The authors have declared that no competing interests exist.
Conceived and designed the experiments: GAH AJB. Performed the experiments: GAH AJB. Analyzed the data: GAH AJB SJ CFB. Contributed reagents/materials/analysis tools: SJ. Wrote the paper: GAH AJB SJ CFB.
Age is one of the most salient aspects in faces and of fundamental cognitive and social relevance. Although face processing has been studied extensively, brain regions responsive to age have yet to be localized. Using evocative face morphs and fMRI, we segregate two areas extending beyond the previously established face-sensitive core network, centered on the inferior temporal sulci and angular gyri bilaterally, both of which process changes of facial age. By means of probabilistic tractography, we compare their patterns of functional activation and structural connectivity. The ventral portion of Wernicke's understudied perpendicular association fasciculus is shown to interconnect the two areas, and activation within these clusters is related to the probability of fiber connectivity between them. In addition, post-hoc age-rating competence is found to be associated with high response magnitudes in the left angular gyrus. Our results provide the first evidence that facial age has a distinct representation pattern in the posterior human brain. We propose that particular face-sensitive nodes interact with additional object-unselective quantification modules to obtain individual estimates of facial age. This brain network processing the age of faces differs from the cortical areas that have previously been linked to less developmental but instantly changeable face aspects. Our probabilistic method of associating activations with connectivity patterns reveals an exemplary link that can be used to further study, assess and quantify structure-function relationships.
Processing the age of faces is a crucial cognitive skill. Facial age varies but not in a volatile manner like eye gaze, lip movements or facial expressions. Contrary to other fixed and variant cues
Developmentally, age is less salient in faces than gender and ethnicity
Here, we generate continuous morphs that introduce independent age and gender changes of face stimuli. These changes are virtual but based on a fully morphable 3D model, similar to
(A) Exemplary keyframes of a video sequence (see Movie S1) morphing a 20 year-old female into a 60 year-old male. Both gradual age and gender changes are illustrated at intervals of 1 second. (B) Line magnitude images of optical flow velocities computed by the Horn-Schunck algorithm. Differential motion/optical flow was quantified as an overall parameter by the sum of flow magnitudes between successive keyframes. (C) Motion-/flow-related activations of hMT+ derived from the group-level analyses (n = 24 subjects, FWER-corrected p<0.05, [−log10 (p)] colorbar) on posterior cortical flat maps of both hemispheres. Additionally, ventral (v) and dorsal (d) visuotopic labels (V1–8, Vp, LO, hMT+) of the SuMS database, transformed from Caret's PALS atlas into FreeSurfer's average surface space, are displayed. Note that according to recent data
We use functional (fMRI) and diffusion-weighted (DWI) magnetic resonance imaging to investigate the following:
We obtained blood oxygenation level dependent (BOLD) fMRI and DWI scans of 24 healthy, right-handed, white Caucasian volunteers (age range 23 to 34, mean age 26, standard deviation 3 years; 12 females), who all gave written informed consent. The Ethics Committee of the University of Würzburg (Faculty of Medicine) approved the study. Handedness was assessed by a variant of the Edinburgh Handedness Inventory
For validation of the tractography results, diffusion-weighted imaging (DWI) data of additional 46 right-handed healthy volunteers (age range 19 to 63, mean age 30, standard deviation 9 years; 25 females) from the database of the Oxford Centre for Functional MRI of the Brain (FMRIB) were analyzed.
Full-front photographs of 121 unfamiliar, unambiguously gendered faces of white Caucasians (age range 2 to 81, mean age 33, standard deviation 15 years; 60 females age-matched to the males; all free of any make-up and beardless, with eye gaze directed at the viewer, wearing no jewellery or piercings, without tattoos; rated as neutral in their expression on a 6-point visual analogue scale by all participants) were matched by a computerized algorithm
Morphing transitions between two faces lasted 6 seconds (see Movie S1), with an additional 1-second still in-between. Face morphs with and without gender transitions (n = 60 each) were arranged in random order, with all morphs across gender proceeding from full-male/-female to the opposite.
Facial age of start and target stimuli was continuously modulated, except for the one-second stills, with varying degrees and age changes being pseudo-randomized according to 10-year intervals. Age changes during morphing were independent of (i.e. orthogonal to) the transitions of gender (Pearson's normalized correlation between the modelled regressors: r = 0.046, p = 0.39) and the psychometric ratings of attractiveness/likeability, which was not significantly modulated (i.e. did not exceed one rating-point difference for all start/target image pairs on a 6-point visual analogue scale). Thereby, morphing eliminates the need for an explicit baseline but nevertheless allows us to separate the effects of age and gender.
