Nineteen healthy, right-handed volunteers (mean age: 25.7 ± 3.8; 14 females) without any neurological or psychiatric disorder completed the study after giving written informed consent. This study was carried out in accordance with the recommendations and the approval of the Ethics Committee of the Freie Universität Berlin. All subjects gave written informed consent in accordance with the Declaration of Helsinki.

The experimental paradigm constituted a 2 × 4 Design with factors CONDITON: perception, imagery and LOCATION: left hand thumb (lH), right hand thumb (rH), left foot big toe (lF), right foot big toe (rF). Vibrotactile stimulation was delivered using 8-dot piezoelectric Braille-like displays (2 × 4 matrix with 2.5 mm spacing) attached to the four body locations, controlled by a programmable stimulator (Piezostimulator, QuaeroSys, St. Johann, Germany). The vibrotactile stimuli were designed to elicit minimal tickling sensation and maximal stimulation intensity to assure a clear percept at fingers and toes. For 8 s a 30 Hz sinusoidal carrier signal was amplitude-modulated with a 2 Hz sinusoidal, with alternating elevation (to maximize stimulation efficiency) of the display’s four rows (Figure 1). Similar vibrotactile stimuli in the flutter range have been used in the study of perceptual and higher order cognitive functions in human and monkey research (Romo et al., 1999; Preuschhof et al., 2006; Spitzer et al., 2010; Schmidt et al., 2017), due to their well characterized, directly stimulus driven processing in SI neurons (Mountcastle et al., 1990; Hernández et al., 1997). For visual guidance through the experiment, visual cues were presented as fixation cross, which changed its color to blue/green to indicate the PERCEPTION/IMAGERY condition, where color-assignment was randomized across participants. Cues indicated where to imagine vibration or where stimulation occurred (upper left, lH; upper right, rH; lower left, lF; lower right, rF; see Figure 1) to match visual sensory stimulation between conditions. During null-events the fixation-cross remained on the screen and no stimulus was applied or imagined. While mounting subjects in the scanner, they were familiarized with the stimuli to enable imagery.

Each of three experimental runs comprised of 24 trials, corresponding to three trials per condition complemented with 6 null-events. The maximum amount of scanning time and thereby the amount of repetitions per trial type was limited due to the demanding nature of the MI tasks. To compensate for the relatively low amount of trials, we applied a block design including null-events to maximize signal changes between conditions. The order of trials was randomized and trials were intermitted by a fixed inter-trial interval of 12 s, which allowed the participants to prepare for the next trial, while no specific preparatory cue was given ahead of each trial. Stimulus presentation was controlled using custom MATLAB code (The MathWorks, MA) and the Cogent 2000 Toolbox (developed by John Romaya at the LON at the Wellcome Department of Imaging Neuroscience). Visual cues were presented on a screen that was visible from the scanner via a mirror system attached to the head coil.

MRI data was acquired in 3 runs of 10.5 min on a 3T TIM Trio (Siemens) at the Center for Cognitive Neuroscience Berlin (CCNB). 310 functional images were recorded per run as T2^∗-weighted gradient-echo EPI: 37 slices; interleaved order; whole brain; TR = 2000 ms; TE = 30 ms; 3 × 3 × 3 mm³ voxel; flip angle = 70°; 64 × 64 matrix. Additionally, a T1-weighted MPRAGE with 176 sagittal slices, TR = 1900, TE = 2.52 ms; 1 × 1 × 1 mm³ voxel was acquired.

Functional magnetic resonance imaging data were pre-processed with SPM8 (Wellcome Trust Centre for Neuroimaging, Institute for Neurology, University College London, United Kingdom). To minimize movement-induced image artifacts each data set was realigned to its mean image. Next, EPI images were normalized to MNI space using unified segmentation (as implemented in SPM8) and re-interpolated to 2 × 2 × 2 mm³ voxel size. Data was smoothed with an 8 mm FWHM Gaussian kernel. Smoothing of data for the Conjunction analyses and contrasts between PERCEPTION > IMAGERY was limited to 5 mm FWHM to preserve a high degree of regional specificity in the group level analysis.

