Introduction

Face recognition is a crucial and highly sophisticated ability in humans. The development of this skill has been argued to occur gradually over childhood, with increases in recognition of face identity and emotion improving through adolescence (e.g., Kolb et al. 1992; Chung and Thomson 1995; Herba et al. 2006). Face processing, in fact, continues to evolve with age, with older adults performing face processing tasks using different strategies than younger adults (Chaby et al. 2011; Gunning-Dixon et al. 2003). Skill in face recognition has been attributed in part to acquiring expertise in processing the configural structure of faces (e.g., Baenninger 1994; Mondloch et al. 2006; Maurer et al. 2002), often assessed with the use of inverted faces, as face inversion disrupts configural more than featural processing of faces (Yin 1969; Rhodes et al. 1993), in children as well as adults (Pascalis et al. 2001; Tanaka et al. 1998; Mondloch et al. 2002).

Some recent investigations have argued, however, that face recognition matures early (e.g., Crookes and McKone 2009; Picozzi et al. 2009), thereby implying that neural processing of faces is qualitatively the same from five years of age; age-related effects may be due to maturation of other aspects of cognitive or perceptual processing (see Mondloch et al. 2006 for discussion), in particular the efficiency of the processing has been suggested to improve with age. There are few brain imaging studies across the period of early through late childhood that have investigated the mnemonic aspects of face processing; thus there is little neuroimaging data to apply to the discussion of the underlying neural development. Measures of brain processes during face recognition that provide spatial and temporal indices would contribute to our understanding of which aspects of brain function mature in the acquisition of this skill over childhood.

Specialised, distributed brain networks have been described for processing faces (e.g., Allison et al. 2000; Haxby et al. 2000), with the brain regions implicated varying with the information retrieved from the faces (e.g., Ishai 2008; Gobbini and Haxby 2006; Platek and Kemp 2009). While considerable research has focussed on basic face processing with various neuroimaging modalities, fewer investigations have examined the neural correlates of face recognition. fMRI studies of face recognition in adults have shown activation in the classic face-processing ventral stream, as well as the anterior cingulate, the inferior frontal gyri and insula (e.g., Ishai and Yago 2006; Dolcos et al. 2007), overlapping regions in the frontal lobes that are found in non-verbal n-back recognition tasks such as the dorsolateral prefrontal and dorsal cingulate regions (for reviews, see Owen et al. 2005; Nyberg et al. 2003). These sources are reliably identified with fMRI, but the timing of the activation remains uncertain. As neural timing or efficiency is a critical aspect of cognitive functions that mature over childhood, understanding the development of neural structures underlying recognition memory requires information on the timing of brain processing.

Several neuroimaging studies have demonstrated changes in the neural correlates of frontal lobe memory function over childhood. Brahmbhatt et al. (2007) found fMRI activation differences between a large series of adolescents and adults in n-back tasks, in the absence of performance differences. In a smaller study with a larger age range, Scherf et al. (2006) showed that younger children relied on more medial frontal regions in a working memory task and, consistent with Brahmbhatt et al. (2007), that significant differences in brain function were found between adolescents and adults. Using a more difficult task, Kwon et al. (2002) found both performance and brain activation differences across childhood and adolescence in frontal and parietal regions. They reported a right hemisphere system involved with visuo-spatial memory and attention and a left hemisphere system implicated in verbal rehearsal and memory; the two neural networks matured in parallel. Ciesielski et al. (2006) found little frontal involvement in 6-year old children in a challenging n-back task, with the emergence of frontal activity in the ten-year-old group and continued development between the 10-year olds and adults. Electrophysiological studies of recognition contribute timing information to the sequence of brain processes. Using an n-back task with faces, Itier and Taylor (2004a, b) showed effects of face repetition and recognition on the early ERP components as well as the more typical frontal late component (250–600 ms) that was most sensitive to memory. Significant age effects were seen across the age range from eight-year-olds to adults, with the frontal component not maturing until mid-adolescence. Configural changes induced by face inversion produced little effect on the frontal component related to recognition (Itier and Taylor 2004a) suggesting a similar working memory system was involved across age groups.

