Participants were recruited from the University of Pennsylvania community to participate in one of three experiments (Experiment 1: n = 8, 4 female, age range: 21–38; Experiment 2: n = 8, 3 female, age range 20–38; Experiment 3: n = 15, 7 female, age range: 20–38). Five subjects that participated in Experiment 1 also participated in Experiment 2 (separated by around 3 months). All subjects that participated in Experiment 1 also participated in Experiment 3 during the same testing session, with Experiment 3 preceding Experiment 1. Subjects had normal or corrected-normal vision and had radiologically normal brains, without history of neuropsychological disorder. All participants provided written consent according to procedures approved by the University of Pennsylvania institutional review board.

Scanning was performed at the Hospital of the University of Pennsylvania using a 3T Siemens Trio scanner equipped with a 32-channel head coil. High-resolution T1-weighted images for anatomical localization were acquired using a three-dimensional magnetization-prepared rapid acquisition gradient echo pulse sequence [repetition time (TR), 1620 ms; echo time (TE), 3.09 ms; inversion time (TI), 950 ms; voxel size, 1 × 1 × 1 mm; matrix size, 192 × 256 × 160]. T2*-weighted images sensitive to blood oxygenation level-dependent (BOLD) contrasts were acquired using a gradient echo, echoplanar pulse sequence [TR, 3000 ms; TE, 30 ms; flip angle 90°; voxel size, 3 × 3 × 3 mm; field of view (FOV), 192; matrix size, 64 × 64 × 44]. Visual stimuli were displayed by rear-projecting them onto a Mylar screen at 1024 × 768 pixel resolution with an Epson 8100 3-LCD projector equipped with a Buhl long-throw lens. Subjects viewed stimuli through a mirror attached to the head coil.

Each experiment consisted of two 5 min 25 s fMRI scan runs in which subjects viewed stimuli from four conditions that were chosen to test specific hypotheses about PPA function (see “Stimuli” Section). Scan runs were divided into sixteen 15 s blocks; in each, subjects viewed 15 stimuli from the same condition presented one at a time for 600 ms each followed by a 400 ms interstimulus interval. Stimuli had a visual extent of approximately 13 × 13 degrees and were presented on a gray background. The experimental blocks were interspersed with five 15 s fixation blocks in which a black fixation cross was presented at the middle of a uniform gray screen. Subject attention was maintained by asking them to perform a one-back image repetition detection task during the experimental blocks. Stimulus repetitions occurred twice per block; thus, there were 52 unique stimuli per condition.

Following the experimental runs, subjects completed two functional localizer runs in which they viewed scenes, objects, faces, and scrambled objects in separate blocks. Data from these runs were used to identify the location of the PPA and other scene-selective regions. These runs had the same length, design, timing, and task as the main experimental runs.

Functional MR images for both the main experiments and functional localizer were preprocessed using the following steps. First, they were corrected for differences in slice timing by resampling slices in time to match the first slice of each volume. Second, they were corrected for subject motion by realigning to the first volume of the scan run using MCFLIRT (Jenkinson et al., 2002). Third, the timecourses for each voxel were high-pass filtered to remove low temporal frequency fluctuations in the BOLD signal that exceeded lengths of 100 s. Data from the functional localizer scan were smoothed with a 5 mm full-width at half-maximum (FWHM) Gaussian filter. Data from the experimental scans were smoothed with a 5 mm FWHM Gaussian filter for all region of interest analyses and 8 mm FWHM for all whole-brain group analyses.

We examined univariate responses within several regions of interest (ROI) known to be involved in visual processing. ROIs were defined individually for each subject using data from the functional localizer scans. In addition to the PPA, we defined ROIs for two other scene-responsive regions [retrosplenial complex (RSC) and occipital place area (OPA); Hasson et al., 2002; Bar and Aminoff, 2003; Dilks et al., 2013], and early visual cortex (EVC). The OPA, but not the RSC or EVC, has been previously reported to show a similar rectilinearity bias to PPA (Nasr et al., 2014). ROIs were defined using a contrast of scenes > objects for PPA, RSC, and OPA, and scrambled-objects > baseline for EVC, and they were further constrained by a group-based anatomical map of scene- or scrambled-object-selective activation derived from a large number (42) of localizer subjects that had been previously obtained in our lab (Julian et al., 2012). Specifically, each ROI was defined as the top 100 voxels in each hemisphere that exhibited the defining contrast and fell within the group-parcel mask for that ROI. The group-parcel mask for EVC was defined based on a scrambled-objects > intact-objects contrast. The voxels comprising each ROI did not need to be contiguous. This method ensured that all ROIs could be defined in both hemispheres in every subject and that all ROIs contained the same number of voxels. All ROIs were combined across hemispheres unless otherwise noted. All contrasts were performed in the native anatomical space for each subject and the group-parcel map was mapped into that space using a linear transformation.