Statistical analysis was performed according to a standard general linear model (GLM) approach. The first level design was specified to model the 8 task-conditions as independent regressors, the six null-events were split into two separate regressors (to allow independent baseline contrasts in the conjunction analysis) and a constant for each run. To test for activation clusters shared by PERCEPTION and IMAGERY we computed first-level contrasts of task-condition against null-events, with independent null-event regressors for the PERCEPTION and IMAGERY contrasts. The resulting contrast images were forwarded to a second level flexible-factorial design. We computed four conjunction analyses between perception > null-events and imagery > null-events contrast images, individually for each body location. Conjunction analyses were evaluated against the conjunction null hypothesis as implemented in SPM (Friston et al., 2005; Nichols et al., 2005).

All reported coordinates correspond to MNI space. The Anatomy Toolbox was used to establish cytoarchitectonical references where possible (Geyer et al., 1999, 2000; Grefkes et al., 2001; Eickhoff et al., 2005, 2006, 2007). Activation clusters are reported at p < 0.05, family wise error (FWE) corrected at the voxel level, with a cluster size threshold of 22 voxel, corresponding to the cluster size threshold of false discovery rate correction. Thresholded statistical parametric maps were rendered on a standard 3D brain template using MRIcron (by Chris Rorden; Version 6 6 2013).

To extract contrast estimates we used the activation clusters of the PERCEPTION conditions and masked them with SI subregion specific anatomical masks of the Anatomy Toolbox.

To test what brain regions are related to the imagery process independent of the exact location of MI, we computed the contrast IMAGERY > PERCEPTION, pooled over conditions (Figure 2). The depicted network reflects task demands of generating a mental representation from memory. The network spans regions of the task-positive network, specifically, supplementary motor areas (SMA) and frontal eye fields (FEF)/premotor cortex (preMC). Further we found bilateral medial frontal gyrus and bilateral inferior frontal gyrus (IFG). Activation in the intraparietal sulcus (IPS) was left-lateralized (Table 1).

We first tested what regions are locus-specifically activated during perception. For each of the four PERCEPTION-conditions the first level contrast against null-events was contrasted against the three other PERCEPTION-conditions in a second level analysis. Figure 3A displays the results that reflect the well-established somatotopic organization within SI as to separate finger and foot representations and to reflect activation in SI contralateral to the body side of stimulation.

When testing in the same way for imagery related activity, namely by contrasting each of the four IMAGERY conditions against the three other IMAGERY conditions, we found weaker activation than in the PERCEPTION condition. Following our a priori hypothesis, we found somatotopic activation within a bilateral SI/SII mask, however, at an uncorrected significance level of p < 0.001. These contrasts show a shift of activation toward more posterior aspects of SI (Figure 3). This effect is apparent for left and right finger representations, while the less pronounced activation in the toe-representing regions do not allow to observe such a shift. To test if the IMAGERY activation clusters overlap somatotopically with the activation found in the PERCEPTION condition, we computed overlays between the PERCEPTION > other PERCEPTION condition contrasts and IMAGERY > other IMAGERY condition contrasts in Figure 3C.

To formally test for overlapping activation between PERCEPTION and IMAGERY, we computed locus specific conjunction analyses between PERCEPTION and IMAGERY contrasts against null-events (Figure 4A). The conjunction for lH, and rH revealed clusters within SI (p < 0.05 FWE corrected). Inspecting the conjunction analyses for the left and right foot toe at p < 0.001 uncorrected also revealed a contralateral cluster in the SI foot region (Table 1), demonstrating a content-specific overlap of primary cortices activated during perception and imagery.

The well-pronounced activation clusters of the lH and rH conditions allowed us to test for their specificity within SI subregions. To this end, we tested what aspects of SI were activated more strongly for PERCEPTION than for IMAGERY. The results are compared to clusters activated in both conditions, as revealed by the conjunction analyses. Figure 4B shows that BA1 (anterior aspects of the post-central gyrus) is activated mainly during PERCEPTION. The activation cluster revealed by the conjunction PERCEPTION & IMAGERY was limited to the posterior portion of SI, matching the probability maps of BA2 according to the Anatomy Toolbox.

Besides the study of MI, working memory (WM) studies deal with the question of what brain regions are activated to represent mental content. Actually both MI and WM are defined as the representation and manipulation of information in the absence of perceptual stimulation (Tong, 2013). The influential multi-component model of WM proposed by Baddeley suggested a visuospatial sketchpad as a type of blackboard or buffer on which visual information is stored (Baddeley and Hitch, 1974). Major aspects of such sketchpad-like representations are thought to be realized by visual cortices. So called sensory recruitment models of WM (Pasternak and Greenlee, 2005) emerged from observations of delay-activity in sensory regions and were well compatible with reports from the study of MI about re-activation of perceptual regions during mental reconstruction across all modalities (McNorgan, 2012; Tong, 2013; Schmidt et al., 2014). Also the MI literature emphasized the importance of topographically organized early sensory cortices for implementing a type of visual buffer (Kosslyn, 2005). Kosslyn compares it’s function to a pegboard, where different types of local information are represented topographically.

