Participants

Twenty-six healthy volunteers participated in the MEG experiment (8 males, mean age 24, 19–34 years) and sixteen participants (5 males, mean age 24, 18–34 years) took part in a separate behavioural experiment. All participants were native English speakers, right handed, had normal vision and reported no history of language disorders, neurological or psychiatric illness. The current study was approved by the Research Ethics and Governance Committee of the York Neuroimaging Centre, University of York, UK, and written informed consent was obtained from all participants. Six MEG datasets were excluded from analysis due to excessive movements or artefacts (see data acquisition and pre-processing for details).

Experimental design and procedures

This experiment employed a word-picture verification task in which pictures of items from two semantic categories (animals–e.g., zebra and manipulable manmade objects–e.g., screwdriver) were identified as members of either these general categories, or at a more specific level (i.e., using their specific names zebra and screwdriver). This gave a 2X2 design in which semantic category determined the relevance of visual and action features to identification, while superordinate- and specific-level trials were compared to manipulate the importance of accessing detailed visual and motor features (cf. [45]). The verbal label (e.g., ‘animal’ or ‘Dalmatian’) defined the level of processing for the target concept. Each condition comprised 120 trials where the word was congruent with the picture; these 480 matching trials were pseudo-randomly intermixed with 180 trials (45 trials per condition) where the word and picture did not match. Examples of matching and mismatching trials are shown in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

(a) Experimental Design. We used a word-picture verification task in which pictures of animals or manmade objects were identified as members of either these superordinate categories, or at a more specific level. (b) Trial Structure. Words and pictures were presented simultaneously and participants were asked to press a button with their left hand when the picture and word did not match. These mismatching trials requiring an overt response occurred on 30% of trials and were not included in the subsequent MEG analysis.

https://doi.org/10.1371/journal.pone.0169269.g001

We used simultaneous presentation of words and pictures, in order to provide a clear onset for semantic retrieval. Similar paradigms have been extensively studied behaviourally, with data suggesting that word processing is faster and dominant over picture processing [56–59]. For example, the picture-word interference paradigm presents words superimposed on pictures and shows that word reading is relatively immune to the simultaneous presentation of semantically related pictures, while semantically-related words influence picture naming [59]. Therefore, in our paradigm, we can be confident that processing the semantic content of pictures is influenced by the level of specificity defined by the words presented on each trial.

The stimuli used for the experiment consisted of colour photographs of 60 animals and 60 manmade manipulable objects, each of which was presented twice using two different images of the same object. These images were selected from a large dataset comprising at least three different images of the same object. Ratings of different semantic and psycholinguistic characteristics of these concepts/images were also collected using separate online surveys employing five independent groups of participants (not included in the MEG study). On 5-point Likert scales, participants rated how much a given picture was a good fit to a particular concept (Image Agreement), the extent to which that concept was representative of the category to which it belongs (animal or manmade object) (Typicality), how familiar the concept was (Familiarity) and the extent to which visual, action and visual-motor features defined the concept (Semantic Features). Finally, for the two pictures of each concept with the higher image agreement score, we assessed how consistently each picture was named at the most specific level (Concept Agreement) and used those labels for the specific-level condition. The lists of words used for defining the picture at specific level were matched between conditions in terms of number of letters (p = .84) and lexical frequency (p = .71) obtained from the SUBTLEX_UK database [60](for details see S1 Table). Similarly, the stimuli used for the two conditions were matched for image agreement (p = .20), name agreement (p = .35) and familiarity (p = .23), but also for the number of non-white pixels as a measure of image complexity [61](p = .72). The items in the Animal and Manmade conditions were significantly different in terms of their predominant semantic features: animals scored more highly for visual (p < .001) and visual-motion features (p < .001), whereas manmade objects scored more highly for action features (p < .001) (see S1 Table for details). The two lists also differed for typicality (p = .01), because animals were generally rated as more typical members of the superordinate category than manmade objects (animal = 4.18, manmade object = 3.86); this is likely to reflect the greater similarity between the features of animals than manmade objects. The colour photographs of the animals and manmade objects selected for the MEG experiment were presented on a back projection screen (60Hz) in a dark room. Images appeared in a 14.4° × 14.4° region in the centre of the screen (when viewed from the standard distance of 60cm) which was set to a mid-grey level. The average luminance of the stimulus display region was in the mesopic range. As illustrated in Fig 1, a pair of red nonius lines was presented in the same mid-grey region, helping participants to maintain a steady fixation. The nonius lines were present throughout each experimental run, except when replaced briefly by stimulus pictures which appeared in the middle of the patch for 300ms. The contrast in the stimulus pictures was reduced sufficiently to ensure that the superimposed dark grey text was easily visible (Arial Monospace). Each letter subtended ~0.75° horizontally.

