Microstructural Properties of the Cerebellar Peduncles in Children With Developmental Language Disorder

Abstract Children with developmental language disorder (DLD) struggle to learn their native language for no apparent reason. While research on the neurobiological underpinnings of the disorder has focused on the role of corticostriatal systems, little is known about the role of the cerebellum in DLD. Corticocerebellar circuits might be involved in the disorder as they contribute to complex sensorimotor skill learning, including the acquisition of spoken language. Here, we used diffusion-weighted imaging data from 77 typically developing and 54 children with DLD and performed probabilistic tractography to identify the cerebellum’s white matter tracts: the inferior, middle, and superior cerebellar peduncles. Children with DLD showed lower fractional anisotropy (FA) in the inferior cerebellar peduncles (ICP), fiber tracts that carry motor and sensory input via the inferior olive to the cerebellum. Lower FA in DLD was driven by lower axial diffusivity. Probing this further with more sophisticated modeling of diffusion data, we found higher orientation dispersion but no difference in neurite density in the ICP of children with DLD. Reduced FA is therefore unlikely to be reflecting microstructural differences in myelination, rather the organization of axons in these pathways is disrupted. ICP microstructure was not associated with language or motor coordination performance in our sample. We also found no differences in the middle and superior peduncles, the main pathways connecting the cerebellum with the cortex. To conclude, it is not corticocerebellar but atypical olivocerebellar white matter connections that characterize DLD and suggest the involvement of the olivocerebellar system in speech and language acquisition and development.


INTRODUCTION
Children with developmental language disorder (DLD) show significant and unexplained deficits in producing or comprehending language or both compared with their peers (Bishop et al., 2016).A child with DLD can present with a wide range of problems including problems in speech sound discrimination and phonology, in word learning and vocabulary, expressive and receptive grammar (Conti-Ramsden et al., 2001;Krishnan et al., 2021), as well as pragmatics (Bishop, 2014; see Bishop et al., 2016Bishop et al., , 2017 for detailed description and terminology).Despite it being a highly prevalent disorder, research on the neurobiological basis of DLD is still scarce.The majority of neuroimaging studies conducted so far have highlighted differences between DLD and typically developing (TD) children in the function and structure of cortical structures (Mayes et al., 2015), in particular inferior frontal and superior temporal areas that are considered key nodes of the language network (Fedorenko & Thompson-Schill, 2014).Among the most consistent findings in the literature is underactivation in left frontal and temporal cortical areas while processing language as well as microstructural differences in dorsal anatomical pathways that connect them (for a review see Asaridou & Watkins, 2022).
Another relatively consistent finding has been atypical structure in the basal ganglia (Badcock et al., 2012;Lee et al., 2013;Watkins et al., 2002).The basal ganglia play an important role in procedural learning, including speech motor learning during language acquisition (Karuza et al., 2013) and the ability to extract sequential regularities required for learning phonology and grammar (Krishnan et al., 2016).It has been hypothesized that the problems in DLD stem from deficient domain-general implicit learning mechanisms that involve the basal ganglia (Krishnan et al., 2016;Ullman et al., 2020).Children with DLD show poorer performance in nonlinguistic procedural learning, particularly in sequence-based tasks and probabilistic categorical learning (Gabriel et al., 2013;Hardiman et al., 2013;Lee & Tomblin, 2012).Supporting this hypothesis, a recent study using quantitative MRI revealed differences in striatal myelin in children with DLD (Krishnan et al., 2022).
Corticocerebellar circuits, alongside corticostriatal systems, contribute to complex sensorimotor skill learning, including the acquisition of spoken language (Ziegler & Ackermann, 2017).The cerebellum, together with frontal and temporal cortical areas and the basal ganglia, is part of a FOXP2-dependent circuitry that has been proposed to support speech motor control (Vargha-Khadem et al., 2005).Neural expression of FOXP2, a gene in which a point mutation results in a speech and language disorder in members of the KE family, has been found in both the basal ganglia and the cerebellum (Lai et al., 2003).Affected members of the KE family show atypical basal ganglia structure (Vargha-Khadem et al., 1998;Watkins et al., 2002) as well as atypical cerebellar structure (Argyropoulos et al., 2019).Children with cerebellar agenesis or congenital cerebellar malformations show delayed speech acquisition (Glickstein, 1994;Steinlin, 1998).Cerebellar mutism (lack of speech) and dysarthria can follow brain surgery that directly or indirectly affects the cerebellum in children (Küper & Timmann, 2013).Disruption of cerebro-cerebellar circuits, particularly during development, has also been documented in developmental disorders such as autism and dyslexia (for a review see Stoodley, 2016).
The role of the cerebellum in DLD has received little attention.This is partly because children with DLD perform as well as TD children in tasks such as eyeblink conditioning, which rely on the cerebellum (Hardiman et al., 2013).Performance in other tasks, however, that involve the cerebellum (Miall & Christensen, 2004) and require fine motor skills does differ, with children with DLD performing worse than TD (Brookman et al., 2013;Jäncke et al., 2007;Powell & Bishop, 1992;Zelaznik & Goffman, 2010).
Given the uniform architecture of the cerebellum, functional specialization for language must be determined by inputs and outputs from peripheral sensors and effectors along with cortical and subcortical language circuits (Skipper & Lametti, 2021;Steele & Chakravarty, 2018).These inputs and outputs are mediated by the three main white matter pathways that connect the cerebellum to the rest of the brain known as the peduncles.The inferior and middle cerebellar peduncles provide cerebellar inputs, and the superior cerebellar peduncles Developmental language disorder (DLD): A neurodevelopmental impairment in acquiring language that affects more than 7% of children.The language-learning difficulty is unexplained by hearing loss, lack of exposure or opportunity, or by other neurodevelopmental or medical diagnoses.DLD can include problems with all aspects of receptive and expressive language processing and literacy acquisition.deliver cerebellar outputs.The inferior cerebellar peduncles consist of climbing fibers carrying predominantly afferent input from the inferior olive to the cerebellum (Moberget & Ivry, 2016).The middle cerebellar peduncles consist of mossy fibers carrying afferent input from the cortex to the cerebellum via the pons (Stoodley & Schmahmann, 2010).Lastly, the superior cerebellar peduncles consist of mossy fibers carrying efferent input from the cerebellum to the cortex via the thalamus (Moberget & Ivry, 2016).In TD children, the development of the middle and superior cerebellar peduncles peaks in early adolescence (12-15 yrs) while the inferior cerebellar peduncles can peak later, in mid adolescence (12-17 yrs;Simmonds et al., 2014).
The role of the cerebellar white matter connections in speech and language remains relatively understudied.White matter integrity in the cerebellar peduncles has been associated with reading skills in children and adolescents (Bruckert et al., 2020;Travis et al., 2015), and with speech rate in adults who stutter (Jossinger et al., 2021), while differences in white matter microstructure of the peduncles have been reported between controls and adults and children who stutter (Connally et al., 2014;Johnson et al., 2022).There is some evidence that corticocerebellar white matter connections show different microstructural properties in adolescents (>14 yrs) and young adults with DLD compared with a control group (Lee et al., 2020).Very little is known about cerebellar structure in children with DLD and, to the best of our knowledge, cerebellar connectivity has not yet been tested in this population.
The aim of the current study was to investigate cerebellar white matter connectivity in a large sample of children with DLD.We hypothesized that microstructural properties of the main cerebellar pathways (the inferior, middle, and superior cerebellar peduncles) will differ between children with DLD and TD children.We also hypothesized that microstructural characteristics of the cerebellar peduncles in our sample will be associated with performance on language and fine-motor tasks.

