Twenty-nine adolescents [2 females; ages 11–17 years (mean = 14.3); 2 left-handed] with normal or corrected-to-normal vision were recruited from the Philadelphia community to participate in the main fMRI experiment. All participants gave written informed consent in accordance with procedures approved by the Children's Hospital of Philadelphia Institutional Review Board and were paid for their participation. Prior to the fMRI session, subjects completed a mock scan procedure, allowing the participants to acclimate to the scanner environment and train to minimize movement while scanning. Only participants who were under a minimum movement criterion preceded to the scanning session. None of the participants moved more than 3 mm during any scanning run. Three subjects were eliminated from the connectivity analysis because they did not show activation within the region of interest used to define the seed region.

Imaging data were collected using a 3T Siemens Verio scanner and a 12-channel head coil. Two structural MR images were acquired for the registration of fMRI data to standard space: a high-resolution T1-weighted MPRAGE sequence of the entire brain (176 sagittal slices, isotropic voxel size = 1 mm, TR = 1900 ms, TE = 2.54 ms, flip angle = 9°), and a high-resolution FLASH sequence collected in the same axial plane as the fMRI data (number of slices = 40, slice thickness = 3.5 mm, TR = 300 ms, TE = 2.46 ms, flip angle = 60°). Functional data consisted of two 4-minute runs of whole-brain T2^* weighted BOLD echoplanar images with 107 volumes acquired per run (40 oblique axial slices, isotropic voxel size = 3.5 mm, TR = 2340 ms, TE = 25 ms, flip angle = 90°).

Stimuli of interest were 32 gray scale fearful faces and 32 houses presented within 2° of visual angle into the left lens of MR compatible dual display LCD goggles (Resonance Technology Inc., Northridge, CA). Responses were recorded with a four-key fiber optic response box. Task stimuli consisted of movies of colorful Mondrian images changing at a rate of 10 Hz. Mondrian images were created using Matlab, with each 28-s block movie consisting of 280 unique dynamic noise images, each presented for 100 ms. Images were made into movies using Corel Video, with letters and fixation cross images added to these movies before exporting the movies to Quicktime. Experimental presentation was done with Psychopy (Peirce, 2007, 2009).

A fixation cross appeared in the center of the Mondrian movies, and uppercase letters from the English alphabet appeared in one of four quadrants immediately adjacent to the fixation cross. Letters consisted of 5 vowels (A, E, I, O, U) and 5 consonants (C, H, N, T, S). The task consisted of 12 28-s blocks (12 TRs, 2340 ms each TR). Within a block, letter trials appeared in the right eye and stimulus trails to the left, which was experienced by the subject as one image (see Figure 1A). Following is a description of these trials as incorporated into a block (For visual schematic, see Figure 1B). Each trial was a total of 2340 ms, the length of one TR. Projected through the right lens, a block began with a continuous stream of Mondrian images changing at a rate of 10 Hz. After 2340 ms, the first of 10 letter trials was presented. A letter trial consisted of a 300 ms fixation cross, followed by the appearance of a letter in one of the four quadrants for a duration of 1500 ms. Onset of the letter trials was varied by 300–600 ms from the start, with the difference in onset accounted for at the end of a trial, such that each letter trial was 2340 ms. At the end of 10 trials, only the Mondrians appeared for 2340 ms (no letters or fixation) and then the block was complete. In each block, all 10 letters were presented, with letter order and onset variance randomized between blocks.

Stimuli of interest were projected through the left lens, blocked by stimulus category, with category order counterbalanced across participants. Eight fearful faces, eight houses, or a no-stimulus control were presented in each block. A block began with a black screen for the first 4680 ms. After this period, 8 stimulus trials were presented. A stimulus trial began with a stimulus that appeared after 600 ms. The stimulus was slowly ramped from a contrast level of 0 to 100 over 750 ms and ramped back down over the following 750 ms (total duration: 1500 ms). The left screen was then blank for another 340 ms until another trial began. Following the presentation of 8 trials, no stimulus appeared for another 4680 ms until block completion. Task blocks were separated by 11,700 ms of rest, with a black screen presented to both eyes. It should be noted that for the no stimulus control condition, a black screen was presented to the left eye for the entire 28-s block, while the task still appeared in the right eye.