Other key brain regions for mnemonic functions are the hippocampi. The hippocampal structures have historically been associated with long-term memory retrieval but are now increasingly implicated in short term memory functions as well. Numerous neuroimaging studies have demonstrated the involvement of the hippocampi in a range of memory-related processes, including memory encoding (Davachi and Wagner 2002) and working memory (Ranganath and D’Esposito 2001). Unlike the frontal lobes which show very protracted maturation through to adulthood (Yakolev and Lecours 1967; Shaw et al. 2008; Giedd et al. 2009), development of hippocampi show comparatively less change with age. The size of the hippocampi do not increase significantly over childhood (Ostby et al. 2009), although there are regional and asymmetrical maturational differences, such as an increase in the posterior aspects related to memory function (Gogtay et al. 2006). Finn et al. (2010) demonstrated the interaction of prefrontal cortex and hippocampal regions over adolescence and young adulthood in working memory. Both areas were active in young adolescents; prefrontal activation associated with performance was seen only in late adolescence, while the hippocampal-prefrontal co-activation was seen only in the older adolescents and adults with higher task demands. The right hippocampus was activated in the younger group regardless of load, while the left hippocampus was active during the more difficult task in the older groups. This elegant study highlighted the importance of measuring both cortical and subcortical brain activation to understand the development of working memory; they did not however, include young children nor have accurate measures of the timing of this activity.

Thus, the frontal cortices and the hippocampi are critical to memory function and undergo quite different patterns of age-related maturation; investigations that have assessed these brain regions in memory tasks across childhood are lacking. Considerable evidence has demonstrated the ability of MEG to detect signals from deep brain structures such as the hippocampi during the processes of memory encoding (Breier et al. 1999; Riggs et al. 2009) and working memory (Tesche and Karhu 2000), and to provide timing information of hippocampal function in these memory processes. In both MEG and ERP studies, frontal activation related to mnemonic functions appear at latencies typically above 240 ms (see Rugg and Curran 2007, for a review), while hippocampal activity in memory tasks is often reported around 400 ms (e.g., Barbeau et al. 2008). Earlier activation in the hippocampi have been reported (i.e., Hung et al. 2010) and may depend on the paradigm and the sensitivity of newly applied analysis techniques; we have demonstrated some of the parameters necessary for reliable detection of hippocampal activation with MEG (Quraan et al. 2011).

A few MEG studies have investigated face recognition (Itier et al. 2006; Harris and Nakayama 2008; Schweinberger et al. 2007), largely focussing on early effects on the posterior components. There has been little investigation on the neural components involved in the later effects of face recognition. We recently showed, however, that the immediate recognition of repeated upright faces activates right frontal and right insula sources at 240–250 ms in adults, with bilateral activation seen to recognised inverted faces (Taylor et al. 2011). We found that the contrast between the upright and inverted face recognition showed an anterior cingulate source; Schweinberger et al. (2007) also reported activation in the area of the anterior cingulate to repeated faces.

Despite the ability of MEG to provide both excellent temporal and spatial resolution, no MEG studies have explored the development of face recognition memory. Few developmental MEG investigations have included faces, and the studies that have been reported have not always seen components comparable to the ERP literature (Kylliäinen et al. 2006a, b; Kimura et al. 2004; Taylor et al. 2010). We examined face recognition in children through to young adults to further elucidate the maturation of this ability, with a focus on the frontal and hippocampal brain regions using MEG to localise the activations both spatially and temporally. We expected recognition-sensitive activity in the right hippocampus (Finn et al. 2010; Hung et al. 2010) and in the right frontal regions (Taylor et al. 2011).