Participants
One hundred and seventy-five children aged 10;0-15;11 (yrs; mons) were recruited as part of the Oxford Brain Organisation and Language Development (OxBOLD) study.The study protocol was approved by the University of Oxford's Medical Sciences Interdivisional Research Ethics Committee in accordance with the Declaration of Helsinki.Informed written consent was obtained from parents/guardians as well as from the child prior to study enrolment.Five participants did not complete the behavioral testing or imaging, seven were subsequently found not to meet the study's inclusion criteria, three were excluded due to incidental findings, and one participant was excluded because their cerebellum was not included in the scans (not part of field of view [FoV]).Complete data were therefore available for 159 participants with sample characteristics shown in Table 1.
In order to be included in the study, participants had to pass a bilateral pure tone audiometric screening test to confirm normal hearing and score no more than two standard deviations below the normative mean in two nonverbal IQ tests, the WISC-IV Matrix Reasoning and Block Design tests (Wechsler, 2003).Participants were excluded if they had a diagnosis of a developmental disorder such as Down or William syndrome, or a history of neurological impairments or neurological disorders such as epilepsy.Participants with a diagnosis of autistic spectrum disorder (ASD) or attention deficit hyperactivity disorder (ADHD) were also excluded.Similarly, children who scored more than seven on the hyperactivity subscale of the Strengths and Difficulties Questionnaire (SDQ; Goodman, 1997), or more than 15 on the Social Communication Questionnaire-Lifetime (SCQ; Rutter et al., 2003) were also excluded from the study.All participants grew up in the UK speaking English and met the safety requirements to undergo magnetic resonance imaging (MRI).
Participants in the TD group presented with no history of speech and language problems and scored one standard deviation below the normative mean on no more than one standardized test score of language abilities.Participants in the DLD group presented with a history of speech and language problems and scored more than one standard deviation below the normative mean in two or more of the standardized language tests.Children who presented with a history of speech and language problems, including a previous DLD diagnosis, but failed to meet our DLD criteria at the time of testing, formed a separate group, hereafter referred to as HSL.