We then used general linear models (GLMs) implemented in FSL¹ to estimate the response of each voxel to the four experimental conditions for each experiment. Each condition was modeled as a boxcar function convolved with a canonical hemodynamic response function. To test for the effects of independent factors, we applied repeated-measures analysis of variances (ANOVAs) to the univariate responses within each ROI. Subsequent comparisons between individual conditions were based on paired-sampled t-tests. For tests of the rectilinearity hypothesis, significance was assessed using 1-tailed tests in the direction of the rectilinearity hypothesis (i.e., greater response to high- than low-rectilinearity conditions). For all other tests, significance was assessed using 2-tailed tests.

In addition to the ROI analyses, we also performed a whole-brain group analysis to test for effects of rectilinearity and category in Experiments 2 and 3 outside of our ROIs. For this analysis, data from those subjects who participated in both Experiments 2 and 3 were first combined via a within-subject fixed-effects analysis prior to the group analysis. To generate group-averaged maps, each individual participant’s functional maps were spatially transformed onto the averaged human brain using a spherical transformation in FreeSurfer² (Fischl et al., 1999) and then averaged using random effects models in FSL.

In addition to univariate analyses, we also assessed whether there was information about rectilinearity and stimulus category represented in the multivoxel patterns of response in each ROI in Experiments 2 and 3. To do so, for each participant, we used GLMs to estimate the response pattern evoked by each stimulus condition separately for each of the two fMRI runs. Multivoxel pattern analyses (MVPA) were then performed through split-half pattern comparison (Haxby et al., 2001). Individual patterns were normalized prior to this computation by subtracting the grand mean pattern (i.e., the cocktail mean) for each half of the data. For each ROI, we then computed the correlation between the response patterns resulting from the same stimulus conditions and from different stimulus conditions. To test for coding of rectilinearity controlling for stimulus category, we computed a discrimination index that was the difference in the average correlation between the same rectilinearity condition and the corresponding different rectilinearity condition (i.e., [same rectilinearity, same category] − [different rectilinearity, same category]). This rectilinearity discrimination index was computed separately for scenes and faces. Likewise, to test for coding of stimulus category controlling for rectilinearity we computed a discrimination index that was the difference in the average correlation between the same category condition and the corresponding different category condition (i.e., [same category, same rectilinearity] − [different category, same rectilinearity]). This category discrimination index was computed separately for high and low rectilinearity stimulus conditions. To assess statistical significance, t-tests were used evaluate if the discrimination indices were greater than zero.

In our first experiment, we sought to replicate the PPA rectilinearity bias for simple shapes reported by Nasr et al. (2014). To do so, we scanned participants while they viewed arrays of computer-generated gray-scale squares (high-rectilinearity) and circles (low-rectilinearity) presented at two different sizes (small or large; Figure 1A). We then examined the fMRI response in each of the predefined ROIs.

Consistent with the rectilinearity hypothesis, the PPA responded more strongly to arrays of squares than to arrays of circles (Figure 2). Confirming this observation, a 2 × 2 ANOVA with factors for shape (square vs. circle) and size (small vs. large) found a main effect of shape (F_(1,7) = 7.48, p < 0.05, ηp2 = 0.52). The greater PPA response to squares than circles was significant for large shapes (t₍₇₎ = 2.64, p < 0.05) and marginally significant for small shapes (t₍₇₎ = 1.72, p = 0.07). Thus, our results replicate the basic finding of Nasr and colleagues that the PPA’s response to arrays of 2-d shapes is modulated by rectilinearity.

We also observed a significant effect of rectilinearity, with greater response to arrays of squares than to arrays of circles, in OPA (F_(1,7) = 10.37, p < 0.01, ηp2 = 0.60). In contrast, there was no rectilinearity bias in RSC (F_(1,7) = 0.09, p = 0.77) or EVC (F_(1,7) = 2.94, p = 0.13). The fact that EVC responds equally to squares and circles suggests that the rectilinearity bias observed in PPA and OPA was not simply inherited from this region. The OPA showed a significantly greater rectilinearity effect than EVC (F_(1,7) = 8.64, p < 0.05, ηp2 = 0.55) but the region-by-rectilinearity interaction between PPA and EVC fell short of significance (F_(1,7) = 3.89, p = 0.089).