Each trial started with the nonius red lines. After a variable interval of 700–1000 ms, the target was projected onto the screen for 300 ms. The inter-trial interval (ITI) varied between 900 and 1200 ms, and this was increased by 3000 ms in the event of a button press in the MEG version of the experiment. The behavioural pilot of this experiment required the participants to use the two mouse buttons to indicate whether the compound picture-word either matched or not, while in the MEG experiment, participants were asked to press a button with their left index finger only when picture and word did not match, and these trials were excluded from further analysis of the MEG data. The experiment was controlled using Presentation 16.1 (Neurobehavioral Systems). To familiarize participants with the task, 20 practice trials were performed at the beginning of the experiment but they were not included in the analysis. For each participant, the experiment was administered in six blocks of approximately 7 minutes each, separated by self-paced breaks. Block order was randomized across subjects. For the MEG experiment, participants were instructed to keep still throughout the experiment, and to avoid any movement not related to the task. They were asked to blink only after making a button press.

Data acquisition and pre-processing

Before MEG data acquisition, participants’ head shape and the location of five head coils were recorded with a 3D digitizer (Fastrak Polhemus). The signal from the head coils was used to localise participant’s head position within the helmet before and after the experiment. For each participant, a high-resolution structural T1-weighted anatomical volume was acquired in a GE 3.0 T Signa Excite HDx system (General Electric, USA) at the York Neuroimaging Centre, University of York. The 3D digitized head shape of each participant was used for the co-registration of individual MEG data onto the participant’s structural MRI image using a surface-based alignment procedure [62].

MEG data were collected in a magnetically shielded room using a whole-head 248-channel, Magnes 3600 (4D Neuroimaging, San Diego, California), with the magnetometers arranged in a helmet shaped array. Data were recorded in continuous mode, with a sampling rate of 678.17 Hz and pass-band filtered between 1–200 Hz. MEG signals were subjected to a global field noise filter subtracting external, non-biological noise detected by the MEG reference channels, and converted into epochs of 1300 ms length, starting 500 ms before the target onset. Mismatch trials were discarded from any functional analysis. Each epoch was visually checked and excluded from further analysis in the event of response errors and/or artefacts, such as eye blinks, other movements, or electrical noise. Statistical analyses included only datasets with at least 75% of trials retained after artefact rejection. Twenty datasets reached this criterion. On average, 12% of the trials were rejected from these datasets (min 5%—max 25%).

Analysis strategy

The spatial and temporal resolution of the MEG recordings was exploited in a two-step analysis: first, we examined the response of the whole brain to the task (across conditions) and (in a supplementary analysis) to the main effects of specificity and semantic category, at a coarse frequency resolution and averaging out the temporal component. Secondly, we interrogated the activity of specific cortical regions engaged by the task at a finer frequency and temporal scale.