Behavioral Measures
All participants who passed our first screening, completed a comprehensive neuropsychological test battery, providing a thorough and in-depth assessment of language and related abilities as well as motor coordination (see Table 1 of the Supporting Information, available at https://doi.org/10.1162/nol_a_00142,for a list of domains and tests used as part of the battery; see Krishnan et al., 2021, for a detailed description).Abbreviations: TD = typically developing, DLD = developmental language disorder, HSL = history of speech-language problems, FA = fractional anisotropy, SNR = signal-to-noise ratio, a.u.= arbitrary units for factor scores.p values reported on group differences with one-way analysis of variance and chi-square tests.
To reduce dimensionality and minimize multiple comparison problems, performance in the language and memory tests was summarized using factor scores.The approach used to identify the best weighted combination of these measures to give a language and a memory factor is described in Krishnan et al. (2021).In brief, a two-factor hybrid exploratory-confirmatory approach (Brown, 2015) was applied and compared against a single factor model.Model fit was significantly better (as indicated using Akaike's information and Bayesian information criteria) for the two-factor compared to the one-factor model (see Krishnan et al., 2021, for a detailed description).Group mean and standard deviation for each factor (Language and Memory), nonverbal IQ, and the z-score for performance on the Purdue Pegboard (Brookman et al., 2013;Tiffin & Asher, 1948) can be seen in Table 1.

MRI Acquisition
Imaging data were acquired using a 3T Siemens Prisma scanner with a 32-channel head coil.Participants wore noise-cancelling headphones (Optoacoustics OptoActive II Active Noise Cancelling Headphones), which were held in place using inflatable pads.Foam padding was inserted around the head for comfort and to restrict movement.Highresolution structural images were acquired using a MPRAGE (magnetization prepared rapid gradient echo) T1-weighted sequence (TR = 1,900 ms, TE = 3.97 ms, flip angle = 8°, FoV = 192 mm, 1 mm isotropic voxel size).Diffusion-weighted MRI followed the same sequence acquisition protocol used in the UK Biobank Project (Miller et al., 2016).In brief, sampling in q-space was performed in two shells at b = 1,000 s/mm 2 and 2,000 s/mm 2 (voxel size = 2 mm, multiband factor = 3).For both diffusion-weighted shells, 50 distinct diffusion-encoding directions were acquired, covering 100 distinct directions across both shells.Five b = 0 s/mm 2 images were obtained as well as three b = 0 s/mm 2 images with reversed phase-encoding direction.

Preprocessing
Diffusion-weighted imaging (DWI) data were processed using the FMRIB Software Library ( Version 6.0; FSL; Smith et al., 2004).Prior to any preprocessing steps, all non-diffusion (b0) images for the anterior-posterior as well as the posterior-anterior phase-encoding direction were manually checked for artifacts.The first anterior-posterior and posterior-anterior b0 image was used as a default, however, in the presence of artifacts the best alternative b0 image was chosen.We started by estimating the susceptibility-induced off-resonance field from the pairs of images using the TOPUP correction tool (Andersson et al., 2003).We then performed skull stripping using FSL's brain extraction tool (Smith, 2002) creating a non-diffusion brain mask.Subsequently, we used FSL's EDDY tool to correct DWI data for eddy current-induced distortions and participant head motion (Andersson & Sotiropoulos, 2016).We used the mporder option (Andersson et al., 2017) to correct for slice-to-volume (i.e., within-volume) movement as well as the estimate_move_by_susceptibility option (Andersson et al., 2018) to correct for susceptibility-by-movement interactions with eddy currents.Outlier detection was performed using the repol option (Andersson et al., 2016) to identify slices with signal loss due to motion during the diffusion encoding.Identified slices were replaced by non-parametric predictions using the Gaussian process (Andersson et al., 2016).We further ran automated quality control (EDDY QC) to detect data acquisition and preprocessing issues at subject and group level (Bastiani et al., 2019; Diffusion tensor model: A model that describes the diffusion rates of water molecules in diffusionweighted imaging data.Unimpeded diffusion in all directions (isotropic) is represented as a sphere, while diffusion predominantly along one direction (anisotropic) is represented as an ellipsoid.
see Table 1 for signal-to-noise ratio (SNR) and relative motion displacement descriptives per group).
Fitting the DWI data using tensor and NODDI models After preprocessing, we fitted a tensor model at each voxel using b = 1,000 s/mm 2 sampled data (DTIFIT; Behrens et al., 2003) to compute fractional anisotropy (FA), eigenvalues (including axial diffusivity; AD) and eigenvector maps.The computed eigenvalues were then used to derive maps for radial diffusivity (RD).
Additional microstructural parameters were derived from neurite orientation dispersion (OD) and density imaging (NODDI; Zhang et al., 2012), using both shells of the diffusion data (1,000 s/mm 2 and 2,000 s/mm 2 ).NODDI protocols model diffusion data in three compartments: intra-cellular, extra-cellular, and cerebrospinal fluid (CSF).NODDI parameters were estimated with the CUDA diffusion modelling toolbox (cuDIMOT; Hernandez-Fernandez et al., 2019) using the Watson model with Markov Chain Monte Carlo optimization.Three parameters were estimated: fraction of the data in which diffusion is isotropic (f iso ; the free water fraction, e.g., the amount of CSF in a voxel); fraction of intra-cellular compartment compared to the total intra-and extra-cellular compartment (f intra ; that is, disregarding f iso; is an index of neurite density or packing density); and OD.OD provides a measure of how dispersed the fibers are, bounded from 0 to 1.A low OD indicates that the fibers are aligned; high OD means fiber directions are not aligned.Changes in OD and f intra both impact FA, as would partial volume effects of CSF captured by f iso .NODDI parameters may help disentangle the microstructural differences underlying differences in FA (Zhang et al., 2012).

