White-Matter Development is Different in Bilingual and Monolingual Children: A Longitudinal DTI Study

Although numerous people grow up speaking more than one language, the impact of bilingualism on brain developing neuroanatomy is still poorly understood. This study aimed to determine whether the changes in the mean fractional-anisotropy (MFA) of language pathways are different between bilingual and monolingual children. Simultaneous-bilinguals, sequential-bilinguals and monolingual, male and female 10–13 years old children participated in this longitudinal study over a period of two years. We used diffusion tensor tractography to obtain mean fractional-anisotropy values of four language related pathways and one control bundle: 1-left-inferior-occipitofrontal fasciculus/lIFOF, 2-left-arcuate fasciculus/lAF/lSLF, 3-bundle arising from the anterior part of corpus-callosum and projecting to orbital lobe/AC-OL, 4-fibres emerging from anterior-midbody of corpus-callosum (CC) to motor cortices/AMB-PMC, 5- right-inferior-occipitofrontal fasciculus rIFOF as the control pathway unrelated to language. These values and their rate of change were compared between 3 groups. FA-values did not change significantly over two years for lAF/lSLF and AC-OL. Sequential-bilinguals had the highest degree of change in the MFA value of lIFOF, and AMB-PMC did not present significant group differences. The comparison of MFA of lIFOF yielded a significantly higher FA-value in simultaneous bilinguals compared to monolinguals. These findings acknowledge the existing difference of the development of the semantic processing specific pathway between children with different semantic processing procedure. These also support the hypothesis that age of second language acquisition affects the maturation and myelination of some language specific white-matter pathways.