The main fMRI experiment consisted of two 4 min 20 s scan runs, each of which was divided into 6 task blocks and 7 periods of rest. During each block, participants viewed letters that appeared surrounding a central fixation. They were instructed to press the right button if the letter was a vowel and the left button if the letter was a consonant. Following the presentation of the 2 runs, a catch trial was presented. A catch trial consists of a fearful face or house image presented atop of the Mondrian image to both eyes, in order to mimic break from interocular suppression. This trial is used as a probe to assess whether participants experienced break from interocular suppression earlier in the experiment. Following the catch trial, participants were asked, “Did you notice anything different about the last 2 trials?” All participants reported the presence of a face and a house. They were then asked if they saw any objects or parts of objects earlier in the experiment. All participants reported that they did not see objects prior to the catch trial, indicating successful suppression of the objects for the duration of the experiment.

Following the main experimental scans, a 5-min functional localizer scan was administered, in which subjects detected when a centrally presented white crosshair appeared on full color faces, scenes, objects, and scrambled objects, presented in a blocked design. Four, 14-s blocks of each image category were presented as “superblocks,” in which the stimulus category blocks were presented in succession and separated by 14 s of rest. Each “superblock” sequence was presented four times, with object categories in a different order for each “superblock.”

Image preprocessing and statistical analyses were performed using SPM8 (Wellcome Trust Centre for Functional Neuroimaging, London, UK). Functional images from both experimental and localizer scan runs were initially analyzed separately for each participant. Low-frequency drifts were removed with high-pass filtering with a cutoff period of 128 s and autocorrelations modeled using a first-order autoregressive model. Images for each participant were realigned to the first image in the series (Friston et al., 1995) and coregistered with the structural image (Ashburner and Friston, 1997). The transformation required to bring a participant's images into standard MNI152 space were calculated using tissue probability maps and these warping parameters were then applied to all functional images for that participant (Ashburner and Friston, 2005). The data were spatially smoothed with a 4 mm FWHM isotropic Gaussian kernel.

We first assessed the activation pattern evoked by the conscious task (flashing Mondrian images presented to the right eye). To examine this, we averaged activation across the three covariates (fearful faces, houses, and control) compared to a resting baseline (12 s blocks of rest). Because the Mondrian images are consistent across these three conditions, we expected activation in regions of the central visual system. Indeed, participants activated bilateral lateral geniculate nucleus (LGN) and EVC (Figure 2A; Table 1). We then explored whether there were differences in EVC between three conditions by extracting subject's parameter estimates from each condition, separately, using a mask defined by the regions reaching whole brain significance. (We chose not to explore the LGN signal further, as there is a great deal of anatomical variability in subject anatomy and we would be unable to differentiate the LGN from surrounding structures). We observed a significant effect of condition in EVC, bilaterally [left F₍₂₎ = 6.83, p = 0.002; right F₍₂₎ = 12.01, p < 0.001] (Figure 2B). However, this was driven by stronger activation when there was no stimulus presented to the left eye compared to a fearful face or house stimulus [RIGHT: faces t₍₂₈₎ = 3.92, p = 0.001; houses t₍₂₈₎ = 5.17, p < 0.001; LEFT: faces t₍₂₈₎ = 2.98, p = 0.006; houses t₍₂₈₎ = 3.54, p = 0.001]. We find no significant differences between fearful face and house conditions in EVC.

When contrasting the conditions with a stimulus (fearful faces or houses) with the no stimulus control condition, we find a single cluster of activation that encompasses right lateralized superior colliculus, thalamus, amygdala, and hippocampus, can a reference to Figure 3A and Table 1. These results are consistent with an abundance of previous work implicating these regions in implicit perception and vision without awareness (De Gelder et al., 1999; De Gelder and Hadjikhani, 2006; Tamietto et al., 2009; Stienen and De Gelder, 2011; Van den Stock et al., 2011). However, there are no differences between fearful faces and houses based on mean activation in these subcortical regions. Even when we lower this contrast to an excessively liberal threshold (p < 0.05, uncorrected), the regions showing mean differences to stimulus vs. no stimulus are only in subcortical areas. Based on our a priori hypothesis regarding the amygdala, we examine responses in this region statistically using a region of interest approach, described below.