Tractography
We estimated the probability distribution of diffusion directions using BEDPOSTX (Jbabdi et al., 2012).The transformation from diffusion to standard space (and vice versa) was calculated using nonlinear registration with the subject-specific T1-weighted anatomical image as an intermediate image in the transformation (FNIRT; Andersson, 2010).Probabilistic tractography was performed using XTRACT (Warrington et al., 2020), which transforms standard space masks for seeds and targets into each participant's native diffusion space.The inferior (ICP) and superior cerebellar peduncles (SCP) in each hemisphere as well as the middle cerebellar peduncle (MCP) were identified using a published protocol and available masks by Bruckert et al. (2019).In brief, masks were used as seeds and targets in XTRACT to robustly segment portions of the peduncles.For the left and right ICP, the seed mask was a region of interest (ROI) placed in an axial slice at the level of the pontomesencephalic junction and the target was a second ROI placed ipsilaterally in an axial slice at the level of the medulla oblongata (target mask).The MCP was segmented by placing ROIs in the central portion of the left (seed mask) and right (target mask) MCP in an axial plane at the medial pons level.The left and right SCP were segmented by placing the seed mask in the SCP in an axial plane at the level of the pontomesencephalic junction and the target mask in the dentate nucleus in an axial plane at medial pons level.We added a midline exclusion mask to prevent the streamlines crossing and the seed and target masks were used as stop/termination masks to prevent streamlines continuing past these ROIs.Note that the midline exclusion mask for the MCP allowed pontine streamlines to cross to the contralateral hemisphere.Figure 1 shows the masks for the ICP (A), MCP (B), and the SCP (C) from the coronal, axial, and sagittal perspective.We used XTRACT default settings; the resulting tract segments of the peduncles were thresholded to remove the bottom 10%, binarized, and used to mask the diffusion images.Figure 2 displays Isotropic signal fraction (f iso ): A proxy for the free water fraction in a voxel corresponding to the amount of CSF present.Takes values between 0 and 1, with highest values in cerebral spinal fluid.

Intra-cellular signal fraction (f intra ):
A proxy for the packing density of neurites, also known as the neurite density index.Takes values between 0 and 1 with highest values in major white matter tracts.

Orientation dispersion (OD):
An estimate of the orientational coherence of neurites (axons and dendrites).The lower the coherence, the larger the variation in orientations of neurites and the larger the OD.OD takes values between 0 and 1, with lowest values in major white matter tracts.

Neurite orientation dispersion and density imaging (NODDI):
Refers to a method that models diffusion MRI data to provide additional information to the tensor model: OD, f inta , and f iso .

Radial diffusivity (RD):
A measure of diffusion across the shorter axis of an ellipsoid, used as a proxy for diffusion across the tracts where, for example, more myelin would impede diffusion.

Axial diffusivity (AD):
A measure of diffusion along the predominant axis of an ellipsoid, used as proxy for diffusion along and within the tracts where, for example, axons are tightly packed and highly organized.

Fractional anisotropy (FA):
A diffusion tensor derived measure approximating the ratio of axial to radial diffusivities to indicate how directional (or anisotropic) diffusion is in a voxel.It takes values between 0 and 1 with highest values in major white matter tracts and lowest values in cerebrospinal fluid.the tractography output in a representative TD participant.Means for each tract for FA, AD, RD, OD, f intra , and f iso were obtained along with whole-brain FA for each participant.

Cerebellar volume analysis
We used Freesurfer ( Version 7.2.0;https://surfer.nmr.mgh.harvard.edu/)to perform volumetric segmentation of participants' T1-weighted images (Dale et al., 1999).Preprocessing included skull-stripping, automated Talairach transformation, subcortical segmentation of gray and white matter volumetric structures (including the cerebellum), and automatic labeling of brain volume (Fischl et al., 2002(Fischl et al., , 2004)).Data from three TD participants and one HSL participant were excluded after quality control.Participants' left and right cerebellar gray and white matter volumes (measured in mm 3 ) were obtained from the segmentation output for statistical group comparisons.Variation in whole brain volume was corrected for by statistically regressing out the Freesurfer estimated intracranial volume (ICV) for each participant.A linear mixed effects model (lme4 package; Bates et al., 2014) was run with volume (in mm 3 ) as the dependent variable, and group (TD, DLD), hemisphere (left, right), tissue (gray, white), group-by-hemisphere-by-tissue interaction, and total intracranial volume as predictors (formula = lmer(volume ~group * hemisphere * tissue + ICV + (1|sub), Data).