Introduction
Since multiple language societies are expanding, with increasing numbers of families with bilingual children, the interest of the scientific community to study the effect of dual language acquisition on cerebral structure, functionality and performance in children has grown. Second language acquisition situations (explicit vs. implicit learning, the age of exposure to second language and more) have been shown to cause differences in second language proficiency by many researchers [1]. There is some evidence that speaking more than one language alters cognitive [2] and attention network processes [3]. This ability may have a beneficent effect on the maintenance of brain integrity with age [2,[4][5][6][7][8]. The bilingual's ability to simultaneously activate two languages, to continuously switch between them, and to inhibit the non-target language [9], can be expected to positively affect the brain structures and networks underlying language and cognition processes [10,11].
The structure-function relation of the brain has been investigated in a number of bilingualism studies. In a study of brain integrity in the elderly, it was demonstrated that the integrity of bilinguals' brain has a slower rate of decline compared to monolinguals [6]. In another study examining cerebral atrophy in Alzheimer disease, comparing bilingual patients with monolingual controls, computed tomography (CT) showed increased atrophy in bilinguals in areas used to distinguish AD; this shows that bilingualism alters the structure of specific brain regions [12].
Nevertheless, the study of bilingualism and its impact on brain neuroanatomy remains obscure. One of the unanswered questions refers to the differences in brain microstructure development between bilinguals and monolinguals. Because brain development and maturation occurs to a large extent in childhood, a direct approach to investigating maturational changes of bilingualism in terms of the functionality of the pathways involved needs follow-up studies that compare the microstructure of the brain at different time points in children exhibiting relevant functional differences.
Brain adaptation to a changing environment, known as plasticity [13,14], is altered by bilingualism both in the gray matter volumes [11], the microstructure of white-matter pathways [15], and the cross-sectional area of sub-regions of the corpus callosum [16]. However, none of these studies investigated the developmental changes in the microstructural correlates of bilingualism; and there has been no structural brain imaging study in bilingual children. Further, the age of L2 acquisition is commonly overlooked in present studies. To the best of our knowledge, concerning the impact of bilingualism on the microstructural changes in white-matter, no such quantitative longitudinal studies have been carried out.
Diffusion tensor imaging (DTI) has been used as a non-invasive tool to characterize whitematter microstructure in vivo [17]. It provides quantitative information about the integrity and maturation of the brain through quantities like fractional anisotropy (FA), apparent diffusion coefficient (ADC) [18][19][20] and can be used to study the development of white-matter pathways [21]. Recently, longitudinal DTI has been used to study the effect of normal aging [22][23][24].
In their longitudinal DTI study in children Yeatman et al (2012) have found active biological processes in some white-matter tracts, which significantly differed according to their reading skills. Here they reported that children with higher reading proficiency had a lower MFA value on the pathways lAF and lIFOF which increased over time; while those with below-average reading skills yielded a higher initial MFA with a fast decline over time [25]. In another longitudinal study on dyslexic children, Hoeft et al (2011) predicted reading improvement based on variation in brain function and structure [26]. Treit and colleagues (2013) also reported a correlation between the changes in mean diffusivity and reading ability of patients with FASD (Fetal Alcohol Spectrum disorders) [27]. These findings reinforce the hypothesis that the functionality of specific pathways may impact the way these pathways mature.
In the current longitudinal tractography study, we have examined the development of the microstructure of four language-processing-related pathways in elementary school children using magnetic resonance diffusion tensor imaging (MR-DTI). We included three groups of subjects: simultaneous and sequential bilinguals (SimBils and SeqBils), and monolinguals (Monos). Simultaneous bilinguals were defined as individuals who acquired both their languages in parallel from birth, while sequential bilinguals were exposed to their second language from the age of three (at school). The choice of language-related pathways was motivated in our earlier study [10]. In short, the first pathway is the left inferior occipitofrontal fasciculus (lIFOF), which provides the ventral connection for language processing and is involved in semantic processing [28][29][30]. The second comprises the left arcuate fasciculus/superior longitudinal fasciculus (lAF/lSLF), the dorsal stream [31], connecting Wernicke's area to Broca's area [32]. The third is the bundle connecting the anterior part of the corpus callosum to the orbital lobe (AC-OL). And the fourth is the bundle of fibres emerging from the anterior midbody (AMB) of the corpus callosum that are associated with the premotor and supplementary motor cortices (AMB-PMC). This part of the corpus callosum was reported to have a larger mid-sagittal area in bilinguals as compared to monolinguals [16]. We added right inferior occipitofrontal fasciculus (rIFOF) as a non-linguistic pathway to act a control bundle.
In our previous study of the impact of bilingualism on the microstructure of language-specific pathways [10], it was observed that MFA value of the lIFOF and AC-OL differed significantly between the two groups of bilinguals and monolingual controls. SimBils showed the highest and Monos had the lowest MFA on the lIFOF bundle. While it was demonstrated that the bundle AC-OL has higher MFA in Monos compared to SimBils. In all cases the MFA of the bundles in SeqBils was amid the two other groups.
Our hypothesis driven aim was to supplement our earlier results with data on the diffusion anisotropy after an interval of about two years following the first observations. We expected the differences between the 3 language groups observed at the beginning of the two-year period (as documented in [10]) to be confirmed at the end of the interval. In addition, based on the findings from previous longitudinal studies [25][26][27]33] that report the impact of specific functions on white-matter development, we hypothesized that functional differences between the language groups would affect the maturation rate of their language related pathways. Especially the lIFOF, which is involved in semantic processing, was expected to reflect maturation rate differences: bilinguals have to process more semantic data (in 2 different languages) and are therefore especially dependent upon efficient neural connections in this bundle [34][35][36].

Methodology Participants
Forty healthy male and female children who also served in our first study [10] were reassessed after an average time interval of 2 years (19-27 months, mean: 22.3 months; SD: 2.3 months). The female to male ratio was 1/1 (see Table 1). As before, the children were subdivided in 3 groups with equal variances (14 simultaneous bilinguals, 16 sequential bilinguals and 10 monolinguals). All subjects used French or Dutch as their first language (L1) and the second language (L2) of the bilinguals was restricted to Roman or Germanic languages, both branches of the Indo-European language family.