The amygdala is frequently activated by social information and is thought to play a particular role in guiding orientation responses to visual social stimuli (Adolphs and Spezio, 2006; Adolphs, 2008, 2010). We have previously found amygdala activation to fearful faces (an emotional, social stimulus) in the absence of awareness (Pasley et al., 2004; Troiani et al., 2012). Thus, we expected a differentially stronger response in the amygdala for fearful faces compared to houses. We explored this hypothesis by examining responses in bilateral amygdala, for each of three regions defined by cytoarchitectonic probabilistic maps (Amunts et al., 2005). Contrary to our hypothesis, we did not find amygdala activation that was specific to fearful faces. Instead, in all amygdala ROIs, we observed an effect of condition (stimulus vs. no stimulus) in bilateral superficial amygdala and the right centromedial amygdala [Left SF: F₍₂₎ = 3.18, p = 0.049; Right SF: F₍₂₎ = 7.15, p = 0.002; Right CM: F₍₂₎ = 6.74, p = 0.002], but there were no differences in activation between fearful faces and houses (Figure 3B). Please note that these are relative differences in activation, such that in the control condition, the amygdala is quite suppressed compared to baseline. The amygdala is known to undergo suppression compared to a resting baseline during an attention-demanding task (such as detecting letters in a noise pattern) (Morawetz et al., 2010; Stjepanovic et al., 2011). Thus, we interpret the less negative amygdala response to fearful face and house stimuli as a small break from the suppression of the amygdala. While we did not observe a category-specific response in the amygdala to fearful faces, we go on to explore the connectivity profile of the right superficial amygdala, based on its involvement in social processing (Goossens et al., 2009; Bos et al., 2013; Bzdok et al., 2012).

We previously identified a network of increased coherence with the left amygdala BOLD signal for suppressed fearful face presentations compared to suppressed houses (Troiani et al., 2012). One goal of the current study was to examine whether this network existed with a more robust form of interocular suppression. Based on our finding of right superficial amygdala activation to both faces and houses, we used this region to guide a connectivity analysis. We reasoned that despite the lack of differential mean activation in this region based on the category of the stimulus, perhaps this activation leads to increased connectivity for one stimulus (fearful faces) more than another (houses), based on its motivational value. We used regions of interest from the results of our previous connectivity analysis to guide our search. These seven ROIs included bilateral pulvinar, bilateral insula, left inferior parietal cortex, left frontal eye fields, and EVC (Figure 4A). We find increased coherence between the right superficial amygdala seed and two regions, including the right pulvinar and left inferior parietal cortex (Figure 4B). These results suggest that the pulvinar and parietal cortex may be amongst the earliest regions to differentiate between motivational stimuli, a point we will take up further in the discussion.

The FFA and PPA are regions typically defined based on their category-selectivity (Kanwisher et al., 1997; Epstein and Kanwisher, 1998). In conscious vision, the FFA responds most strongly to faces compared to other objects, while the PPA responds most robustly to scenes or houses and not at all to faces. In studies of non-conscious vision, activation in category specific regions is thought to reflect stimulus awareness, as activation in these regions may indicate that the signal from the visual stimulus has escaped suppression enough to proceed beyond early regions in the visual processing hierarchy and reach higher level processing regions. Although, some studies have found activation in category-specific visual cortex without awareness albeit at much lower levels compared to responses to conscious stimuli (Jiang and He, 2006; Troiani et al., 2012). Given the link between conscious awareness and activation in category-selective cortex, we examine mean responses in the FFA and PPA to all three conditions (fearful faces, houses, no stimulus control). We find no differences between the three conditions in either FFA [F_{(2, 25)} = 0.63, p = 0.54, n.s.] or PPA [F_{(2, 25)} = 1.92, p = 0.160, n.s.], indicating that the stimuli are not escaping suppression enough to reach ventral visual cortex.