Unexpectedly, we also observed size effects in the PPA, OPA, and RSC. All three scene regions responded significantly more to large than small shapes (PPA: F_(1,7) = 41.91, p < 0.01, ηp2 = 0.86; RSC: F_(1,7) = 14.35, p < 0.01, ηp2 = 0.59; OPA: F_(1,7) = 10.01, p < 0.01, ηp2 = 0.67). The reason for this preference for large shapes is unclear. It may indicate a preference for larger objects (Konkle and Oliva, 2012), or it might be driven by uncontrolled variables such as spatial frequency or numerosity. Notably, EVC showed the opposite effect, responding more to small than large shapes (F_(1,7) = 72.97, p < 0.001, ηp2 = 0.91). In no ROI was there was significant interaction between shape and size (PPA: F_(1,7) = 0.51, p = 0.50; RSC: F_(1,7) = 0.01, p = 0.92; OPA: F_(1,7) = 0.15, p = 0.71; EVC: F_(1,7) = 2.36, p = 0.17). Further, an additional 2 × 2 × 2 analysis with hemisphere as a factor found that the shape and size effects did not vary by hemisphere in any ROI (all F_(1,7)s < 2.59, ps > 0.15).

After replicating the rectilinearity bias for shapes in the PPA, we next moved on to test whether there is a rectilinearity effect for naturalistic images, by scanning participants while they viewed images of high- and low-rectilinearity scenes and high- and low-rectilinearity faces (Figure 1B). Importantly, rectilinearity was matched between the scenes and faces; that is, the high-rectilinearity scenes and faces had a similar level of rectilinearity, as did the low-rectilinearity scenes and faces. Thus, our design not only allowed us to examine rectilinearity effects for scenes and faces, it also provided a strong test of the rectilinearity hypothesis. If the preferential response to scenes compared to faces in the PPA is due to the greater rectilinearity of scenes, then the “categorical” effect should be eliminated by rectilinear matching. On the other hand, if the PPA responds strongly to scenes in part because it is tuned to scenes as a category, then it should continue to exhibit a preferential response to scenes even after rectilinear matching.

Results are plotted in Figure 3A. A 2 × 2 ANOVA with factors for category (scene vs. face) and rectilinearity (high vs. low) found a strong effect of category in the PPA (F_(1,7) = 153.65, p < 0.001, ηp2 = 0.96), with greater response to scenes than to faces. Crucially, there was no main effect of rectilinearity (F_(1,7) = 0.49, p = 0.51). There was a significant interaction between category and rectilinearity (F_(1,7) = 11.26, p < 0.05, ηp2 = 0.62); however, this interaction was driven by a lower response to high- than low-rectilinearity scenes (t₍₇₎ = −3.77, p = 0.99) and the numerically converse effect for faces, though the rectilinearity effect for faces was not significant (t₍₇₎ = 1.52, p = 0.17).

Results in the other two scene regions were similar to the PPA: both RSC and OPA responded significantly more to scenes than faces (RSC: F_(1,7) = 31.10, p < 0.01, ηp2 = 0.82; OPA: F_(1,7) = 46.09, p < 0.001, ηp2 = 0.87) but neither region exhibited a main effect of rectilinearity (RSC: F_(1,7) = 0.26, p = 0.63; OPA: F_(1,7) = 0, p = 0.99). Like PPA, OPA also showed a significant interaction between category and rectilinearity (F_(1,7) = 15.45, p < 0.01, ηp2 = 0.69), with a lower response to high- than low-rectilinearity scenes (t₍₇₎ = −3.10, p = 0.99) and a marginal rectilinearity effect for faces (t₍₇₎ = 1.65, p = 0.06). There was no interaction between category and rectilinearity in RSC (F_(1,7) = 2.08, p = 0.19). Comparison of the three scene regions revealed no interaction between region and rectilinearity (F_(2,14) = 0.05, p = 0.95), but an interaction between region and category (F_(2,14) = 15.99, p < 0.001, ηp2 = 0.70; scene-face response difference: OPA > PPA > RSC). Additional 2 × 2 × 2 ANOVAs with hemisphere as a factor found no significant interaction between hemisphere and rectilinearity in the PPA or RSC (both F_(1,7)s < 2.5, ps > 0.16). In the OPA, there was a significant interaction between hemisphere and rectilinearity (F_(1,7) = 5.69, p < 0.05, ηp2 = 0.45), with a numerically greater response to high- than low-rectilinearity in the left hemisphere, and the converse effect in the right hemisphere, although neither hemisphere exhibited a significant effect of rectilinearity (both t₍₇₎s < 0.44, ps > 0.3). There was no interaction between category and hemisphere in any scene region (all F_(1,7)s < 2.61, ps > 0.15).