Statistical Analysis
All statistical analyses were conducted in R ( Version 4.2.1;R Core Team, 2022).We analyzed group differences in FA using generalized linear mixed effects analysis modeled using beta  distribution as implemented in the glmmTMB package (Brooks et al., 2017).Beta regression models, as introduced by Ferrari and Cribari-Neto (2010), are used when the dependent variable is beta distributed, that is, continuous probability distributions defined on the interval [0, 1].This is suited for FA data that range from 0 (isotropic) to 1 (completely anisotropic).Participants' extracted mean FA values for each of the peduncles (ICP, MCP, SCP) were the dependent variable in the generalized linear mixed models.As fixed effects we entered group (TD, DLD, HSL), hemisphere (left and right in the ICP and SCP) and their interaction, as well as whole-brain FA as a covariate.As random effects we entered intercepts for participants (1| subj).We then adjusted each of the models for age, sex, and relative motion displacement during scan acquisition, to assess whether any significant effects would remain significant in the adjusted models.
model ← glmmTMBðFA group * hemisphere þ wholebrainFA þ ð1jsubjÞ; data ¼ data; beta family ðlink ¼ "logit "ÞÞ We used Bonferroni-corrected alpha levels based on the number of models tested to adjust for multiple comparisons where appropriate.

RESULTS
Probabilistic tractography was successful in all 159 participants.We reviewed imaging quality and tractography results for each participant by visually inspecting registration, DTIFIT, NODDI, and tractography output.Data quality was deemed sufficient for all participants and no observations were excluded from further analyses.A table summarizing mean FA in each peduncle for the three groups can be found in Supplementary Material Table S2.

Group Effects on FA in Cerebellar Peduncles
We tested whether FA in the cerebellar peduncles can be predicted by group (TD, DLD, HSL) by fitting three beta-distributed generalized linear mixed effects models-one per peduncle.To adjust for multiple comparisons, we applied a Bonferroni-corrected alpha level of 0.017.

Inferior cerebellar peduncle
The model predicting mean FA in the ICP showed a significant main effect of group (see Table 2).Post-hoc pairwise comparisons revealed that FA in the ICP was significantly lower in the DLD group compared with the TD group (TD − DLD: t(309) = 2.47, p Bonferroni = 0.042).FA in the HSL group did not differ significantly from the TD and DLD group.There was also a significant main effect of hemisphere due to significantly higher FA on the right compared with the left (see Table 2).The group-by-hemisphere interaction was not significant.The same pattern of significant effects remained after adjusting for age, sex, and head motion (full model summaries are available in Table 2).These results remain the same if we exclude the HSL group from the analysis (see Supplementary Material Table S3).For illustrative purposes, FA data broken down by hemisphere are presented in Figure 3.
To gain a better understanding of the microstructure underlying the FA differences in the ICP, we examined RD, diffusivity perpendicular to the axonal tract which is often used as an index of myelination, and AD, diffusivity parallel to the tract which is often used as an index of axonal integrity (see Supplementary Material Table S4 for descriptive statistics).FA approximates a ratio of AD to RD, so changes in either RD or AD (or both) can affect FA estimates.We ran linear mixed effects models using the same predictors as in the FA model (formula = lmer(measure ~group * hemisphere + whole-brain FA + (1|subj), Data)).The model predicting mean RD in the ICP revealed no differences among groups.There was a significant effect of hemisphere in which the right hemisphere had significantly lower RD than the left (Hemisphere R b = −2.67 × 10 −5 , SE = 7.03 × 10 −6 , 95% CI [−4.04 × 10 −5 , −1.29 × 10 −5 ], t = −3.79,p = 0.000213).The model predicting mean AD in the ICP revealed a significant main effect of group (Group DLD b = −2.79× 10 −5 , SE = 9.64 × 10 −6 , 95% CI [−4.66 × 10 −5 , −9.13 × 10 −6 ], t = −2.89,p = 0.004).Post-hoc pairwise comparisons revealed a significant difference between TD and DLD in AD in the ICP such that participants in the DLD group had lower AD (TD − DLD: t(155) = 2.60, p Bonferroni = 0.031).The group-byhemisphere interaction was not significant.For illustrative purposes, AD data broken down by hemisphere are presented in Supplementary Material Figure S5.
We also fitted our diffusion data using the NODDI tissue model (see Supplementary Material Table S6 for descriptive statistics).This allowed us to compare OD, f intra , and f iso in the ICP among groups, using glmmTMB(noddy ~group * hemisphere + whole-brain FA + (1| subj), Data, beta_family(link="logit")) per NODDI measure as these are also bound between 0 and 1.These measures reflect microstructural complexity of dendrites and axons which Note: TD (typically developing) group acts as reference category for the other groups (DLD = developmental language disorder, HSL = history of speech and language impairments).CI = Confidence Interval; Coefficient = unstandardized (b).An asterisk (*) indicates p values that survived Bonferroni correction.