DTI analysis
As in our previous paper [10], within slice registration was performed to correct for eddy current distortion. The DTIStudio program was used to analyze the registered images and to produce corresponding Fractional Anisotropic maps [37]. A detailed description of using the PAR/REC files and deriving FA maps was reported in [10].
The FACT-algorithm (Fibre Assignment by Continuous Tracking) [38] was applied to reconstruct the 3d tracts, responsible for language processing. Each tract started from the centre of a voxel with an FA value exceeding 0.2, proceeded along the eigenvectors of neighbouring voxels and ended when the FA became smaller than 0.2 or when the tract deviated by more than 40 degrees. In order to select the desired fibre bundles, we used a region-of-interest (ROI) based approach (Fig. 1). The four selected bundles were the lAF/ lSLF, the lIFOF and the AC-OL, which were tracked by a two-ROI approach, and the AMB-PMC, which was found by means of a single ROI.
These ROIs had been saved for each subject in the first study (at time T1), and in order to have an accurately identical tracking in the follow-up study (at time T2), the same ROIs were used there. To this end, the 3D anatomical scans for both time points were normalized and registered to a customized paediatric template that was made based on the age and gender distribution of our study sample (at both time-points). In order to minimize the risk of errors due to using available templates that are based on adult population, Template-O-Matic toolbox was used to create the customized template using the age of subjects at both time-points and also the gender distribution [39].
The resulting normalization parameters were used for registering the diffusion images and corresponding FA maps of the both waves of scans to the template. Thus the risk of miss-localization of ROIs due to brain growth was minimized. The ROIs from T1 were then loaded on the images of T2, avoiding any manual redrawing of the ROIs.

Quantitative and statistical analysis
The fibre tracking data were analyzed using SPSS version 20.0 (SPSS Inc, Chicago, IL USA). For each bundle and for each subject, we computed the mean FA-value (MFA) and its standard deviation. The Shapiro-Wilk test was used to assess the normality of the resulting data distribution in each language group. The majority of these distributions were not normal.
To test for gender differences, we used the two-sample t-test to compare the MFA values at time 1, the MFA values at time 2, and the change between the two MFA values (ΔMFA).  (3) The study had been approved by the Ethics Committee of our University Hospital (UZBrussel, Belgium) and informed written consent was obtained from all parents and the documents were recorded. doi:10.1371/journal.pone.0117968.t001 To test our main research question, a repeated measures ANOVA was applied to assess the main effect of language group ('Group'), the main effect of the time interval ('Time') and also the interaction effect ('Time x Group'). A separate analysis was done for each of the four bundles. Finally, to explore the relationship between the time interval between the two observations (ΔT = T2-T1) and the corresponding degree of change in MFA values (ΔMFA), we computed bootstrapped Pearson correlation values.
To deal with non-normality, we used bootstrapping (Efron & Tibshirani, 1994) both to assess the significance of the test statistics, and to report confidence intervals for the mean differences in follow-up post-hoc tests. All post-hoc tests were Bonferroni corrected to control for type l error and in all tests P-values smaller than 0.05 were assumed to indicate significance. Location of ROIS to trace specific fibres; plane1: mid-sagittal slice; plane2: a coronal slice located half-way between the genu of the CC and the anterior margin of the insula; plane3: a coronal slice behind the splenium of the CC; plane4: a coronal slice at the level of the rolandic operculum; plane5: an axial slice lateral to the ventricular trigone; plane6: a coronal slice between the anterior edge of the CC and the anterior end of cerebral hemisphere Cyan ROIs (A: plane4 and B: plane5) are those to trace lAF/lSLF [81], green ROIs (F: plane6 and G: plane1) to trace AC-OL [82], the yellow ROI (C: plane1) to trace AMB-PMC [83] and the red ROIs (D: plane3 and E: plane2) to trace lIFOF [44] [Taken from our previous study [10]]. doi:10.1371/journal.pone.0117968.g001

Results
Four participants (3 SeqBils and 1 SimBil) were excluded from further analysis due to the fact that the ROIs on the FA maps for both time points could not be perfectly matched and thus the bundles could not be traced correctly.