Psychophysical changes of age and gender were parametrically modeled according to Steven's power law (cf.
(A) Facial age difference ratings (magnitude of age-gradients spanned by morphing rated on a 6-point visual analogue scale, maximum scaled to 3.0 arbitrary units [a.u.]) followed Stevens' (∧0.3) better than Weber-Fechner's law (log10) or a natural logarithmic transformation (ln) of start and target age. All face stimuli (n = 121) were morphed to an average-aged male face of 33 years, the morphing sequence was randomly played forwards or backwards for the rating (circles with error bars; n = 24 subjects). (B) Facial aging (x-axis; objective age in [years]) increased the variability of subjective age ratings (y-axis; SD, standard deviation of estimated age in [years] across n = 24 subjects). Rating accuracy of factual (n = 121 stimuli of real faces) and interpolated age (n = 80 intermediate face stimuli from the morphing algorithm; one randomly selected for each annual increment between 2 and 81 years of age) did not differ significantly (p = 0.97). (C) Face gender ratings (on a 6-point visual analogue scale, maximum scaled to 1.0 arbitrary units [a.u.]) along temporal morph continua (n = 60) across faces of clearly different sex. Subjective ratings by (n = 24) subjects (boxes with error bars, blue line) were augmented above linear transition values (dashed line with black dots), reflecting the tendency to apperceptive gender categorization.
Both were time-binned at the video frame rate (24 fps) and scaled to maxima of 1.0 arbitrary units [a.u.]. (A) Differential age change encoded according to Stevens' law of psychophysics (using a power exponent of 0.3;
The morphing video was presented at 24 frames per second (fps) using a fMRI-compatible LCD screen, scaled to the maximum resolution of the presentation equipment (640×480 pixels VGA). The paradigm contained the entire set of 120 face morphs between different pairs of start and target faces, separated by 1-second stills, and was presented to the subjects in a single run lasting 14 minutes. Thus, the inter-stimulus interval (ISI) between the offset of a 6-seconds morph and the onset of the next was 1 second, resulting in a stimulus onset asynchrony (SOA) or inter-trial interval (ITI) of 7 seconds with no stimulus repetitions. In order to sustain attention and to monitor compliance, subjects were instructed to press a key with their right index finger whenever the target face appearance of the morphing sequence was anticipated. Speed and accuracy were not emphasized. The explicit task was of no particular interest and only ensured that the participants attended to the paradigm. The associated activations are not reported because they are contaminated by co-activations of executive and motor functioning from key pressing. While watching the face morph of Movie S1, for example, the subject realizes that a young female is changed into an older male, and the characteristics of the target face can be anticipated before it finally appears. This may coincide with the detection of a subjective identity transition although identity virtually changes continuously in a slow but permanent manner during all morphing episodes. Note that age and gender changes were continuously modeled over the entire morphing period while subjects generally pressed the key once in anticipation of the target face, i.e. explicit task performance was sufficiently independent of implicit age and gender processing as we have modeled it. Key-press responses were recorded by a fMRI-compatible keyboard and logged by Cogent 2000 v125 (2003, Wellcome Laboratory of Neurobiology, London,
We acquired fMRI time-series and T1-weighted anatomical images of the (n = 24) healthy volunteers in one session, and whole-brain DWI data and explicit behavioral post-hoc ratings of the same set of subjects in a second session within two weeks. In order to assess the relation between fMRI activations and age rating competence at the second level, age-rating performance was used to discriminate most accurate (n = 5) from average (n = 14) age-raters.
Functional and T1-weighted structural MRI data were acquired on a 3 Tesla TimTrio scanner (Siemens, Erlangen, Germany) using a 12-channel head coil. Whole-brain T2*-weighted BOLD images were recorded by a single-shot 2D gradient-echo EPI sequence with interleaved slice acquisition (TR = 2400 ms, TE = 30 ms, resolution 3×3×4 mm3, including 25% interslice gap, 30 sagittal slices of 3.2 mm thickness). After discarding the four initial scans, 350 volumes acquired during visual paradigm presentation were analyzed. In order to unwarp geometric distortions of BOLD EPIs, we used gradient-echo fieldmaps (TR = 500 ms, TE1 = 4.30 ms, ΔTE1/2 = 2.46 ms). In addition, a T1-weighted 3D anatomical image using a MPRAGE sequence (TR = 1560 ms, TE = 2.26 ms, resolution 1×1×1 mm3) optimized for segmentation and surface reconstructions and, for basic screening, a T2-weighted 2D axial FLAIR sequence were acquired.