Neurobiology of Language 783
Cerebellar connectivity in DLD contributes to diffusion tensor indices such as FA.No differences among groups were found for f intra or f iso .We did, however, find a significant group effect in the model predicting mean OD in the ICP (Group DLD b = 0.11, SE = 0.04, 95% CI [0.047, 0.182], z = 3.32, p < 0.001).
Post-hoc pairwise comparisons revealed a significant difference between TD and DLD in OD in the ICP (TD − DLD: t(309) = −3.07,p Bonferroni = 0.007), such that participants in the DLD group had higher OD compared with participants in the TD group.We also found a significant effect of hemisphere where the right hemisphere had significantly lower OD than the left (Hemisphere R b = −0.079,SE = 0.026, 95% CI [−0.129, −0.028], z = −3.06,p = 0.002206).In the absence of a significant group-by-hemisphere interaction no post-hoc pairwise group comparisons by hemisphere were performed (OD data broken down by hemisphere are presented in Supplementary Material Figure S7).
In summary, analysis of the ICP data indicated a significantly lower FA in the DLD group relative to the TD group, which was explained in part by lower AD rather than higher RD.Furthermore, the DLD group had higher OD than the TD group.These group differences did not differ significantly between hemispheres (i.e., there was no interaction).FA was significantly higher and RD and OD significantly lower on the right than on the left across all groups.The data in the HSL group were not significantly different from either the TD or the DLD group.

Middle and superior cerebellar peduncles
The model predicting mean FA in the MCP showed no effect of group.Similarly, no effect of group was found in the model predicting mean FA in the SCP (see Figure 4).

Associations Between Behavioral Outcomes and FA in the ICP
Having identified group differences in the inferior cerebellar peduncles using a categorical approach, we tested for associations between FA in these tracts and behavioral outcome measures including language, memory, and motor coordination.In this continuous analysis approach, we fitted multiple linear regression models predicting behavioral performance from FA in the left and right ICP across groups.We included whole-brain FA as a control covariate in these models.To adjust for multiple comparisons, we applied a Bonferroni-corrected alpha level of 0.008 to all six regression models.While the three groups differed significantly in their performance on measures of language, memory, and motor coordination (see Table 1), we found no statistically significant relationship between these measures and FA in the left and right ICP.
Testing for Previously Reported Group-by-Age Interaction effects A previous study by Lee et al. (2020) tested FA in the MCP and SCP in adolescent and young adults with DLD (14-28 yrs) and found significant group and group-by-age interaction effects.The interaction revealed distinct developmental trajectories in DLD versus TD participants: While FA increased with age in TD participants, it decreased in DLD participants (Lee et al., 2020).We attempted to replicate these findings in a younger but overlapping age range.We ran beta-distributed generalized linear mixed effects models for each tract separately with group (TD, DLD), age, gender, handedness, and nonverbal IQ as fixed effects, and intercepts for participants as random factor (formula = glmmTMB(FA ~group * age + gender + handedness + nonverbal IQ + (1|subj), Data, beta_family(link="logit")).We further expanded this model to the left and right ICP, which were not tested in Lee et al. (2020).To adjust for multiple tests (five regressions), we applied a Bonferroni-corrected alpha level of 0.01.
No significant group or group-by-age interaction effects were found in FA in the MCP, SCP (left and right).There was, however, a significant main effect of group and a significant age * group interaction effect in the right ICP (see Table 3 and Figure 5).Similar effects were found in the left ICP but they did not survive correction.Full ICP model summaries can be found in Table 3.It is worth noting that running multiple linear regression models as per Lee et al. (2020) did not change the pattern of the findings.The significant interaction was probed using the emmeans package (Lenth, 2022) to assess the statistical significance of simple slopes against zero and simple slope differences for age across the two groups.The slope of age on FA in the TD group was 0.025, 95% CI [0.002, 0.048] while in DLD it was −0.03, 95% CI [−0.056, −0.004].The 95% confidence interval did not contain zero for either group, so the simple slope was significant for both.The difference in slopes was also significant (TD − DLD: t(121) = 3.16, p = 0.002) with TD having a higher slope than DLD.

Volumetric Differences in the Cerebellum
We also examined differences in cerebellar gray and white matter in DLD, using Freesurfer anatomical segmentation output (see Supplementary Material Table S8 for descriptive statistics).There was a significant effect of group (Group DLD b = −1541, SE = 749.1,t(430.9)= −2.06,p = 0.040).This was primarily driven by the DLD group having significantly smaller cerebellum that the TD group although the post-hoc pairwise comparison did not survive Bonferroni correction.There was also a significant effect of tissue type (Tissue WM b = −42642.93,SE = 645.59,t(453) = −66.05,p = 2 × 10 −16 ) whereby white matter volume was overall smaller than gray matter, and a significant effect of the ICV covariate (b = 0.012, SE = 0.002, t(150) = 7.51, p = 4.88 × 10 −12 ).The three-way interaction of group × hemisphere × tissue type was not significant so no follow-up tests were conducted.Visual inspection of the data suggested that differences lay primarily in the left and right cerebellar white matter rather than gray matter (plotted in Figure 6).