Gender
The two-sample t-tests did not reveal any significant differences between males and females. Only the MFA values of the lIFOF in the second time point had a marginally significant difference (F>M), t (36) = 2.02, P = 0.051.
One way ANOVA revealed a significant group difference for this bundle in the second wave of scans F (2, 35) = 3.455, P = 0.043. Bonferroni corrected Post-Hoc comparisons for lIFOF-2nd showed a significant difference between Monos < SimBils (P = 0.038).
The interaction effect 'Time x Group' was also significant, F (2, 35) = 3.432, P = 0.044 with an effect size of η 2 = 0.164. Bonferroni corrected Post-hoc comparisons revealed a significant group difference between simultaneous > sequential bilinguals (P = 0.028). The comparison between SimBils> Monos also led to a significant difference (P = 0. 006). But no significant difference could be observed in the comparison between Monos and SeqBils (P = 0.919).
The correlation analysis indicated that the ΔMFA value of lIFOF pathway was positively correlated with ΔT (correlation coefficient = 0.722, P = 0.03).
ANOVA test also showed a significant group difference for ΔMFA value of the bundle lIFOF F (2, 35) = 3.432, p = 0.044. Table 2 summarizes the results of Post-hoc Bonferroni corrected tests. (Fig. 2B shows the Delta IFOF value for the bundle lIFOF in 3 groups). These show that the MFA value of the bundle lIFOF has changed the most in SeqBils compared to the 2 other groups.

Discussion
Based on our hypothesis that speaking more than one language affects the maturation of the brain, we conducted a 2 year longitudinal DTI study examining the mean fractional anisotropy, MFA, of four distinct language-related pathways in simultaneous and sequential bilingual and monolingual children. The first wave of scans was conducted at an age around 9 years (T1) and the second wave after a time interval of about two years (T2).
We have limited our investigations of the fibres AF/SLF and IFOF to the left hemisphere based on the left lateralization of language in more than 95% of right-handed people [40,41] and also the fact that the role of lIFOF and lAF in language processing is well established [30,[42][43][44][45].

The results of the 1st cross-sectional study
The data obtained at T1 already gave rise to an earlier cross-sectional report focused on the bilingualism-related differences in the white-matter of children between the ages of 8-10 years [10].
Using parametric tests it showed that the lIFOF and AC-OL were characterized by MFA values that were significantly different between the monolinguals and the two groups of bilinguals. Bootstrap results are based on 1000 stratified bootstrap samples. Fig. 2A represents the MFA values of the lIFOF for the three groups and two runs. The results of the follow-up study With the exception of the AC-OL, all bundles showed at least marginally significant changes in MFA over the period of two years. For the AC-OL, the group comparisons confirmed the difference in MFA between the language groups encountered at T1 in our previous study. However, this difference could no longer be established at time T2. The absence of significant group differences in the MFA of lAF/ lSLF and PMC at T1 was confirmed at T2. The MFA of the lIFOF showed significant differences between the language groups at both T1 and T2. Also ΔMFA for this bundle differed significantly between the groups. Post-hoc tests showed these differences could be restricted to the comparisons SeqBil vs. SimBil in the case of ΔMFA and SimBil vs. Mono in the case of MFA2.
lIFOF Among widely so many areas in the temporal, parietal and frontal parts of the brain which are involved in semantic processing [29,46,47], the left inferior fronto-occipital fasciculus (IFOF) shows significant correlation with mostly all semantic procedures [48]. The lFOF is part of an important subcortical structure underlying the "semantic processing network" [29,49].
This longitudinal DTI study indicates that bilingual children, who have different semantic processing procedures compared to monolinguals, have a different rate of developing the bundle IFOF which is a semantic processing specified pathway. As at T1 we showed significant bilingualism-related differences in the MFA values for this bundle [10], The bundle lIFOF was of main interest of the present study as we expected it to be most affected by bilingualism. IFOF connects the prefrontal brain regions to the occipital lobes and is of major importance for language processing. The dorsolateral prefrontal cortex is a part of language-control network which is specifically activated in bilingual language tasks [50]; and the temporal and occipital lobes process the visual and audio language stimuli. Subcortical stimulation studies have demonstrated the participation of IFOF in semantic processing [29], while a picture-wordmatching experiment has confirmed its role in visual-verbal incongruence judgment [51]. Its established role in word recognition and reading ability [52] and our current findings further confirms the importance of this pathway for language processing.
Recently, a clear correlation has been reported between the integrity of inferior longitudinal fasciculus (ILF) and object recognition in children [53]. This pathway is also known to carry visual information from the occipital brain regions to the temporal lobes [54] and to link the representation of objects to their lexical labels [55,56]. As there is a strong correlation between ILF and IFOF [57], this indicates that IFOF may also play a role in the process of object recognition [58]. Researchers of early language pay special attention to children under 2 year playing with objects [59][60][61][62]. They link the substitution of objects in this play (e.g. using a stick as a sward) [59], to early language development [60,61,63]. The significant group difference in the ΔMFA values of the lIFOF bundle suggests different maturation rates of this bundle for the three groups.
In this longitudinal study we found significant changes in MFA value of IFOF over 2-year time; reflecting the alterations in fibre characteristics during this period. The findings in the second wave of scans are consistent with the previous cross-sectional study [10] that showed significantly higher MFA values of the bundle lIFOF in bilinguals compared to monolinguals. Furthermore, Lebel and colleagues have reported that many fibres-such as SLF and IFOFreach their maximal anisotropic changes between 13 and 20 years [64]. Compatible to their findings, in our study, the lIFOF bundle also showed significant increases in MFA value. Moreover, our results suggest that these changes are associated with linguistic skills and with age of SLA in children. In other words, maturation of this fibre seems to be related to semantic differences in bilingual children's linguistic skills and also associated with the duration of being bilingual. Nevertheless, a direct assessment of semantic skills and correlation with MFA changes would be necessary for a more precise interpretation.