A previous study found that faces and houses presented in the absence of awareness were associated with distinct multi-voxel patterns in object-selective cortices (Sterzer et al., 2008). These results suggested that some amount of information escapes suppression and reaches object-selective cortex differentially by object type (i.e., FFA for faces and PPA for houses). To examine whether object-related information was present in our own data, we employed a multi-voxel pattern analysis. We find no evidence that signals in subject-specific FFA or PPA are able to discriminate between fearful faces, houses, or control [FFA left: t₍₂₀₎ = 1.32, p = 0.203; FFA right: t₍₂₀₎ = 0.93, p = 0.365; PPA left: t₍₂₀₎ = 1.09, p = 0.288; PPA right: t₍₂₀₎ = 0.992, p = 0.332]. In conjunction with the null univariate results described above, these results suggest that stimulus information does not escape suppression enough to reach higher-level cortex in the current experiment.

Although the combination of flash suppression and rivalry used in our previous study (Troiani et al., 2012) is referenced as a form of CFS, there are a few, important differences. In our previous design, we used a single red/blue rivalrous image that was viewed through anaglyph glasses. Because it is difficult to exactly match the colored lenses of the anaglyph glasses and the color of the rivalrous stimuli, it is possible that certain wavelengths can “leak through” from the suppressed image into the dominant eye. Here, we used MR compatible goggles with a dual LCD display, which allowed for stimulus presentation directly into one eye without the possibility of wavelength-based “leak through” of information. Previously, we induced motion suppression through the use of a centrally presented word/checkerboard stimulus that moved around the screen. In practice, this was quite suppressive—and participants were still not explicitly aware of the stimuli. However, completely changing a colorful, dominant stimulus at a rapid rate [as in the type of CFS described by Tsuchiya and Koch (2005)] is a much stronger form of suppression. Indeed, there are published reports indicating that not all forms of rendering a stimulus unconscious produce similar results. Although backward masking (BM) has been used for years to render stimuli invisible, the use of backward masking and CFS in otherwise identical paradigms produce different results. In a behavioral study using BM in one experiment and CFS in another, non-conscious affective priming was achieved for both happy and angry faces in the BM study and only for angry faces using CFS (Almeida et al., 2013). Thus, it appears that using the less robust BM technique, more information is processed despite equivalent phenomenological suppression. CFS can also be implemented in multiple ways. One method is to use a spectrally separated image and anaglyph glasses. Another is to present completely independent inputs: this can be done by separating the eyes with a piece of cardboard and using prism goggles or using a dual display head mounted device that presents separate images to each eye. Although the suppression strength of (1) CFS using a single, spectrally separated image and anaglyph glasses, and (2) CFS using independent visual inputs has not been explicitly compared using the same paradigm, findings from previous work also suggests potential differences between these implementations. Fang and He (2005) used anaglyph glasses and spectrally separated images in a CFS paradigm to demonstrate that categorically-distinct information can reach the dorsal stream but not the ventral stream. Furthermore, images of suppressed tools evoked stronger dorsal stream activation than suppressed faces. However, Caplovitz et al. (2010) fail to replicate this effect using a CFS experiment with independent displays. One possible explanation for these discrepancies is that CFS implemented with spectral images and anaglyph glasses is a less robust form of suppression than CFS implemented with completely independent visual inputs (due to wavelength-based “leak through” of information from one image to the opposite eye). Similar to the reported differences in suppression strength between BM and CFS, we suggest there may also be differences between using CFS-induced suppression with anaglyph glasses and a single, spectrally altered image compared to CFS-induced suppression with completely independent visual inputs.

In our previous study, stimuli could be differentiated based on mean amygdala activation. That is, unseen fearful faces activated more amygdala than unseen houses. Accompanying this greater amygdala activation was left parietal activation for suppressed fearful faces compared to suppressed houses and increased connectivity with multiple regions involved in attention. Thus, we expected to replicate our previous finding of category-specific activation in the current study, despite employing several methods to further prevent escape from suppression. In contrast to our hypothesis, we show equally robust amygdala activation to both fearful faces and houses presented outside of awareness. At the whole brain level, both suppressed stimulus categories activated the right superficial amygdala, a result that was confirmed with a more thorough analysis of amygdala subregions.