Like the scene regions, EVC responded similarly to high and low rectilinearity stimuli (Figure 3A; F_(1,7) = 0.85, p = 0.39), and there was also no significant interaction between category and rectilinearity (F_(1,7) = 0.98, p = 0.36) or hemisphere and rectilinearity (F_(1,7) = 0.66, p = 0.44). Moreover, there was no significant region-by-rectilinearity interaction between EVC and any of the scene regions (all F_(1,7)s < 2.52, ps > 0.18). However, EVC responded more to scenes than faces (F_(1,7) = 23.33, p < 0.01, ηp2 = 0.77), and there was a significant interaction between hemisphere and category (F_(1,7) = 11.29, p < 0.05, ηp2 = 0.62). Although PPA and OPA showed a greater scene-preferential response than EVC (all F_(1,7)s > 39.95, ps < 0.001, ηp2s > 0.85), the presence of category effects in EVC nonetheless indicates that there are some low-level differences between the scene and face categories despite our efforts to control for rectilinearity, overall visual extent, contrast, and luminance. There was no region-by-category interaction between EVC and RSC (F_(1,7) = 0.44, p = 0.53).

We considered the possibility that a subregion in the vicinity of the PPA, but not the most scene-selective part of the region, might be selective for the presence of right angles. Such a subregion might not be included in the PPA as defined by the top 100 voxels showing scene-selectivity in each hemisphere in the functional localizers. To address this possibility, we compared rectilinearity-selectivity (high vs. low rectilinearity contrast t-statistic) and category-selectivity (scene vs. face category contrast t-statistic) for each participant for each voxel in both hemispheres in the group-defined PPA parcel, which is larger than the individually-defined ROIs (Julian et al., 2012). Figure 3B shows the results of this comparison. Only a small fraction of voxels was more rectilinearity-selective than scene selective. Further, in each participant the number of category-selective voxels far exceeded the number of rectilinearity-selective voxels (Figure 3B). Thus, it is unlikely that the failure to find rectilinearity-selectivity in the PPA was due to a bias in ROI definition induced by our analysis methods.

We also considered the possibility that the PPA might be sensitive to rectilinearity at the representational level. That is, even though the overall level of activity in PPA did not distinguish between high vs. low rectilinearity scenes and faces, such distinctions might be apparent in multi-voxel response patterns. To test this, we performed split-half MVPA to see if it was possible to distinguish between stimuli based on rectilinearity (Figure 4A). We did not find strong evidence for this: in PPA and OPA, scenes with the same level of rectilinearity were no more representationally similar than scenes with different levels of rectilinearity (both t₍₇₎s < 1.20, ps > 0.26), and an absence of rectilinearity information was also observed for faces (both t₍₇₎s < −0.03, ps > 0.51). In RSC and EVC, there was significant information about rectilinearity for scenes (both t₍₇₎s > 2.41, ps < 0.05), but not faces (both t₍₇₎s < −0.82, ps > 0.78). In contrast, both high- and low-rectilinearity stimuli were more representationally similar if they were drawn from same category than if they were drawn from different categories in all ROIs (all t₍₇₎s > 5.83, ps < 0.001; Figure 4B). These results reinforce the idea that scene regions are driven more by category than by rectilinearity.

In sum, our data in this case did not support the rectilinearity hypothesis. Not only was the categorical effect maintained after rectilinear matching, but no rectilinearity effect was observed for scenes and faces. Moreover, MVPA could not distinguish between stimuli that differed only in rectilinearity independent of category.

One possible reason for the lack of a rectilinearity effect in Experiment 2 may have been that difference between the high- and low-rectilinearity conditions was too subtle to be noticed by participants. Although the high- and low-rectilinearity conditions in Experiment 2 differed on rectilinearity according to the rectilinearity index designed by Nasr et al. (2014), this index may fail to capture the most perceptually salient rectilinearity dimensions. Further, the greater response to scenes than faces in EVC in Experiment 2 emphasizes that there were other uncontrolled low-level differences between the categories, complicating the interpretation of the category results. To address these concerns, participants in Experiment 3 viewed images of scenes and faces with artificially enhanced or degraded rectilinearity (Figure 1C). Images were decomposed into square pixels (pixelated) to increase rectilinearity or round points (pointillized) to decrease rectilinearity. We then examined the fMRI response in each of the predefined ROIs.