lAF/lSLF
In our previous study [10], we had found that the bundle lAF/lSLF could not be reconstructed in all cases. This fact was explained by the incomplete myelination in the subjects involved, a result of the relatively slow rate of maturation of this specific tract (the myelination process of this bundle was reported to continue until the 3 rd decade of life [65,66]. Although this bundle did not show significant difference between the 3 groups in either of the two waves of observations, the fact that the MFA values of the two groups of bilinguals changed significantly over the two years, while it did not change in monolinguals supports the idea that the maturation of this bundle occurs faster in bilinguals compared to monolinguals.

AC-OL
Our observation that the AC-OL bundle did not differ significantly between the three groups at T2, in spite of the higher MFA in monolinguals compared to bilinguals at T1, could indicate that the bilinguals have caught up with the monolinguals regarding the development of these fibres passing from the anterior part of the corpus callosum to the orbital lobe, however more specific and precise interpretation of the lack of previously seen differences is only possible by replication of this study with more participants, longer time intervals between the two scan time points. With a relatively short (~2 years) interval between T1 and T2, certain tracts may undergo late or protracted myelination in young children.
The role of this bundle in language processing is still ambiguous, however, there is a reported relationship between the degree of brain lateralization and the size of anterior corpuscallosum [67]. The established correlation between language lateralization and the size of the corpus-callosum [68,69], and also the difference in the degree of language lateralization between bilinguals and monolinguals [70], strengthens the contribution of the bundle AC-OL to bilingualism. The proven role of orbitofrontal cortex (OFC) in understanding others' mental state [71], receiving input from audio stimuli [72], and sentence completion tasks [73] makes this part of the brain particularly important for linguistic studies. Moreover, in a condition of deliberately selecting a word that fits a sentence and obeys semantic or syntactic rules, the OFC activation is clearly observed [73]. Monitoring two languages at the same time and the need to select the suitable word between the two languages makes this selection process more crucial for bilinguals and causes difference in the maturation rate of fibres passing this area of their brain compared to monolinguals.

AMB-PMC
In case of the bundle PMC, the tests did not result in significant difference between the MFA values of the 3 groups at the time T1 and T2. This shows that challenging our assumptions, we may have to reject the hypothesis that bigger cross-sectional anterior mid-body of the CC [16] is originated from different diffusion characteristics of fibres. However, the possible lack of statistical power may increase the risk of type 2 errors in the interpretation of the findings of this bundle [74]. Luders et al (2010) investigated the structural changes of the corpus callosum in a large group of children and adolescents aged 5-18years. They found significant correlations between callosal thickness and age suggesting that the alternating size of corpus callosum may be related to permanent change of fibres connecting homologous cortices in this range of age [75]. Other study groups, also considering wide ranges of age and a large populations, and have reported a continuous maturation of the CC throughout the lifespan [76,77].

rIFOF
The right IFOF in this study was chosen as a non-language pathway to serve as a control tract. Having no evidence about the role of rIFOF in language circuitry one might consider the maturation of this bundle independent of linguistic differences among the three groups. The significant change in the MFA value of this pathway over the two years and the lack of observed group differences indicates that the maturation process of this bundle over the two-year time interval is not influenced by either bilingualism or age of SLA.