For both analyses, the neural sources of the brain activity were reconstructed with a modified version of the vectorised, linearly-constrained minimum-variance (LCMV) beamformer described by Van Veen et al, 1997 [63] and referred by Huang et al., 2004 [64] as Type I beamformer, implemented in the Neuroimaging Analysis Framework pipeline (NAF, York Neuroimaging Centre), using a multiple spheres head model [65]. An MEG beamformer (spatial filter) allows an estimation of the signal coming from a location of interest while attenuating the signal coming from other points in the brain. This is achieved by constructing the neuronal signal at a given point in the brain as the weighted sum of the signals recorded by the MEG sensors. Independent beamformers were reconstructed for each point in the brain, in each of three orthogonal current directions, separately. In our analysis, the covariance matrix used to generate the weights of each beamformer was regularized using an estimate of noise covariance as described in Prendergast et al., [66] and Hymers et al., [67]. This procedure was performed separately for each condition and/or analysis window, in order to obtain an optimal sensitivity to the effect of interest [68, 69]. The outputs of the three spatial filters at each point in the brain (referred to as a Virtual Electrode) were summed to generate the total oscillatory power, thus combining both phase locked (“evoked”) and non-phase locked (“induced”) signal components [70]. For the whole-brain analysis, a noise normalised volumetric map of source total power was produced over a given temporal window and within pre-specified frequency bands. For the region of interest analysis, the time course information at the location specified was reconstructed and the time-frequency decomposition was computed using Stockwell Transforms [71].

This analysis strategy and the parameters used for the current study were similar to those used in recent MEG studies of visual word recognition and object naming [51, 72, 73]. All information necessary to reproduce these analyses is stated below and the analysis pipeline is also in the public domain (http://vcs.ynic.york.ac.uk/docs/naf/index.html).

Whole-brain beamforming

The brain’s response to the task was characterised within broad frequency ranges across 500 ms (averaging out the temporal component). The purpose of this analysis was to identify brain regions important for the task in general terms, so that relevant sites could be investigated in more detail in a regions-of-interest analysis (see below).

A 3D lattice of points was constructed across the whole brain with 5-mm spacing, and beamformers were used to compute the total power at each point using the Neural Activity Index (NAI) [63]–an estimate of oscillatory power that takes account of spatially-inhomogeneous noise–at each point independently, within the following frequency pass-bands: 5–15 Hz, 15–25 Hz, 25–35 Hz and 35–50 Hz. Filtering was achieved with 4th order Butterworth filters with automatic padding to eliminate edge artefacts. These frequency ranges represent a subdivision of the frequency spectrum in step of 10Hz (or 15Hz in the case of the gamma band). The frequency bands roughly matched the frequencies of alpha, low and high beta and low gamma band although their purpose was to characterise strong sources of oscillatory power across the whole brain in general terms, to support the selection of point-of-interest to interrogate in the second step of analysis in which we could examine responses across the full range of frequencies in a fine-grained and continuous way. A similar approach was used in previous MEG studies of reading [72, 73], to describe the brain dynamics underlying lexical-semantic processing. We examined total power, which combines evoked (phase-locked to the stimulus) and induced (non-phase locked) components, in each frequency band, comparing an active period (0-500ms following stimulus onset) to a baseline passive period (from -550 to -50 ms before the stimulus was presented). For each individual participant and each frequency band, this analysis produced an NAI volumetric map for the active and passive period. A paired-samples t-statistic was used to characterise the difference between active and passive windows at each point in space in these maps. Individual participant's t-maps were transformed into standardized space and superimposed on the MNI template brain with the cerebellum removed using MRIcroN software [74] (see group t-maps in Fig 2 and S1 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. 3D rendered cortical representations showing significant activity above baseline across conditions, during 500 ms post-target onset in four frequency bands (15–25 Hz, 25–35 Hz and 35–50 Hz).

t-Maps are thresholded at p<0.01 (corrected). All the activations represent event related desynchronization. Significant event related desynchronization was only observed between 5–15 Hz in a region of the right fusiform gyrus that overlapped with the activity observed at 15-25Hz, at a reduced threshold (p = 0.05) and thus this frequency band was omitted from this figure, although all of the whole-brain maps can be accessed from Neurovault (http://neurovault.org/collections/1937/). Arrows indicate the locations selected for the VE analysis.