Multiple recent fMRI-WM studies using multivariate pattern analysis (MVPA) have provided support for sensory recruitment models by demonstrating that visual stimulus features can be decoded from visual cortices during WM (Christophel et al., 2015; Lee and Baker, 2016). These findings have lead to the view that sensory regions implement a memory buffer for all types of visual information (Albers et al., 2013; D’Esposito and Postle, 2015), which could consecutively also being speculated to realize the representation of mental content generated during MI.

In the tactile modality, delayed EEG activity over somatosensory regions during WM retention periods indicated distinct modality specific activation in somatosensory regions (Katus and Eimer, 2015; Katus et al., 2015). Results of a recent tactile fMRI MVPA study point in the direction that it is not primarily sensory regions that code mental contents (Schmidt and Blankenburg, 2018). This study revealed that hierarchically higher, modality-independent regions retained information throughout a 12 s WM delay phase, when participants memorized the spatial layout of a Braille-like stimulus. In contrast, somatosensory cortex was found to represent stimulus information only briefly during an early, potentially stimulus encoding, phase. Another tactile working memory study tested for the retention of the frequency of vibrotactile stimuli (Schmidt et al., 2017). Parametric multivariate codes of working memory content were found rather in prefrontal than in sensory regions, as suggested by previous electrophysiological studies (Romo et al., 1999; Spitzer et al., 2010). In sum, WM and MI studies do not allow a final conclusion about what aspects of mental images, or mental representation in general, are represented in somatosensory cortices. Results from research on visual MI and WM suggest that the levels of abstraction and sensory-detail determine whether lower order hierarchical sensory regions are recruited during MI and WM.

Here, we used simple vibratory stimuli that delivered standardized, vibratory stimulation to elicit a clear percept. In contrast to an earlier tactile MI study (Schmidt et al., 2014), the applied stimuli did not contain any relevant spatial information. Consequently, the employed task did not necessitate to generate mental images with fine-grained spatial sensory details. This might elucidate why in the current study not all parts of SI are recruited during MI, but only the higher subregion BA2 gets top-down recruited during MI. Our results could therefore also be interpreted along the view that it is the degree of sensory vividness that determines to what degree sensory regions are recruited during MI (Kosslyn and Thompson, 2003). Specifically, the more abstract information of a mental representation is represented less activation will be found in low order sensory regions. Testing this suggestion will require studies that ideally include trial-by-trial based assessments of MIs vividness. Further, different types of tactile stimuli are required to understand what stimulus features determine how these are represented in lower sensory regions. Finally, research might focus on the distinction between stimulus information being represented as categorical information, e.g., after intensive training and naming of stimuli, or rather in sensory formats as during perception.

Similarly as in previous MI studies in different modalities (McNorgan, 2012), we found the activation strength in sensory cortices to be smaller during MI than the activation during perception (Figure 4B). While vibratory stimuli have been intensively used in human WM studies (Auksztulewicz et al., 2011; Spitzer and Blankenburg, 2011; Spitzer et al., 2014), it remains an open question in how far the MI of such stimuli reaches a mental vividness to necessitate recruitment of lowest level sensory regions, which might be higher for other types of tactile stimuli. The question of what exact subprocesses are shared by imagery and perception and what subpopulations of neurons realize them is to be addressed in future research.

The construction and maintenance of a mental image can be subdivided in different cognitive subprocesses (Kosslyn, 2005; Buckner and Carroll, 2007; Hassabis and Maguire, 2009). These include the access of long-term memory and the re-activation of a mental content by making it available for conscious processing. Most of these processes operate closely entangled with attentional mechanisms. It has become a debate whether it is possible to dissociate attentional mechanisms from neuronal processes that reflect the mere representation of a mental content, or the “pure” mental image. This challenge has been discussed in different process models of MI (Kosslyn, 2005) and it has been argued that it is not meaningful to consider them as independent psychological constructs (Gazzaley and Nobre, 2012).