DISCUSSION
Corticocerebellar circuits, alongside corticostriatal systems, contribute to complex sensorimotor skill learning, including the acquisition of spoken language (Stoodley, 2016;Ziegler & Ackermann, 2017).In this study, we hypothesized that in corticocerebellar circuits are associated with language learning problems in development.To this end, we investigated differences in cerebellar white matter connectivity in children with DLD, a common developmental language disorder, using DWI.We found that FA across the left and right ICP was lower in children with DLD, and that this was primarily driven by decreased AD and increased OD in DLD compared to TD children.We also found atypical age-related changes in the right ICP such that while FA increased with age in TD children it decreased in DLD; put another way, the group difference in FA in the right ICP is only evident at later ages and FA values in the two groups are overlapping at the younger end of the age range.We did not, however, find any associations between FA in the ICP and language, memory, or motor performance.These findings indicate typical corticocerebellar (MCP and SCP) but atypical olivocerebellar connectivity in DLD and suggest that the olivocerebellar system might play in important role in spoken language development.
Considering previous research findings in DLD, which have implicated atypical perisylvian and striatal structure and function, as well as the broad range of difficulties that children with DLD present with (expressive and receptive language, vocabulary, grammar, narrative, etc.), we anticipated differences primarily in the corticocerebellar pathways (i.e., MCP and SCP).Instead, we identified differences in the ICP, a pathway that feeds the cerebellum with sensory inputs (including proprioceptive and visual inputs) from neurons located in the periphery, including somatosensory input from speech organs (oral and facial muscles; Moberget & Ivry, 2016).It has been proposed that this input regarding the current sensory state is combined with feed-forward input from the cortex to allow computation of predictions (or internal models) and prediction-errors (Moberget & Ivry, 2016).These cerebellar computations ultimately contribute to performance refinement and learning via error correction (Caligiore et al., 2016;Koziol et al., 2014).
Animal studies have highlighted the role of the olivocerebellar pathway in motor learning and motor control (Lang et al., 2017) and are corroborated by patient studies that show that lesion in this pathway impairs performance in motor adaptation tasks (Martin et al., 1996).More recently, a study reported that motor adaptation performance in neurotypical controls was negatively correlated with FA in the left ICP (Jossinger et al., 2020) offering more evidence that adaptation relies on afferent cerebellar input.With respect to speech and language disorders, the ICP has been repeatedly implicated in developmental stuttering, a common developmental disorder characterized by speech dysfluency.FA was reduced in the ICP as well as the other cerebellar peduncles in adults who stutter compared to controls (Connally et al., 2014).Reduced FA in the right ICP, but not in other peduncles, was found in children who stutter, and FA in this tract showed a negative relationship with stuttering frequency (Johnson et al., 2022).Lastly, a negative association was found between speech rate and FA in the left ICP in adults who stutter, potentially facilitating a hyperactive error-correction process from afferent inputs and resulting in lower speech rates in this group (Jossinger et al., 2021).
Only a handful of studies have considered the role of the cerebellum in DLD-perhaps unsurprisingly given the scarcity of neuroimaging studies on DLD.One recent study reported larger gray matter volume in the right cerebellum in children with DLD compared with TD-an unanticipated finding as the authors hypothesized (and did not find) differences in the caudate nucleus, the inferior frontal gyrus, and the superior temporal gyrus instead (Pigdon et al., 2019).Our current findings point at reduced cerebellar gray and white matter volume in DLD and are in agreement with the broader literature of atypical (reduced) cerebellar volume reported in a number of developmental disorders such as ASD, ADHD, and dyslexia (Stoodley, 2016).With respect to cerebellar connectivity in DLD, one recent study reported lower FA in the SCP and MCP bilaterally (the ICP was not tested) in DLD; this was again an unforeseen finding as the authors anticipated differences in corticostriatal tracts (Lee et al., 2020).The discrepancy between Lee et al.'s (2020) findings and ours could be due to differences in age (their sample consisted of adolescents and young adults with DLD), sample size, as well as imaging and analysis protocols.
Previous findings on the SCP and MCP in children with DLD emphasized the importance of age as a moderator when considering group differences (Lee et al., 2020).Indeed, we found that while FA increased with age in TD children, it did not in children with DLD in the right ICP (with the left ICP following a similar trend).Longitudinal studies have shown that the MCP and SCP reach their white matter maturation peak between 12 and 15 years, and the ICP between 12 and 17 years (Simmonds et al., 2014).The age range of our sample (10-16 yrs) was therefore well suited for investigating the developmental trajectories of these tracts, indicating an atypical time course of maturation in the ICP in children with language learning problems.Alternatively, it could be the case that older children with DLD were simply more "severe" in their language profile, a possibility we cannot exclude in our cross-sectional approach.