Once again, we failed to find an effect of rectilinearity on PPA response (Figure 3C). A 2 × 2 ANOVA with factors for category (scene vs. face) and rectilinearity (pixelated vs. pointillized) found greater response in the PPA to scenes compared to faces (F_(1,14) = 100.19, p < 0.0001, ηp2 = 0.88) but no difference between pixelated and pointillized stimuli (F_(1,14) = 0.93, p = 0.35). This lack of a rectilinearity bias was found for both scenes (t₍₁₄₎ = 1.20, p = 0.25) and faces (t₍₁₄₎ = 0.20, p = 0.84). There was no interaction between category and rectilinearity (F_(1,14) = 0.46, p = 0.51). Further, comparing rectilinearity selectivity (pixelated vs. pointillized contrast t-statistic) and category selectivity (scene vs. face category contrast t-statistic) for each voxel in both hemispheres of the group-defined PPA parcel (Julian et al., 2012), there were few PPA voxels that were more selective for right angles than for scenes, and each participant exhibited substantially more voxels selective for scenes than right angles (Figure 3D). Thus, the failure to find rectilinearity-selectivity in the PPA in the present experiment was not due to a bias induced by our method of defining the ROI.

Results in the other scene regions were similar. Both RSC and OPA responded more to scenes than faces (RSC: F_(1,14) = 96.60, p < 0.001, ηp2 = 0.87; OPA: F_(1,14) = 90.15, p < 0.0001, ηp2 = 0.87), but neither region showed a significant main effect of rectilinearity (RSC: F_(1,14) = 1.02, p = 0.33; OPA: F_(1,14) = 2.96, p = 0.11), although the nonsignificant trend in OPA was in the predicted direction. There was no interaction between category and rectilinearity (RSC: F_(1,14) = 0.69, p = 0.42; OPA: F_(1,14) = 1.79, p = 0.20). Comparison of the three scene regions revealed no interaction between region and rectilinearity (F_(2,28) = 2.26, p = 0.12), but an interaction between region and category (F_(2,28) = 45.19, p < 0.001, ηp2 = 0.78; scene-face response difference: OPA > PPA > RSC). EVC did not show effects of category (F_(1,14) = 1.39, p = 0.26) or rectilinearity (F_(1,14) = 0.38, p = 0.55), and no interaction between category and rectilinearity (F_(1,14) = 2.61, p = 0.13), indicating the category effects in scene regions could not have been inherited from EVC. There was no significant region-by-rectilinearity interaction between EVC and any of the scene regions (all F_(1,14)s < 1.15, ps > 0.30), but all scene regions showed a greater scene-preferential response than EVC (all F_(1,14)s > 12.50, ps < 0.004, ηp2s > 0.47). 2 × 2 × 2 ANOVAs with hemisphere as a factor found no significant interactions between hemisphere and rectilinearity or category in any ROI (all F_(1,14)s < 3.98, ps > 0.07).

To test whether the scene regions distinguished between pixilated and pointillized stimuli at the level of multivoxel patterns, we again performed split-half MVPA (Figure 4C). There was no significant information about rectilinearity in the RSC or OPA for either scenes or faces (both t₍₁₄₎s < 1.22, ps > 0.12). The PPA showed marginal information about rectilinearity for faces (t₍₁₄₎ = 1.59, p = 0.07, but not scenes (t₍₁₄₎ = 0.72, p = 0.24). All ROIs contained significant information about category for both pixelated and pointilized stimuli (all t₍₁₄₎s > 6.46, ps < 0.0001). Notably there was significant information about rectilinearity in EVC for scenes (t₍₁₄₎ = 3.28, p < 0.01) and marginal information for faces (t₍₁₄₎ = 1.52, p = 0.075), indicating that this region was sensitive to the difference between the pixilated and pointillized stimulus conditions.

To test for effects of rectilinearity and category outside of our ROIs, we performed a whole-brain group analysis, aggregating data from Experiments 2 and 3 to maximize power to detect effects of rectilinearity and category. This analysis revealed very strong category effects throughout high-level visual cortex (Figure 5): the PPA, RSC, and OPA responded more strongly to scenes, whereas lateral and ventral occipitotemporal regions responded more strongly to faces. By contrast, we observed no rectilinearity effects that survived correction for multiple comparisons, although notably there was sensitivity to rectilinearity observed near the posterior right PPA and the left OPA at uncorrected statistical thresholds.