In order to avoid potential DWI signal-loss artefacts
DWI data of the independent database were acquired on a 1.5 Tesla Sonata scanner (Siemens, Erlangen, Germany) with similar sequence parameters at slightly lower slice thickness (resolution2×2×2 mm3, 72 axial slices). Three sets of DWI data were recorded for subsequent averaging to improve the signal-to-noise ratio (total scan time 45 minutes).
All MRI data were processed using FSL 4.1 (
In order to take advantage of surface-based registrations and statistical analyses, FreeSurfer was used for segmentation and surface reconstructions of the structural T1-weighted MRIs. Employing boundary-based registration (using bbregister, part of FreeSurfer,
Relative response magnitudes were quantified based on individual mean within-cluster contrast-of-parameter estimates (COPEs) normalized to the respective minimum, see
(A) Group-level (n = 24) functional activations1 related to age and gender change, respectively. (B) Quantification and between-cluster/-hemisphere comparisons of observed effect sizes evoked by facial age and gender changes across (n = 24) subjects. Individual values of each cluster's mean activation (± error bars across subjects) were normalized to the lowest average of corresponding response magnitudes (as extracted from the first-level analyses). (C) Increased age-related activations1 of the most accurate (n = 5) above average age-raters (n = 14). The corresponding cortical flat map is outlined by the borders of the left age-responsive pANG cluster. (D) Relative to average post-hoc raters (avg, n = 14), high explicit age-rating accuracy (upper quintile P80, n = 5) was accompanied by almost five times the response magnitude during implicit age-change processing within left pANG (p<0.001, based on mean individual activation levels of the sub-cluster shown in
DWI data were processed using FMRIB's Diffusion Toolbox (FDT, part of FSL). Up to two fiber orientations were modeled and the corresponding probabilistic distributions of diffusion parameters were built up at each voxel (using bedpostx, part of FDT). Probabilistic modeling of multiple fiber orientations
Ventral portion of Wernicke's perpendicular fasciculus (WpF) connecting pANG and pITS (average probabilistic path distributions connecting the functional clusters; n = 24, 3D-tract volume rendering thresholded at ≥100 connecting samples passing through each voxel, displayed on sagittal [x = −36 mm] and coronal [y = −54 mm] projection view planes in MNI standard space). pANG, posterior angular gyrus area; pITS, posterior inferior temporal sulcus; orientation labels: L, left; R, right.
Utilizing spatial cross-correlations between functional activation probability values and tractography-based connectivity scores, we examine if the activation pattern of one area that processes facial age is predicted by its intrinsic structural connectivity with another, i.e. evidence for two selected areas directly interacting with each other as connections to the latter determine activations of the former and vice versa. The rationale behind this analysis was that if the spatial profile of a connection between A and B predicts the activation profile in A, then this suggests that the connection between A and B is indeed involved in brain processes producing the activation in A. Because fiber pathways, even when connecting A and B, do not have to participate in the processing, and because functional activations of A and B can be associated with each other in the absence of direct structural connectivity, we don't expect perfect spatial correspondence of functional and connectional probability profiles. But if detectable, significant correspondence of functional and connectivity profiles should emphasize the functional importance of a tract between A and B. Gender processing is again used for within- and across-condition comparison.
Vertex-wise spatial cross-correlations between functional and structural profiles provide a quantitative measure for the association of the two (cf.
Spatial cross-correlation plots (± SEM)1 between activation probabilities ([−log10(p)]) and structural connectivity scores ([log10(cs/ns)], with [(cs/ns)] reflecting ratios of connecting samples to the number of samples sent out from each vertex) for pANG, pITS and FFG (cf.
Since face stimuli underwent continuous temporal changes during the morphing, the explanatory variables of interest were modeled according to their change over time. Age and gender change were time-binned at the video frame-rate (24 fps). Scaling of each regressor was set to a relative maximum of 1. In order to determine accurate stimulus response functions, especially for age and gender, we extensively evaluated our paradigm and the stimuli employed by various psychometric ratings (see Behavioral Results). Thereby, we empirically identified unbiased stimulus response functions for age and gender, later used for modeling in the fMRI analysis.
As illustrated in
Optical flow
This section is divided into two parts. The first part focuses on the psychometric behavioral data and, based upon these, on modeling the psychophysical changes of the main explanatory variables.
