-RAGE sequence (1.0-mm isotropic voxels in humans; 0.35-mm isotropic voxels in anesthetized monkeys) was also used for high-resolution anatomical imaging from the same subjects. Functional and anatomical data were preprocessed and analyzed using FreeSurfer and FS-FAST (http://surfer.nmr.mgh.harvard.edu/).

25For each subject, the inflated cortex and flattened cortical patches were reconstructed from MRI-based anatomical images. All functional images were motion corrected, spatially smoothed (unless otherwise noted) using a 3-D Gaussian kernel (2.5-mm half width at half maximum in humans and 1-mm half width at half maximum in monkeys), and normalized across scans. For intensity normalization, the BOLD image was first skull-stripped using the FSL (http://www.fmrib.ox.ac.uk/fsl/) Brain Extraction Tool to create a mask of brain-only voxels. Within this mask, the mean intensity of all voxels across all time points was computed. The value at each voxel at each time point was then scaled by 1,000 divided by the mean intensity, effectively forcing the mean in the mask to be 1,000.

26The estimated hemodynamic response was defined by a γ function, and then the average signal intensity maps were calculated for each condition. Voxel-wise statistical tests were conducted by computing contrasts, based on a univariate general linear model. To avoid sampling gaps in the (high-resolution) monkey fMRI, we sampled and averaged the functional activity from all voxels located within the gray matter, along the surface normal. Finally, the significance levels were projected onto the inflated/flattened cortex after a rigid co-registration of functional and anatomical volumes. Functional maps were spatially normalized across sessions (in monkeys) and across subjects (in humans) using a spherical transformation, then averaged using a fixed-effects model.

27Stimuli were generated on a PC (Windows XP) and presented via LCD projector (Sharp XG-P25, 1,024×768 pixels resolution, 60-Hz refresh rate) onto a rear-projection screen. The projector luminance was gamma-corrected using a hardware procedure. The gamma setting of the graphics card was set to the point closest to linearity. Additionally, the projector settings were adjusted to fine-tune the linear luminance transfer. The background luminance was set at 268 cd/m2. Matlab 7.0 and Psychophysics Toolbox (http://psychtoolbox.org/) were used to program the experiments.

28The cube/sphere stimuli (∼12° in size), including “smooth,” “textured,” and “bumpy” versions, were all equated for surface reflectance, illumination, 3-D volume, and center location, and were presented on a random dot background.

29The face/place stimuli (20° in size) were based on 20 grayscale face mosaic (“group photograph”) and place (indoor scene) images, which were qualitatively matched in visual complexity and natural image statistics (e.g., quantity of objects and degree of clutter). The initial face/place images (and also the black-white checkerboard images, which were 25° in size) were then spatially filtered (using a Hamming-windowed, causal linear-phase FIR filter with symmetric impulse response), and presented on a uniform gray background. The SF-filtered stimulus set included the following: low-pass (<1 c/deg for face; <1 c/deg or <0.41 c/deg for place; <0.5 c/deg for checkerboard), band-pass (1–5 c/deg for face/place; 0.5–5 c/deg for checkerboard), high-pass (>5 c/deg for face; >5 c/deg or >3.45 c/deg for place; >5 c/deg for checkerboard), and normal (unfiltered) images. Table 1 shows the root mean square contrast of these stimuli.

30The scene stimuli (20° in size, grayscale) were selected from a large scene image database (see Figure S10 for details).

31All stimuli were presented in a blocked design. Within a functional run, the first and last blocks were null (fixation only) epochs, and the remaining stimulus blocks were ordered pseudo-randomly. In the cube/sphere experiments, the stimulus (e.g., a smooth cube) was presented unchanged throughout a given block (“single stimulus imaging”), except for the location of a virtual light source, which was systematically varied every second to refresh the fMRI activity (four source locations, ranging within ±32° in azimuth and ±25° in elevation). In the checkerboard experiments, the stimulus (e.g., high SF checks) was presented in different phases, within a fixed spatial envelope (the phase of the stimulus was systematically varied every 500 ms). In the face/place and scene experiments, multiple examples/images of a particular stimulus condition (e.g., high SF faces) were randomly presented within a given block, with each image presented for 1 s (in the face/place experiment) or 2 s (in the scene experiment).

32To control for variations in attention, we had the human subjects detect/report a briefly presented red dot (0.2°×0.2°), which could appear anywhere on the cube/sphere shapes (Figures 1 and 2) and H/L scenes (Figure 7) during scanning at unpredictable times. Thus, to achieve a stable performance, subjects had to covertly attend to the full visuotopic extent of each stimulus. The detectability of the dot converged to 75% correct, based on a staircase modulation of dot luminance relative to the local background luminance (roughly, dot saturation). The subjects' performance was thus equated across different stimulus conditions. In the monkey experiments, attention was controlled by the fixation task, using juice reward.

33Each human scan session consisted of 12 functional runs, with each run containing 18 blocks (block duration = 16 s). Each monkey scan session consisted of 20 functional runs, with each run containing 10 blocks (block duration = 24 s). In total, we collected 66,223 functional volumes (43 sessions) in humans, and 35,840 functional volumes (14 sessions) in monkeys.