https://doi.org/10.1371/journal.pone.0169269.g002

In order to determine whether the difference between active and passive periods was statistically significant for each point on the lattice, we built up a null distribution by randomly relabelling the two time points for each participant and each voxel, using the permutation procedure developed by Holmes et al. [75]. We established the maximum t-value obtained with random relabelling across 10000 permutations. We then compared the real distribution of t-values in our data with the maximum t-value obtained from the permuted active and passive windows. Maximum statistics can be used to overcome the issue of multiple comparisons (i.e. controlling experiment-wise type I error), since the approach uses the highest permuted t value across the brain to provide a statistical threshold for the whole lattice of points, over which the null hypothesis can be rejected [69]. Fig 2 and S1 Fig show those voxels in the brain with t-values equal or higher than the top 5% or 1% t-values present in the null distribution. We repeated this analysis with all four conditions collapsed together, to characterise the general response of the brain to the task, and also (in a supplementary analysis) examined the main effects of specificity and category, with these conditions being compared with their own passive baseline.

Time-frequency analysis: points of Interest

Separate beamformers were used to reconstruct the neural activity for three points of interest (POI), in order to characterise the response of these regions to our experimental manipulation over time and frequency with greater precision. The MNI coordinates for these POIs were defined within these pre-specified regions using local peaks of maximum activation across all conditions in the group level, whole brain analysis.

(1) One POI was identified within lateral posterior fusiform gyrus (FG, MNI coordinate: -50, -70, -14), since this region is involved in visual object identification [21, 46]. Visual processes within this region should make a critical contribution to the recognition of both animals and man-made objects; however, we would anticipate greater engagement for animals than manmade objects at the specific level, since specific visual features are thought to play a greater role in distinguishing between animals with highly-overlapping visual features [17, 76]. Previous research has linked posterior fusiform to the visual discrimination needed to distinguish between different types of animals which have highly overlapping visual forms (i.e., four legs and a tail) [36, 77]. A similar site showed a greater response for animals than tools in the meta-analysis of Chouinard and Goodale [12]. We would anticipate that this site supports visual aspects of semantic processing.

(2) A second POI was selected within central sulcus (CS, MNI coordinate: -54, -22, 42). The motor and somatosensory hand regions to either side of this sulcus have been shown to be activated by tool concepts and their associated hand actions [10, 12, 21–23, 29]. Therefore, for this site, we would expect greater engagement for manipulable manmade objects than for animals if action features are an important component of our conceptual knowledge about tools. While our main analyses focus on putative spokes within visual and motor cortex, the literature on tool semantics suggests two additional sites that could also make a greater contribution to the identification of manipulable manmade objects than animals. First, left premotor cortex is associated with tool and action comprehension [22, 23, 29, 78], although this site is also likely to be influenced by the control demands of semantic tasks [79, 80]. We present results for this site in S2 Fig. Second, left inferior parietal cortex is associated with tool use and hand praxis [22, 78]. We do not present a POI analysis for this location because there was no clear response to the task within this region in the whole-brain beamforming results (see below).

(3) A POI within the anterior inferior temporal lobe (ATL, MNI coordinate: -51, 6, -39) was defined using coordinates taken from Binney and colleagues [81]. Atrophy in this region is linked to impaired semantic processing in SD patients, and ATL has been shown to be recruited by semantic tasks across categories in normal participants using distortion-corrected fMRI and transcranial magnetic stimulation [5, 37, 38, 81–83]. Within the whole-brain beamforming data, the ATL response fell within an area of significant activity in the group level analysis, although there was no clear local peak. In order to confirm that the pattern of results observed were not selective to this site, time frequency analysis was also performed on another region within the medial anterior temporal lobe taken from a recent MEG study of visual object recognition [45]. The results from this analysis are reported in S3 Fig.

We elected to examine left-hemisphere sites since (i) fMRI and patient studies reveal a greater contribution of the left hemisphere to semantic processing in general [84]; and (ii) given our participants were right-handed, the motor simulation elicited by single-handed tools was expected to be left-sided. Moreover, right motor cortex might have shown irrelevant responses related to the preparation of button presses with the left hand, even though button presses were only required on mismatching catch trials which were excluded from the analysis.