In our previous work, we also identified a network of region that increased in coherence with the amygdala for fearful faces compared to houses (pulvinar, insula, inferior parietal, frontal eye fields, and EVC). We partially replicate this finding in the current study: the amygdala connectivity profile showed differential increases in connectivity for suppressed fearful faces compared to suppressed houses. Specifically, we find increased task-specific coherence between the amygdala and two regions that are part of the attention network identified in our previous work: the right pulvinar nucleus of the thalamus and left inferior parietal cortex. Taking the results of both studies together, we speculate that under the less robust suppression induced previously, more information was able to escape suppression and activate a broader network involved in preattentive stimulus processing. With this greater information breaking through, feedforward, and feedback signals between regions in this network may strengthen their communication and lead to the mean activation differences observed in our previous study.

We also examined mean activation and multivoxel pattern differences in cortical regions associated with category-specific processing of faces (FFA) and houses (PPA). In our previous study, we found that fearful face-specific amygdala activation was accompanied by activation in the left FFA, but no activation in PPA for either suppressed faces or houses. In the current study, we find no category-specific activations to the suppressed stimuli. Furthermore, there seems to be no information at all about the presence of a stimulus in high-level visual cortex, as there were no activation differences in either FFA or PPA for the presence of a suppressed stimulus vs. no stimulus. Additionally, neither the FFA nor PPA could discriminate between the presence of a stimulus vs. no stimulus based on multi-voxel patterns, providing further evidence that stimulus information was not reaching high-level visual cortex and indicating that these stimuli were robustly suppressed from awareness.

Because we additionally included a no stimulus control condition in the current design, we were also able to compare activity in visual processing regions in response to the presentation of a suppressed stimulus compared to no stimulus. Unsurprisingly, we show that the main task activates bilateral LGN and EVC, consistent with information processing by a retino-geniculate-cortical pathway. When further exploring activation in EVC to each condition separately, we find significant differences between stimulus presentation and control. More specifically, the control condition correlated with more activation in EVC than the two suppressed stimulus conditions. V1 is the first stage in the visual processing hierarchy at which the information from both eyes is combined. In previous studies examining the neural bases of binocular rivalry, activation in V1 has been concomitant with awareness. That is, when subjects were asked to report whether they perceived one rivalrous stimulus compared to another, activation in V1 strongly correlated with the reported percept (Polonsky et al., 2000; Tong and Engel, 2001; Lee et al., 2007). When stimuli are reliably suppressed, this is associated with suppression in V1 (Lee and Blake, 2002). Because observers remained unaware of the stimuli presented to their left eye for the duration of the study, this pattern of activation in EVC likely reflects successful suppression of the fearful face and house stimuli.

These results are also informative with regard to the idea of parallel visual pathways. Visual signals originate from the retina and project to the LGN and onto the primary visual cortex (V1), located in the posterior occipital lobe, surrounding the calcarine fissure. It is thought that a parallel pathway exists which projects from the superior colliculli to the thalamus, and onto the amygdala. In our data, we show that the presence of a stimulus appears to reduce activation in EVC. In contrast, we show that stimulus information activates regions of the superior colliculus, thalamus, hippocampus, and amygdala, indicating that information has reached structures of the superior colliculus-pulvinar-amygdala pathway. This suggests that information can reach subcortical regions and influence the amygdala without corresponding information representation in higher-level visual regions (FFA/PPA) or even lower level cortical visual regions (EVC).

We also find hippocampal activation when stimuli are present (but suppressed). This finding is consistent with models of fear conditioning that implicate hippocampal-amygdala connections in contextual fear conditioning. For example, Alvarez et al. (2008) found right anterior hippocampus and bilateral amygdala activation for the conditioned stimulus in a foot shock fear conditioning paradigm, but only when preceded by the associated context. Amygdala-hippocampal connectivity increases bidirectionally when retrieval of emotional information is relevant to the current behavior (Smith et al., 2006). Furthermore, unseen primes have been shown to generate predictive signals related to stimulus history and influence the percept selected in a binocular rivalry paradigm (Denison et al., 2011). Thus, it may be that predictive signals are generated by the hippocampus even with the minimal amount of information that escapes interocular suppression. Such a predictive signal would aid in the prioritization of particularly relevant stimuli.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.