Patterns of neural response in scene-selective regions of the human brain are affected by low-level manipulations of spatial frequency

Highlights

• Distinct patterns of neural response to different categories of scenes

• Importance of image properties in generating these patterns not understood

• Low-level spatial filtering has a significant effect on patterns of response.

• Scene-selective regions are sensitive to low-level properties of the image.

Abstract

Neuroimaging studies have found distinct patterns of response to different categories of scenes. However, the relative importance of low-level image properties in generating these response patterns is not fully understood. To address this issue, we directly manipulated the low level properties of scenes in a way that preserved the ability to perceive the category. We then measured the effect of these manipulations on category-selective patterns of fMRI response in the PPA, RSC and OPA. In Experiment 1, a horizontal-pass or vertical-pass orientation filter was applied to images of indoor and natural scenes. The image filter did not have a large effect on the patterns of response. For example, vertical- and horizontal-pass filtered indoor images generated similar patterns of response. Similarly, vertical- and horizontal-pass filtered natural scenes generated similar patterns of response. In Experiment 2, low-pass or high-pass spatial frequency filters were applied to the images. We found that image filter had a marked effect on the patterns of response in scene-selective regions. For example, low-pass indoor images generated similar patterns of response to low-pass natural images. The effect of filter varied across different scene-selective regions, suggesting differences in the way that scenes are represented in these regions. These results indicate that patterns of response in scene-selective regions are sensitive to the low-level properties of the image, particularly the spatial frequency content.

Introduction

Despite their spatial complexity and heterogeneity, human observers are able to reliably categorise real world scenes even when images are presented rapidly (Greene and Oliva, 2009, Potter, 1975) or visually degraded (Torralba, 2009, Walther et al., 2011). This capacity is thought to be based on neural activity in regions of human visual cortex that are selectively responsive to visual scenes (Aguirre and D'Esposito, 1997, Dilks et al., 2013, Epstein and Kanwisher, 1998, Maguire, 2001, Nasr et al., 2011). Whilst studies using univariate fMRI analyses have reported comparable levels of response within these regions to different images of scenes (Epstein and Kanwisher, 1998), more recent reports employing multivariate techniques have shown that there are distinct patterns of response to different categories of scene (Walther et al., 2009, Walther et al., 2011) suggesting a finer-grained organisation that might underpin perceptual discriminations. However, the functional dimensions that shape these patterns have not been fully resolved.

Some reports have argued that patterns of response reflect high-level, categorical differences amongst scenes (Walther et al., 2009, Walther et al., 2011). For example, Walther et al. (2011) showed that the ability to decode scene categories from fMRI data was similar for photographs and line drawings, suggesting some level of invariance to the low level properties of images. However, other studies have suggested that patterns of response in scene-selective regions may be better explained in terms of visual properties of scenes such as spatial layout (e.g., Kravitz et al., 2011, Park et al., 2011, Watson et al., 2014). This latter account is consistent with the sensitivity of the amplitude of response in these regions for orientation (Nasr and Tootell, 2012), spatial frequency (Musel et al., 2014, Rajimehr et al., 2011), visual contrast (Kauffmann et al., 2015), rectilinearity (Nasr et al., 2014), and visual field location (Arcaro et al., 2009, Golomb and Kanwisher, 2012, Levy et al., 2001). Nevertheless, these studies employed univariate analyses, so it remains unclear whether these modulations in the amplitude of response also affect the pattern of response.

In a recent study, we demonstrated that low-level properties of visual scenes (defined by the GIST descriptor; Oliva and Torralba, 2001), predicted patterns of neural response in scene-selective regions (Watson et al., 2014). However, images drawn from the same scene category are likely to have similar low-level properties (Oliva and Torralba, 2001). So, reliable category-specific patterns of response are expected under both categorical and image-based accounts. Therefore, it remains unclear whether patterns are determined primarily by membership of a common category or by the shared low-level image statistics characteristic of that category.

In the current study, we provide a direct comparison of the relative importance of image properties and category in determining patterns of response in scene-selective regions. Participants viewed images from two different categories of scene (indoor and natural) that are known to have distinct image properties (Oliva and Torralba, 2001) and to elicit different patterns of response in scene-selective regions (Walther et al., 2009, Watson et al., 2014). Low-level visual properties of the scenes were manipulated by filtering the images by orientation (Experiment 1) and spatial frequency (Experiment 2) as previous reports have suggested functional biases for these properties (Nasr and Tootell, 2012, Rajimehr et al., 2011). Using multi-voxel pattern analysis (MVPA), we compared the similarity of the patterns of neural response to each condition across the core scene regions (PPA, RSC, OPA). Our prediction was that if scene-selective regions are sensitive to image properties, then some degree of similarity should be seen between conditions sharing the same filter. If scene-selective regions are solely sensitive to category, then conditions sharing the same category should elicit similar patterns of response regardless of the low-level manipulation. The use of pattern analysis allows us to determine whether image properties are an important organising factor in the topography of this region of the brain.