The second part covers the related functional activations, as derived from the same set of (n = 24) subjects, and the structural connectivity between them, i.e. the neuroimaging results of the study.
For the psychometric assessment, we instructed our volunteers to rate their subjective impression of how much facial age actually changed across morph sequences spanning an age spectrum similar to the original fMRI paradigm. In order to minimize the potential rating bias, the (n = 121) face stimuli were morphed to another average-aged male face of 33 years not contained in the original stimulus set. Therefore, the subjects were familiar with the set of faces used as start and targets in the fMRI video paradigm but not with the particular morphs displayed in the psychometric rating. Single start-to-target morphs were randomly played forwards or backwards for the rating, and the corresponding results are plotted in
On a 6-point visual analogue scale, subjective age-gradient ratings for the (n = 121) separate face morphs were best encoded according to Stevens' law of psychophysics
In a second rating, the task was to estimate the age of (n = 201) face stills in years. For this purpose, stills of all (n = 121) real faces and of (n = 80) interpolated age models were displayed randomly. In both of these stimulus samples, i.e. the real and interpolated age models, facial aging increased subjective age rating variability across our (n = 24) subjects similar to the group-ratings of young and old faces reported by Ebner
Based on the actual distribution of rating errors accumulated over all stills, we trichotomized according to upper and lower quintile cutoffs in order to relate age rating performance to fMRI activations (see below). Because implicit age-change processing (during the fMRI experiment) is rather unlikely to strongly correlate with explicit age-rating accuracy in post-hoc assessments of a limited sample size, this was not more rigorously modeled. The upper quintile of most accurate age-raters (n = 5) was compared to average performers (n = 14). The lower quintile of below-average raters (P20, n = 5) was excluded. Confirmed by their own verbal report, their compliance and motivation was limited at the second session so that age-rating performance of these subjects did not correspond to their actual capacities. This was reflected in disproportionally higher post-hoc rating errors and an increased rating variability (low accuracy ≤P20: 9.1±0.8, average accuracy >P20/<P80: 7.1±0.5 and high accuracy ≥P80: 5.7±0.5 years, as averaged over trials).
In order to investigate the psychophysical processing of gender, especially at intermediate ambiguous levels, (n = 119) sample faces along temporal morph continua across gender from the (n = 60) transsexual morphing sequences were rated by the volunteers on a 6-point visual analogue scale according to their subjective impression of facial gender/androgyny levels. In accordance with previous reports
Inclusion of optical flow as a nuisance variable enabled us to account for low-level configural and featural changes during face morphing which would otherwise have confounded the analysis. Optical flow was associated with functional activations in the motion-sensitive cortex (hMT+), see
hemi | cluster | size | CWP | Max | VtxMax | MNI X, Y, Z | vE/BA | annotation | |||
|
left | pANG | 2309 | 0.0001 | 8.739 | 142332 | −41.0 | −74.5 | 27.0 | P |
inferior parietal |
pITS | 526 | 0.0001 | 7.307 | 40331 | −54.1 | −55.8 | −8.5 | P |
inferior temporal | ||
right | pANG | 1177 | 0.0001 | 8.465 | 117820 | 46.7 | −59.3 | 19.5 | P |
inferior parietal | |
pITS | 367 | 0.0004 | 6.025 | 5665 | 54.2 | −53.5 | −9.4 | P |
inferior temporal | ||
|
left | pANG* | 32 | 0.0133 | 5.045 | 146872 | −42.4 | −56.6 | 25.5 | P |
inferior parietal |
|
left | DLPFC | 370 | 0.0005 | 5.948 | 29235 | −36.9 | 19.5 | 22.1 | F |
inf. front. sulcus |
FFG | 228 | 0.0020 | 5.437 | 92500 | −39.1 | −67.9 | −17.2 | P |
fusiform | ||
right | LOT | 862 | 0.0001 | 5.988 | 35952 | 42.4 | −77.3 | −5.4 | O |
lateral occipital | |
FFG | 202 | 0.0021 | 5.107 | 28527 | 36.4 | −56.0 | −16.4 | P |
fusiform | ||
|
left | hMT+ | 605 | 0.0001 | 8.816 | 551 | −42.6 | −79.6 | 0.1 | O |
middle occipital |
right | hMT+ | 1204 | 0.0001 | 8.554 | 91197 | 46.7 | −58.9 | 0.4 | P |
middle temporal |
Clusters significantly activated by changes of facial age, gender and motion/optical flow (FWER-corrected p<0.05 for n = 24 subjects)
hemi, hemisphere; size in [mm2], CWP, cluster-wise probability (non-parametric cluster mass inference over the entire surface; [ ]); Max, peak activation probability (absolute log10-maximum of uncorrected p-values: −log10(p); [ ]); VtxMax, vertex of Max on Freesurfer's average surface; MNI, coordinates in MNI standard space [mm]; vE/BA,
Age change-related activations were centered on the posterior inferior temporal sulcus (pITS), lateral to the fusiform gyrus (FFG), and on the posterior angular gyrus area (pANG) of both hemispheres; see