Methods

Subjects

Ninety subjects participated in the study, divided into the following age groups: 6–7 years (n = 10, mean 6.8 years), 8–9 years (n = 11, mean 9.1 years), 10–11 years (n = 10, mean 11.0 years), 12–13 years (n = 10, mean = 12.9 years), 14–15 years (n = 9, mean 15.2 years), 16–17 years (n = 13, mean 17.0 years), 18–19 years (n = 7, mean 19.3 years) and young adults (n = 20, mean 25.3 years). A further 20 participants (17 children, three adults) had been run on the paradigm, but post-hoc analyses showed excessive ferromagnetic contamination despite screening and exclusion for metallic implants or braces. The male:female ratio was 43:47, thus almost evenly distributed across age groups, but with variability by age group. Within each group the Ns were too small and variable to analyse for sex differences. Likewise, handedness was not analysed, as only 12% of the participants (spread across the groups) were left handed. None of the included participants had a history of behavioural, psychiatric or neurological disorders and all had normal or corrected-to-normal vision. The children were in the age-appropriate grade at school. All subjects gave informed consent; the young children provided informed assent, while their parents signed the written consent. The study was approved by the Research Ethics Board at The Hospital for Sick Children.

Stimuli

Greyscale photographs of 240 different faces (120 male) were used; faces had no earrings, glasses or facial hair. Two blocks of 180 faces were presented. In each block, one-third of the faces repeated immediately in a 1-back task (i.e., 60 faces never repeated, 60 faces would repeat after a first presentation). One block contained upright faces and the other block of 180 contained inverted faces; the two blocks were presented in counter-balanced order across participants. Faces presented in one block did not ever appear in the other block. The use of inverted faces was to increase difficulty while still allowing the tasks to be completed by young children.

Procedure

MEG data were recorded on a CTF/MISL Omega 151 channel system. Data were recorded continuously at 625 Hz with an on-line bandpass of 0–200 Hz, and filtered off-line to 0.5–30 Hz. Subjects lay supine with their head in the dewar of the MEG, inside the dimly lit magnetically shielded room. Prior to data acquisition, three fiducial coils were placed at the nasion and pre-auricular points to localize the subject’s head relative to the MEG sensors. The system allows head localisation to an accuracy of 1 mm. Head localisations were completed before and after each block to confirm that participants had not moved; tolerance was <5 mm in adults and older children, and 10 mm in the youngest children.

Stimuli were back-projected onto a screen in front of the subjects by a video projector situated outside the room, using Presentation software (Neurobehavioral Systems, Albany, CA). Faces were presented on a black background for 500 ms with an inter-stimulus interval randomly varied between 1200 and 1500 ms; a white fixation cross was presented in this interval. Subjects pressed a button as quickly and accurately as possible when they recognized that a face had repeated and another button when the face was new—i.e., when they did not recognise the face.

To ensure accurate MEG-MRI co-registration, after completion of the MEG, the fiducial coils were replaced with MRI contrast markers in the same locations and an anatomical MRI was acquired for all subjects on a 1.5 Tesla Signa GE Advantage scanner, with a T1 3D-SPGR-pulse sequence.

MEG Data Analyses

Sensor Analyses

For both the upright and inverted conditions there were 60 repeat trials (R) and 120 non-repeat trials (N). These non-repeat trials were divided into N1 (never repeated) and N2 (would subsequently repeat), both having 60 trials. For each subject, all trials at the sensor level were averaged by condition, and differences computed for R-N2, the recognition effect, and N1-N2, the control. The conditions were subtracted from each other to remove common mode visual activations, which are large and can obscure the smaller activations related to recognition. It is prudent to remove these dominant visual activations at the sensor level before the source localisation is performed (Quraan et al. 2011) as the interaction between dominant sensory functions and smaller, cognitive functions is known to cause correlation and leakage problems when source localizing using an adaptive beamformer (Quraan and Cheyne 2010; Van Veen et al. 1997; Reddy et al. 1987) or the linearly constrained minimum variance (LCMV) vector beamformer (Sekihara et al. 2001; Robinson and Vrba 1999) used in these analyses. Grand averages for each age group were then calculated for the R-N2 (recognition) and N1-N2 (control) contrasts.