Neurobiology of Language
FA is sensitive to white matter microstructural properties, including myelin content; however, differences in FA do not necessarily translate to differences in myelination (Jones et al., 2013).The group differences we observed in FA in this study are driven by reduced AD and increased OD in DLD.This suggests that reduced orientation coherence of neurites is driving the effect rather than RD and neurite density, through myelination and axon density (Jespersen et al., 2010).Unlike FA, OD is relatively stable across development with only a weak (negative) correlation with age (Lynch et al., 2020).A recent study has also shown that OD in the peduncles is highly heritable (Luo et al., 2022).Our results may therefore reflect causal factors at play in DLD, rather than a consequence of learning deficits.
The main limitation of the present study is the cross-sectional design, which does not allow us to disentangle whether the differences we have observed are the cause or consequence of DLD.Future longitudinal studies are required to test the developmental trajectories of white matter pathways in DLD and to make any strong developmental claims.Another important limitation is that we used DWI tractography to identify and quantify the cerebellum's white matter connectivity in vivo.While its utility is undeniable, tractography is inherently limited and can only offer an indirect and crude approximation of the anatomical fiber pathways (Jeurissen et al., 2019;Schilling et al., 2019;Thomas et al., 2014).Lastly, differences in segmentation protocols have a direct effect on tractography results (Schilling et al., 2021) as do acquisition parameters and tractography approach.Our study is not an attempt to precisely replicate prior work identifying the cerebellar peduncles given the methodological discrepancies.Nevertheless, if the difference revealed by our analysis is a real effect, we would expect that other protocols should be sensitive to it, though quantitative comparisons (e.g., of FA values) between ours and other studies should be made with caution.
To conclude, we found lower FA in the ICP, the white matter pathway that carries sensory input from the periphery to the cerebellum, in children with DLD.Our results point to a potentially suboptimal transfer of sensory input in individuals with profound language-learning difficulties.Unlike cortical and striatal mechanisms, cerebellar contributions to learning might be more important early on during skill acquisition (Stoodley, 2016;Ziegler & Ackermann, 2017).According to this view, the information carried and computations facilitated by the ICP might be selectively consequential for motor learning early on in speech acquisition.Alternatively, the ICP might support motor speech functions across the lifespan, a view partially supported by findings from noninvasive stimulation studies on the cerebellum which introduce disruptions to different motor, but also speech and language functions, in neurotypical adults (Arasanz et al., 2012;Lesage et al., 2012;Runnqvist et al., 2016).The two views are not mutually exclusive, but in order to understand whether the cerebellum selectively affects typical and atypical language development more longitudinal data are needed.Finally, future studies should adopt a network approach to investigate the interactions between the cerebellar and the corticostriatal systems known to be atypical in DLD.Such an approach may be crucial in understanding the neurobiological underpinnings of this complex disorder.statistical advice; Amy Howard for advice on NODDI fitting; Nilgoun Bahar for performing the FreeSurfer quality control; and Dorothy Bishop for providing guidance and feedback throughout this project.

Figure 1 .
Figure 1.Delineation of the Bruckert et al. (2019) masks used to extract (A) the left and right inferior cerebellar peduncles (ICP), (B) the middle cerebellar peduncle (MCP), and (C) the left and right superior cerebellar peduncles (SCPs) from the coronal (left), axial (middle) and sagittal (right) perspective displayed in standard space.Seed masks are colored in red, target masks in blue, and exclusion masks in yellow.An additional stop mask was placed at seed positions (not displayed in the figure).

Figure 2 .
Figure2.The thresholded binarized ICP (left in red, right in blue), MCP (green), and SCP (left in brown, right in yellow) overlaid on the FSL_HCP065_FA image in a typically developing (TD) participant.See Supplementary Material for participant with developmental language disorder (DLD) and participant with a history of speech and language problems (HSL).

Figure 3 .
Figure 3. Mean fractional anisotropy (FA) in the ICP by group (TD in blue, DLD in yellow, HSL in green) and across hemispheres.Graphical representation of ICP in red overlaid on a standardized template in sagittal view.The box is drawn from first to third quartile with the horizontal line drawn in the middle to denote median FA.Whiskers show 1.5 * interquartile range with data beyond the end of the whiskers (group outliers) being plotted individually.

Figure 4 .
Figure 4. Mean fractional anisotropy (FA) in the SCP and MCP by group.Graphical representations of SCP (cyan) and MCP (yellow) overlaid on a standardized template in sagittal and axial view.See Figure 3 for boxplot details.There were no significant differences among groups in these tracts.

Figure 6 .
Figure 6.Mean gray and white matter volume (in mm 3 ) in the right and left cerebellum per group, adjusted for intracranial volume (TD in blue, DLD in yellow, HSL in green).
This research was supported by the National Institute for Health and Care Research (NIHR) Oxford Health Biomedical Research Centre.The views expressed are those of the authors and not necessarily those of the NIHR or the Department of Health and Social Care.The Wellcome Centre for Integrative Neuroimaging is supported by core funding from the Wellcome Trust.

Table 1 .
Sample characteristics of the OxBOLD study

Table 2 .
Model summaries for fractional anisotropy (FA) in the inferior cerebellar peduncles (ICP)

Table 3 .
Lee et al. (2020) fractional anisotropy (FA) of the fiber tracts in the inferior cerebellar peduncles (ICP) as tested byLee et al. (2020)using group (TD vs. DLD), age, sex, handedness, and nonverbal IQ as fixed effects Figure 5. Mean FA in the right ICP by age and group (TD in blue, DLD in yellow).The slopes significantly different from zero also significantly different from each other.Plot presents actual rather than fitted values.