Within left pANG, a sub-cluster above the superior temporal sulcus (STS) discriminated the upper quintile (P80, n = 5) of best explicit age raters from average performers (P20–80, n = 14) by higher activations, see
Gender change-related activations were detected within FFG bilaterally, see
For each subject, probabilistic tractography was run between all age and gender change-related clusters on individual brain surface reconstructions. An association tract, the ventral portion of Wernicke's perpendicular fasciculus
3D model illustrating how the ventral stream, pITS in particular, may interact via Wernicke's perpendicular fasciculus (WpF) with the posterior magnitude-encoding and approximate number system
The connectivity between functionally defined seed- and target-clusters was quantified by counting how many connecting samples arrived at every vertex of the target (from any vertex of the steed), yielding an index for every target vertex. In order to characterize the extent to which different structural seed-to-target connectivities relate to activation patterns, we then examined the vertex-wise spatial cross-correlations between surface connectivity scores and activation probabilities, i.e. the above connectivity index was correlated (plotted) against the activation probability of the target vertices in
At the group level, by far most consistent spatial cross-correlations between activation probabilities and average connectivity scores were detected for pANG and pITS (cf.
Our investigation provides the first description of a distinct brain network associated with processing the age of faces and its underlying structural connectivity. Although facial age is of high social relevance and has been shown to influence medial prefrontal activations when presumed personality characteristics are rated
Wernicke's perpendicular fasciculus (WpF) has been recognized as a separate cerebral association tract but so far largely escaped further description and attention
Given that tractography cannot reveal directions of information processing, our data do not ensure that it is pITS that recruits pANG and not vice versa. Continuous morphing, despite its strength of presumably augmenting change-sensitive neural responses and supporting our explicit psychophysiological model, precludes reliable detection of temporal delay differences between pITS and pANG. Thus, we acknowledge that these clusters may influence each other reciprocally, as schematically indicated in
Due to the inherent smoothness of fMRI data, which is further pronounced on the average surface, and because connectivity probabilities tend to increase the closer seed and target are located, association of probabilistic connectivity and activation probabilities for extremely short association fibers and cortical regions located very near to each other (e.g., right LOT and pITS
A unique combination of fMRI and diffusion tractography measurements enabled us not only to track anatomical connections between peak activations but also to uncover significant spatial cross-correlations between functional activations on the one hand and structural connectivity probabilities on the other. The demonstration of such associations between functional and probabilistic connectivity measures, which has only recently been highlighted by a different hypothesis and approach
Our results allow us to propose the first coherent model (illustrated in
Considering the extraordinary relevance of age judgements for the interpersonal domain, e.g. to establish peer communication, attractiveness and even empathy, the cognitive processing of facial age and aging reflects an intrinsic core capacity of the human “social brain”. Its functions have also been related to the temporo-parietal junction. Attractiveness, which has not been modulated in our study, is influenced by facial age
Future investigations and lesion studies are required to further elucidate cognitive age processing. Our analysis may be broadened by other approaches examining distributed patterns of neural age-encoding in their selectiveness and specificity but this would have been beyond the scope of this study. More elaborate insights can be anticipated investigating age discrimination upon face inversion, processing the age of non-face objects, adaptation to age, own- vs. other-age effects including associated visual processing strategies and their potential center-periphery bias, cross-modal integration of age information, age processing in the blind, dissociation of non-abstract and numerical age representations, and the development of age-recognition expertise. Even though our study highlights pANG as one key component for age processing, its precise role in this context is still speculative and needs further investigation. Our model, illustrated in
The notion of a distributed neural core system processing fixed attributes vs. changeable aspects of faces
(MOV)
We would like to thank E.K. Warrington for her inspiration and helpful notes, as well as T.E. Behrens, A. Biller, B. Fischl, D. Greve, M. Hanke, T. Höll, M. Jenkinson, D. Dierker, F. Oltmanns, and S.M. Smith for their substantial advice. M. Bendszus and L. Solymosi provided continuous support.