After the time-series of each POI was reconstructed epoch by epoch, for each subject, by means of separate beamformers [64], time-frequency plots showing total power were computed using Stockwell transforms [71] over a time window from -500 to 800 ms (to avoid edge effects) and a frequency range from 5–50 Hz (frequency resolution 1.33 Hz). The Stockwell transform, implemented in the NAF software, uses a variable window length for the analysis which is automatically adapted along the frequency range according to the sample rate and the trial length (4^th order Butterworth filters with automatic padding). The time-frequency representations were normalized, separately for each condition and for each participant, by dividing each time-frequency bin by the mean power per frequency bin in a baseline period prior to the start of trials in that condition (-250 to -50 ms). This window length was also used in earlier studies [51, 72, 73, 85], since it provides a compromise between the minimum length sufficient to estimate power at the lowest frequency we report (i.e., 5Hz) and the requirement to characterise the state of the brain immediately before the onset of each trial.

To compare the time frequency representations between experimental conditions, we computed generalized linear mixed models (GLMM) using PROC MIXED in SAS (SAS Institute Inc., North Carolina, US). Time-frequency plots of percentage signal change were treated as two dimensional arrays of small time-frequency tiles, indexed in the model by three main effects, each of which is defined as a class variable: time, frequency and the interaction between time and frequency. Therefore, random effects were included in each GLMM to account for the fact that each participant’s time-frequency plot is made up of multiple time-frequency tiles. We also controlled for time-frequency (or spatial) co-variance in the spectrogram by assuming the estimates of power followed a Gaussian distribution: consequently a Gaussian link function was used in the model. The time-frequency (spatial) variability was integrated into the model by specifying an exponential spatial correlation model for the model residuals [86]. In order to account for inhomogeneity in spatial covariance in the time-frequency spectrograms, we run separate GLMMs for three broad frequency band (6–15, 15–40 and 40–50 Hz); this procedure ensure an optimal Gaussian smoothing parameter for each model. The data were resampled at a frequency resolution of 2Hz and time resolution of 25 ms, the smallest time and frequency bin consistent with model convergence. This time-frequency resolution proved optimal in other similar published studies [51, 72, 73]. Finally, we compared every full GLMM, as outlined above, with its empty equivalent model to test overall model fit. To do this we checked that there was a statistically significant reduction in -2 residual log likelihood comparing the full and empty models, as well as a substantial reduction in the Akaike Information Criterion (AIC). Both of these criteria were fulfilled for every model fitted (see S2 Table).

PROC MIXED constructs an approximate t test to examine the null hypothesis that the LS-Mean for percentage signal change between conditions was equal to zero in each time-frequency tile, and the procedure automatically controls for multiple comparisons (i.e. controlling experiment-wise type I error). This method has been used in multiple peer-reviewed papers (for example [51, 72, 73]). The statistical contours on the percentage signal change figures encompass time-frequency tiles fulfilling both of the following criteria: a) the difference between conditions reached p < 0.05; b) any region in the time-frequency plot defined by (a) also showed a response that was significantly different from zero in at least one of the two contributing conditions.

Behavioural results

The results from the behavioural pilot are reported in S4 Fig. Participants were slower when making a superordinate as opposed to a specific level judgement, and also when categorising manmade objects. There was an interaction between these factors–the slowest responses occurred in the manmade superordinate condition, perhaps reflecting the featural diversity of manmade objects relative to animals.

During the MEG scanning, participants only responded in the case of a mismatch between picture and word (see Fig 1). The behavioural data confirm that participants maintained attention to the task. The overall accuracy was 84% (superordinate animal labels with manmade pictures = 82%; superordinate manmade labels with animal pictures = 88%; specific animal labels with mismatching animal pictures = 83% and specific manmade labels with mismatching manmade pictures = 83%). The percentage of false alarms was below 1% in all conditions.