Principal Components Analysis was used on these contrasts to reduce the data in this multi-factored study, before source analyses were performed. For each group the recognition (R-N2) condition grand average was subjected to singular value decomposition. Only the first component was kept per group. In ascending order (6–7 years, 8–9 years,… adult) for the recognition effect in upright faces, the percent variance explained by the first component was: 0.31, 0.57, 0.44, 0.46, 0.67, 0.71, 0.62 and 0.65. In the same age order, for the recognition effect in inverted faces, the percent variance explained by the first component was: 0.45, 0.27, 0.46, 0.38, 0.52, 0.43, 0.35 and 0.43. This first component was then multiplied back by the original data to yield a time-course showing where in time this component was expressed. The first component was also multiplied by the control (N1-N2) data to yield a time-course as a null comparison.

Source Analyses

Multisphere head models fit to the inner skull of each participant’s MRI were used to model the forward solution to the beamformer. LCMV vector beamformer was used to estimate source activity. The window 280-680 ms was chosen visually from the first component time-course across all groups, as this encompassed the first PCA component (see Fig. 1). The source images were then created for all conditions by averaging the Pseudo-Z source activation within this window. Each subject’s source images were warped to Talairach space using SPM2 and then put back into their grand average source images. The maximum value from the control (N1-N2) grand average was used as a threshold for the recognition (R-N2) grand averages.

Fig. 1
figure 1

Right Topographic plot of 1st principle component for each age group for upright face stimuli. Left Time course of 1st principle component after multiplication with original data (red line) compared to multiplication with control (N1-N2) (blue line) for upright faces

Full size image

Spatial Overlap

Spatial consistencies across groups were considered by overlapping the activation maps for the grand averages. For each group image, any voxels surviving threshold were given a value ‘1’ and all others a ‘0’. Thus, voxels with value ‘8’ indicated that all eight age groups activated in that voxel; each value was assigned a different colour code. The final summed image demonstrates where, and how many, groups have common spatial activations. For this paper, spatial locations with more than five participating groups were pursued.

Virtual Sensors

Estimated time courses, or virtual sensors, were created for all peak locations which passed a threshold of five on the overlap described above (i.e., activation was spatially consistent in at least five age groups). These sensors are estimates of the true activation at the identified location and are in units of Pseudo-Z. To reduce the data, as each location led to a plot of eight groups with their various time-courses, we averaged the activation for each subject across the activation window, 280–680 ms. This gives one number per subject per condition which, when regrouped gives a measure of summed activation in the time window for each location by age.

Correlations

Lastly, each subject’s value from the virtual sensors was plotted against their accuracy for the repeat condition to assess whether signal strength predicted recognition accuracy.

Results

Behavioural Data

Table 1a shows reaction time data for each age group and for each condition. An 8 × 2 × 2 repeated measures ANOVA was computed for age group × (upright vs. inverted) × (repeat vs. control). This showed a significant main effect for age [F(7.82) = 2.90; P < 0.009] and a two way interaction for age × repetition [F(7.81) = 2.97; P < 0079]. This interaction was due to the finding that the youngest age group had the longest reaction time for repeated versus first presentation; regardless of upright or inverted presentation; the older groups did not differ.

Table 1 Behavioural result
Full size table

Table 1b shows accuracy data for each age group and for each condition. An 8 × 2 × 2 repeated measures ANOVA for age group × (upright vs. inverted) × (repeat vs. control) showed significant main effects for upright versus inverted (F(1.82) = 52.43; P < 0.000) and repeat versus control (F(1.82) = 6.53; P < 0.0124). Both inverted and repeat conditions showed significantly lower accuracy. There was also a significant two way interaction of age × (upright vs. inverted) with the 14–15 year old group showing poorer accuracy on the inverted condition, an ‘adolescent dip’ (e.g., Chung and Thomson 1995).