Age of Second language acquisition and its impact on maturation of fibres
The observation that sequential bilinguals showed the highest degree of change in MFA among the three groups is in line with our hypothesis that the age of SLA affects the microstructural development in the brain of children. Sequential bilinguals learn their second language at the age of three or four from scratch, and pass four stages to reach full proficiency in the L2 which are: home language use, observational and listening stage (non-verbal stage), telegraphic and formulaic speech, and finally productive language use [78,79]. Completing these four stages may last from six month to two years taking into account the individual skills and the quality of L2 learning [80]. Thus at the age of 8-10, when we conducted our first wave of scans, it had been around 3-6 years that SeqBils had full proficiency in their L2 and were considered balanced bilinguals, whereas SimBils at that age had been balanced bilinguals for 8-10 years. At the time of our second wave of scans, bilinguals had 2 more years of full proficiency in their L2 and this counted up to 5-8 years for SeqBils and 10-13 years for SimBils. Compared to the first wave, the proportional difference in the length of proficiency was higher in SeqBils than in Simbils indicating stronger changes over two years in these parts of the brain in SeqBils compared to SimBils. In short, the magnitude of change in MFA of language fibres, and most evident the lIFOF is directly related with the years of being bilingual. Sequential bilinguals presented the highest degree of change in the MFA value of their language pathways. In Seq-Bils, compared to SimBils, the proportional duration of being bilingual Years of being bilingual at T2 Years of being binlingual at T1 À Á was higher reflecting a higher ΔMFA value for this group.

Conclusions, future perspectives and Limitations
The results of this study prompt the important conclusion that not only speaking more than one language, but also the age of L2 acquisition affects maturational differences and changes in the white-matter pathways involved in language processing. In spite of the relatively short time-interval between the two waves of observations, these observations support our research hypothesis indicating that language workloads has a significant impact on their myelination and aging process. Language groups that have a non-similar way of using specific pathways tend to exhibit differences in both their MFA at a given age and the rate at which these MFA change with age.
The most important limitation of this study was the time-interval of 2 years between two waves of the scans, which was chosen in such a way that, at the same time, it allowed detectable changes in the brain and fitted perfectly in our 4-year research project.
A more balanced group design, in larger populations and denser and longer time coverage would increase our understanding of bilingualism-induced developmental differences in the microstructure of white-matter throughout life.
Another point of interest in such longitudinal studies is to look into gender differences and the possible interaction of such differences with the change in the maturational characteristics of the brain. Having larger samples and more powerful statistical comparisons, this may provide a clearer insight about the influence of sex differences in brain maturation.
Another limitation of the current study was that, as the participants were chosen to be balanced proficient bilinguals in the first wave of experiments, the language tests were not re-assessed in the second run. It could be sequential bilinguals became more balanced in their knowledge or use of both languages even in relation to simultaneous bilinguals, and these changes might have been correlated with WM microstructural changes. However, as we did not include the appropriate behavioural measurements in run two this assumption should be interpreted cautiously.
Two methodological limitation regarding fibre-reconstruction in human brain are relative dependence on initial position of ROIs and also dealing with kissing, crossing and fanning fibres. Tract-editing based on prior anatomical knowledge of the specific pathways makes reconstruction results less dependent on the location of ROIs. Together with prior knowledge, terminating the tracking procedure by encountering non-cigar shaped anisotropy (we used FA>0.2) can overcome the problem of crossing and fanning fibres to a reasonable extent [38] Of note, the pathways currently studied here are surely not the only important ones in language processing and bilingualism related issues. Further studies may do well to examine additional pathways to improve our understanding of bilingualism-induced changes on the maturation of white-matter microstructure.
In order to get a better insight about the developmental differences between the groups in more specific language regions, whole brain analysis techniques e.g. Voxel based morphometry (VBM) or TBSS could be applied.