Whole brain beamforming

All of the whole-brain maps generated by this stage of the analysis can be accessed from Neurovault (http://neurovault.org/collections/1937/). Changes in oscillatory power in response to the task were seen most clearly in the frequency bands 15–25 Hz, 25–35 Hz and 35–50 Hz. In all these frequency bands, a statistically significant (p = 0.01) reduction in total oscillatory power was observed when the task was compared to the passive baseline period, across a widely distributed set of cortical areas linked to semantic cognition and visual processing (see Fig 2). Event related desynchronization was also observed between 5–15 Hz in right posterior fusiform gyrus, within the response observed at 15-25Hz, at a reduced threshold (p = 0.05). Power reductions in similar frequency bands have been consistently reported in studies investigating language, memory and semantic processing [51, 72, 73, 85, 87, 88], alongside power increases at higher frequencies (high gamma, > 50Hz) and at lower frequencies (theta < 5Hz), which our methods are not well-suited to investigate. Since the response to the task reflected a reduction in total oscillatory power across all sites and conditions, a straightforward interpretation is that the task elicited neural activity that was not phase-locked to the onset of the stimulus and/or that was variable in phase across trials and participants. Event-related reductions in oscillatory power, relative to oscillations at rest (especially in in a mid-frequency range from 5–30 Hz), have been linked to event-related desynchronization [89]. This type of non-phase locked response, at a similar frequency to that observed in this study, has been shown to be correlated with task-related BOLD responses in fMRI [90, 91].

Brain regions responding across conditions included (i) the anterior temporal lobes bilaterally (with a peak in anterior STG/temporal pole), (ii) the entire length of the ventral visual stream bilaterally (reaching ventral ATL), (iii) left inferior frontal gyrus extending into premotor cortex, and (iv) bilateral intraparietal sulci; cortical areas that are all known to contribute to semantic cognition (Fig 2). In addition, we observed (v) activity in right motor cortex, consistent with motor preparation for left-hand button responses, plus a small response in left central sulcus close to the motor hand area, and (vi) extensive activity in right parietal cortex which might reflect visual attention to the complex stimuli we used (picture-word combinations). These changes in oscillatory power across conditions were used to identify the locations for the POI analysis. We chose two regions of response within the ventral visual stream—ventral ATL and posterior fusiform–since these regions are thought to be important for visual object identification [45, 46]. As noted above, we also placed POIs in left central sulcus and left premotor cortex, to examine the potential motor contribution to the task.

In order to provide a more detailed report of the dataset, whole brain beamforming was also used to examine the brain’s response to the main effects of specificity (i.e., for superordinate and specific trials, relative to their own passive periods prior to these trials) and category (i.e., for animals and manmade objects). The results of this analysis are reported in S1 Fig.

Points of Interest

For each POI (ATL, FG, CS) and for each participant, we computed time-frequency (TF) plots of total power for each condition. Fig 3A shows the data for superordinate and specific judgements (i.e., the main effect of specificity). Fig 3B shows the data for animals and manmade objects (i.e., the main effect of category). The responses for each condition individually are provided in S5, S6 and S7 Figs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Time-frequency plots for each cortical site (ATL, FG, and CS) are presented in each column.

(A) Main effect of specificity. (B) Main effect of category. In both (A) and (B), the first and second row report the percentage signal change in total power for each condition, relative to their passive periods. The third row shows differences between the two conditions. The black lines in the time-frequency plots indicate regions showing significant differences between the two conditions (p < .05). See text for details.

https://doi.org/10.1371/journal.pone.0169269.g003

Overall, for all three sites, we observed increases in power (yellow-red) relative to baseline between 50 and 150 ms (less pronounced for ATL), and decreases in power (cyan-blue) relative to baseline from 200 ms onwards. This is consistent with the possibility that power increases correspond to neural responses aligned to the presentation of the stimulus (characterising the brain’s response relatively early in time), while total power decreases correspond to neural activity not well-aligned to the presentation of the stimulus (potentially characterising later responses when differences between participants and trials have accumulated).

Another striking feature is the overall similarity at each site between the superordinate and specific conditions (in Fig 3A) and between the animal and manmade conditions (in Fig 3B). This suggests semantic processing arises from co-ordinated activity throughout the network, rather than, for example, particular nodes switching on and off discretely, at different points in time, or for different stimulus conditions. Nevertheless, the patterns of significant differences between conditions at each site (bottom row in Fig 3A and 3B) also suggest that, superimposed on this co-ordinated network activity, is a pattern of stimulus- and task-specific differences that arise as a result of varying the relative strength of the contributions from different nodes at different points in time.