MEG Data

Principal Components Analysis of Sensor Data

Figure 1 shows a topographic plot by age of the 1st principle component (PC) itself, and then the time course of the 1st PC multiplied by the original sensor data for the upright condition on the repeat (R-N2) versus control presentations. The repetition effect is clearly expressed as a slow wave component between 280 and 680 ms across all age groups, although it is not as clear in the youngest group. This component failed to be expressed in the N1-N2 contrast, as expected, as this was the control. The same pattern of results was seen for the inverted condition, with the 1st PC being expressed in the 280–680 ms window for the older five age groups; the component was not as clearly expressed in the youngest three groups.

Grouped Source Images

To identify source localisations that were consistent across age groups for each condition, thresholded data were grouped by age and overlaid onto a glass brain. Figure 2 shows the spatially overlapping components activated by repetition for the upright (left panel) and inverted (right panel) conditions. For the upright condition, all eight age groups showed activation in right hippocampus and right parietal area, while six age groups showed activation in right inferior frontal gyri. For inverted faces, six of eight age groups showed activations in left parietal while five showed activations in right hippocampus and right inferior frontal region. Table 2 shows the pattern of activations by age group and location.

Fig. 2
figure 2

Number of age groups showing spatially overlapping components activated by repetition for the upright (left panel) and inverted (right panel) face stimuli. For the upright faces, eight groups (black) showed activation in right hippocampus and right parietal areas, while six groups (red) showed activation in right inferior frontal gyrus

Full size image
Table 2 Groups with significant activations
Full size table

To further examine these effects, virtual sensors at the four consistent locations, left parietal, right parietal, right inferior frontal and right hippocampus, were created and summed across the window 280–680 ms for each age group and condition. Figure 3 plots the summed virtual sensors for each age group for the upright (left panel) and inverted (right panel) conditions. There is a clear repetition effect (red line) compared to control (blue line) in the right hippocampus for all age groups for upright faces, as was suggested in Fig. 2. The right and left parietal sources showed a similar pattern across the age range, with the repeated faces producing greater activation than the control faces. For the inferior frontal source, the repetition effect begins to separate from control activation at 10–11 years of age for the upright faces, becoming clearly distinct at 14–15 years. For the inverted faces (right panel) the emergence of a repetition effect was seen only in mid childhood, consistent with the inverted faces being a more difficult task, particularly for the young children.

Fig. 3
figure 3

Summed activations, in the 280–680 time window, averaged within age groups for the repetition (red line) versus control (blue line) conditions, for upright (left panel) and inverted (right panel) face stimuli, at the four most consistent brain locations. Inset is an image from the 10–11 year old group, showing averaged activation in the two brain areas of focus for this paper, the right hippocampus (Talairach coordinates: 39.8-17.9-7.5) and right inferior frontal gyrus (Talairach coordinates: 39.8-23.1-12.7)

Full size image

Correlational Analyses

The summed virtual sensor activations for each individual from each of the four consistent locations were correlated with task accuracy. The only significant correlations were for the right inferior frontal virtual sensor × accuracy for both upright (R = 0.277; P = 0.008) and inverted (R = 0.4546; P = 0.0001) conditions, and the right hippocampus virtual sensor × accuracy for upright (R = 0.2652; P = 0.011) and a trend for inverted (P = 0.08) conditions. Figure 4 plots the strongest of these correlations: right inferior frontal virtual sensor × accuracy for the inverted condition.

Fig. 4
figure 4

Individual summed virtual sensor activations from right inferior frontal gyrus plotted against performance accuracy for the inverted condition, showing the significant positive correlation between frontal activation and accuracy

Full size image

Discussion

In this large developmental series, including children from six through 19 years of age and adults, we showed distinct maturational profiles in two key brain regions implicated in mnemonic function. The right hippocampus showed increased activation to repeated faces in the 1-back tasks, across the age range, whereas the right frontal activation increased over mid-childhood through to adulthood; there was no difference in the youngest three age groups in the frontal activity between new and repeated faces. As we were interested in the neural correlates of recognition memory, we subtracted out the dominant sensory responses and thus, did not investigate the early visual processes related to face processing per se; this was described in an earlier paper with a somewhat smaller sample (Taylor et al. 2010).