Within the WM literature it was suggested to dissociate the representation of mental content from so called cognitive control mechanisms (Riggall and Postle, 2012; D’Esposito and Postle, 2015). It is argued that activation in task-positive networks, including the posterior parietal regions, rather reflect the focus of attention than the content of WM (Postle, 2015). However, it appears equally plausible to consider the content of MI (the mental image as such) as an attentional mechanism of re-activating long-term memory. Some authors even term imagery processes reflective attention (Chun and Johnson, 2011). More specifically, the mental representation of stimulus features might recruit attentional mechanisms that elicit activation patterns dissociable between stimuli, while not directly reflecting the stimulus properties as such. Puckett et al. (2017) recently demonstrated with high-field fMRI that attention modulates neuronal responses to tactile stimuli in a somatotopic fashion. In how far the neuronal populations modulated by attention during perception overlap with the neuronal populations activated by MI remains a central question for future MI studies. Our study will hopefully inspire high-field fMRI studies with their improved spatial resolution to map subject-specific attentional and imagery related somatotopic maps, possibly even in the subregions of SI.

Our study is limited with regards to clarifying the contributions of spatial attention to the imagery process. To limit the complexity of the study design, we did not include an attentional control condition to assess the effects of spatial attention toward a body location without imagination. Such an extension in future studies could reveal how spatial attention and MI differ in their neuronal implementation.

Besides the activation in SI it is most likely that multiple mental codes jointly represent mental content in distributed interacting cortical networks (Schlegel et al., 2013; Larocque et al., 2014; Lewis-Peacock et al., 2015; Postle, 2015; Lee and Baker, 2016). While much neuroimaging work on MI is focused on sensory cortices, there is a demand for conceptual improvements in the definition of the individual processes and psychological constructs that contribute to the MI process, particularly refined definitions of attentional contributions. Until then, it is difficult to assign clear functional labels to all the regions involved in MI. Here, we computed the main effect of IMAGERY versus PERCEPTION. This contrast reveals a distributed attention-related network of regions. The individual functional contribution of these regions remains an open question.

Within the last few decades, the predictive coding (PC) framework has been promoted, most famously in the work of Karl Friston (Friston, 2009, 2010; Friston and Kiebel, 2009). This theory of global brain function has gained notable popularity beyond the neuroscientific community (Hohwy, 2014). Until now, most work on predictive mechanisms has been carried out in the context of action and perception. Very recently, however, Parr and Friston (2017) also proposed how WM might fit into this framework. They suggest WM as a process of evidence accumulation, with the main purpose of optimal policy selection for acting within the environment. This idea could be applied to the domain of MI in a very similar way.

The basic assertion of PC is that our brains continuously generate predictions about future sensory events. These predictions rely on the interaction between bottom-up and top-down signaling within the cortical hierarchy. Neuronal signals coding sensory predictions are propagated through the hierarchy via top-down connections. Sensory signals that reach the cortex from the sensory organs are propagated via bottom-up connections. If accurate predictions are generated, they match the sensory input signals. If predictions do not match the bottom-up signals, a prediction error is generated. This error signal is used to update the generative model by backpropagation to higher-order regions. The continuous influx of information from our senses requires an equal continuity of predictions. The temporary retention of information during WM and MI has to act against this continual flow of new information. Somehow, the overwriting of information as a result of the continuous sensory influx needs to be overcome in order to temporarily retain information about a specific content. Since hierarchical processing is, however, an evolutionarily old and computationally efficient principle, it is likely that the retention of mental content, as necessary for MI, is realized within the constraints of these processing principles, rather than independently of them. Depending on the level of abstraction in which a mental image is generated, an internally generated top-down signal will generate prediction errors at corresponding hierarchical levels. The energy consuming process of error generation could be the neuronal correlate of the BOLD signal changes as found in the study at hand. As this activation is driven by top-down signals and generated by different neuronal populations than bottom-up driven activation, this could also explain the difference in activation strength between imagery and perception.

Different MI and WM tasks require mental representations with different levels of detail. More abstract mental images will elicit such signals on hierarchically higher brain regions, while those mental images that have vivid sensory details will induce activation in lower order sensory regions. This PC perspective on MI can explain the different findings on sensory recruitment and explain the data from visual MI that more vivid mental images activate regions lower in the hierarchy (Kosslyn and Thompson, 2003). The tactile MI study at hand supports this principle. While imagery of a simple stimulus revealed only the hierarchically higher SI subregion BA2, the imagery of fine-grained layout information in our earlier tactile MI study (Schmidt et al., 2014) activated the hierarchically lowest cortical regions BA1 and BA3b.