For the main effect of specificity, the comparison of superordinate and specific conditions revealed stronger power reductions for specific judgements in the beta and low gamma frequency bands in all three POIs (Fig 3A). In ATL, task-related reductions in oscillatory power extended into higher frequencies (25–35 Hz) for specific judgements compared with superordinate judgements from 300–500 ms post-target onset. There was a similar effect of specificity in posterior fusiform cortex, with a stronger response for specific trials from 200–500 ms, as well as for superordinate judgements between 250–300 ms at 40–50 Hz. In the central sulcus, there was a stronger task-related power decrease in the specific condition from 15–25 Hz, extending in time from 200–400 ms, plus a stronger power increase for superordinate-level matching in the first 200 ms. Thus, effects of specific > general on the strength of task-related power decreases were striking from around 200 ms across sites.

With respect to the main effect of category, the comparison of animals and manmade objects revealed earlier and stronger power reductions for the manmade category in the central sulcus as predicted, from 200–600 ms in the beta and low gamma bands (Fig 3B). Stronger power changes in low gamma starting from around 250 ms were also observed in fusiform cortex and ATL for manmade objects compared with animals. Fig 4 shows the effect of category at the specific level (i.e., pigeon versus guitar). As noted in the Introduction, we would expect to see larger category effects in posterior fusiform cortex for specific-level trials, since the distinctive visual properties of animals are thought to be important for distinguishing between these items that generally share many features. Consistent with this prediction, posterior fusiform cortex showed greater task-related decreases in oscillatory power for the specific animal condition contrasted with the specific manmade condition (100–200 ms, 30–50 Hz). The reverse pattern was observed in central sulcus: the specific manmade condition showed a stronger response at this site (200 and 400 ms at 30 Hz). Consequently, across these visual and motor sites, differences between animals and manmade objects were in opposite directions, with stronger power reductions for animals compared with manmade objects in fusiform cortex, and stronger power reductions for the manmade than the animal condition in central sulcus. Finally, there were stronger power reductions in left ATL for tools than for animals from around 100 ms post-onset at 40 Hz, which persisted until around 550 ms (Fig 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Percentage signal change in total power for animal and manmade objects judgements at the specific level are reported in the first and the second row.

The third row shows differences between the two conditions. The black lines in the time-frequency plots indicate regions showing significant differences between the two conditions (p < .05). Each cortical site (ATL, FG, and CS) is presented in each column. See text for details.

https://doi.org/10.1371/journal.pone.0169269.g004

We also assessed the possibility of an interaction between specificity and category at each site (see S5, S6 and S7 Figs). For this analysis, we first examined the specificity effect for each category separately and then compared these effects (i.e. a difference of differences) in order to establish if there was a larger effect of specificity for one category compared with another. In ATL, there were specificity effects for both categories (greater reductions in total oscillatory power for specific identification), and this effect was stronger for manmade objects briefly at around 500 ms and 30 Hz. Left posterior fusiform cortex (visual site) also showed an effect of specificity for both animals and manmade objects: again, this effect resulted in greater decreases in total oscillatory power, and this response was stronger for animals in the first 200 ms above 40 Hz, consistent with the prediction that visual processes are particularly important in distinguishing between animals which have highly-overlapping visual features at the specific level. The left posterior fusiform POI also showed a stronger contribution to the identification of manmade objects at general level, as shown by the ‘reverse specificity effect’ for manmade objects observed in Figs 3A and 4 and S6 Fig (this time reflected in greater power increases relative to baseline for general judgements). This effect could conceivably reflect the greater visual diversity of manmade objects, increasing the contribution of the visual system to their identification at a general level. The central sulcus showed a complex pattern of response: although this site showed a main effect of category (manmade > animal concepts) and a main effect of specificity (specific > superordinate), it also showed a ‘reverse specificity effect’, with a stronger response in the superordinate compared to the specific condition for the animal category at 35 Hz and for the manmade category at 20 Hz around 500 ms post-onset. This elicited a significant interaction.