The frontal activity seen in the present series was only reliable in the right hemisphere, consistent with other studies that report right-sided processing for faces (e.g., Bayle and Taylor 2010; Jackson et al. 2008) or visual spatial stimuli in general (Kwon et al. 2002). Spatial working memory tasks have been associated with the right inferior frontal regions (e.g., Clark et al. 2007; Chase et al. 2008) and nonverbal working memory tasks also activate right frontal areas preferentially (for a review see Owen et al. 2005). As unknown faces are nonverbal stimuli and faces contain spatial information, this asymmetry is consistent with these prior studies. Some fMRI studies with face stimuli report bilateral frontal activation (e.g., Ishai and Yago 2006) or bilateral with larger effects in the right frontal regions (Dolcos et al. 2007), suggesting that this asymmetry is to some extent task and/or imaging modality dependent. The fact that this asymmetry was the same across the age range once it had emerged reliably at 10–11 years of age, suggests that the underlying processing is qualitatively similar from this age on. This is consistent with the conclusions in some studies both from fMRI and ERP (e.g., Itier and Taylor 2004a; Kwon et al. 2002; Ciesielski et al. 2006), that found significant but quantitative changes between late childhood and adults. However, others have reported that frontal activation shifts with age (e.g., Scherf et al. 2006), although this fMRI study had children only as young as 10 years of age and a sample size of nine in the 10–13 age group. Furthermore, they used a visual search task where subjects were instructed to remember the direction of a target; probably, this involved verbal encoding. Kwon et al. (2002) reported a left lateralisation when verbal rehearsal and memory were involved. The effects were similar in pattern but weaker for the inverted faces, showing an apparent later emergence of the repetition effect with inversion, consistent with greater difficulty of inverted face processing. The presence of a comparable frontal pattern, however, regardless of configuration is concordant with earlier work (Itier and Taylor 2004a) that found configural effects on the early processing but not on the later frontal slow component.

In contrast to the frontal repetition effect, which emerged in mid-childhood, the right hippocampal activity to repeated faces was reliably seen across all eight age groups. This suggests that the involvement of the right hippocampus in memory function is established early in life, consistent with the early maturation of the hippocampi. The hippocampal structures undergo the most maturational change in the first year of life (Insausti et al. 2010) with only slight increases in size over childhood (Otsby et al. 2009). The current functional neuroimaging data do not have the sensitivity unfortunately, to reflect the more subtle changes in subregions of the hippocampi over childhood (Gogtay et al. 2006), that may correlate with memory functions. Consistent with these findings, the hippocampal-frontal lobe networks undergo significant myelination in late adolescence (Benes 1989), which may well interact with the developmental trajectories of regional change within the hippocampi.

Our results build upon the current knowledge that the frontal cortices and hippocampi are critical to memory function. Our findings that these two key regions undergo different functional patterns of maturation suggest that children and adults employ different strategies when completing these tasks. Specifically, the early maturation and involvement of the right hippocampus suggests that it subsumes basic memory processing, but more complex or complicated memory tasks will require the additional resources offered by the frontal system. This idea is further supported by our finding of a correlation between MEG signal strength in the right inferior gyrus and accuracy.

In conclusion, this investigation of components immediately following the visual responses in a face recognition task in a large series of children and adults, allowed us to identify two key brain regions implicated in mnemonic function and track their developmental profile. Overall, our findings demonstrate the value of using MEG to elucidate the neural processes and their developmental trajectories; despite essentially equivalent performance, children and adults utilize brain regions differently to achieve face recognition.