This study included data from an Autism Center of Excellence funded by the National Institutes of Health. The Infant Brain Imaging Study (IBIS), the parent network, is an ongoing study of infants at risk for autism. Four clinical data collection sites are associated with the study: University of North Carolina, Chapel Hill; University of Washington, Seattle; Children's Hospital of Philadelphia; and Washington University, St. Louis. Data were coordinated through the Montreal Neurological Institute (MNI) at McGill University, and data processing was performed at the University of North Carolina and the Scientific Computing and Imaging Institute at the University of Utah. The parent study enrolled and assessed 6-month-old high-risk infants who were seen for follow-up assessments at 12 and 24 months of age. Written informed consent was obtained from parents or legal guardians before enrollment, and the study procedures were approved by institutional review boards at each site.
The exclusionary criteria were 1) a significant medical condition known to affect brain development, 2) sensory impairment, 3) low birth weight (<2,200 g) or prematurity (<36 weeks gestation), 4) perinatal brain injury secondary to maternal complications or exposure to specific medications or neurotoxins (e.g., alcohol) during gestation, 5) non-English-speaking family, 6) contraindication for MRI (e.g., metal implant), 7) adoption, and 8) first-degree relative with idiopathic intellectual disability, psychosis, schizophrenia, or bipolar disorder.
The present study group included all high-risk infant siblings who received diffusion-weighted MRI scans at 6 months and for whom behavioral assessments were completed at age 24 months as of June 2011. Symptoms of ASDs were measured at 24 months by using the Autism Diagnostic Observation Schedule (37), which was completed by research-reliable administrators to maximize agreement across sites. On the basis of classifications from this scale, the high-risk infants were divided into two groups: ASD-negative (below the ASD cutoff) and ASD-positive (above the cutoff). At 24 months, 28 infants met the criteria for ASDs and 64 did not. To characterize differences in autism symptoms between groups, a severity score was generated for each participant. Symptom severity scores on the Autism Diagnostic Observation Schedule range from 1 (least severe) to 10 (most severe), with scores of 4 or higher consistent with the presence of an ASD (38). The Mullen Scales of Early Learning (39) were administered at each visit, and the early learning composite score at 24 months of age was used to characterize general cognitive ability. The composite score at 12 months was substituted in the data for four subjects who were missing complete 24-month data for the Mullen scales.
MRI brain scans were completed at the clinical sites on identical 3-T Siemens TIM Trio scanners (Siemens Medical Solutions, Malvern, Pa.) equipped with 12-channel head coils during natural sleep. The diffusion tensor imaging sequence was acquired as an ep2d_diff pulse sequence with a field of view of 190 mm (6 and 12 months) or 209 mm (24 months), 75–81 transversal slices, a slice thickness of 2 mm isotropic, 2×2×2-mm3 voxel resolution, a TR of 12,800–13,300 ms, a TE of 102 ms, variable b values between 0 and 1,000 s/mm2, 25 gradient directions, and a scan time of 5–6 minutes. Intra- and intersite reliabilities were initially established and regularly evaluated by scanning traveling volunteers, or “phantoms,” at all sites within the same week (40).
Data from diffusion-weighted imaging were screened by using DTIprep software (41), which automatically detects artifacts, corrects for motion and eddy current deformations, excludes images with artifacts, and generates a full report. Expert raters manually removed scans with clear residual artifacts. Data sets with fewer than 18 (72%) gradient diffusion-weighted images after this quality procedure were excluded from further processing owing to a low signal-to-noise ratio.
Group analysis of the data from diffusion-weighted imaging, processed by means of diffusion tensor estimates, employed an improved processing pipeline (42). This processing overcomes the major challenge in implementing tract-oriented statistics in large study groups, which is finding consistent spatial parametrization within and between groups. This includes a computational anatomy approach for nonlinear coregistration of the diffusion tensor imaging data to a template reference coordinate frame, a process to parameterize fiber tracts to functions of length, and the mapping of individual tract geometries into common coordinates.
Computational Anatomy Mapping
Unbiased atlas building (43) was used to provide one-to-one mapping between the image data and the template atlas, wherein the atlas is built from the population of data as the centered image with the smallest deformation distances. Registration proceeds in two steps (42). The first applies linear, affine registration of diffusion-weighted imaging baseline images to a structural weighted T2 atlas by using B-spline registration and normalized mutual information (44). This is followed by an unbiased, deformable atlas-building procedure (43) that applies large-deformation diffeomorphic metric mapping transformations. The procedure relates individual data sets to the study-specific atlas template space by means of nonlinear, invertible transformation. Tensor maps were calculated from the diffusion-weighted imaging data sets by using weighted least-squares estimation and were transformed into the atlas space with tensor reorientation by the finite strain approach (45). The transformed tensor images were averaged by using the Riemannian framework (46), resulting in a final three-dimensional average tensor atlas for tractography and tract parameterization. Longitudinal diffusion-weighted imaging data for the subjects covering ages 6 to 24 months were mapped into a common atlas space by using the preceding procedure.
Seed label maps were created according to existing tractography methods (47, 48) and drawn in the combined atlas for regions of interest by using 3D Slicer (www.slicer.org). A secondary check of regions of interest was made by using an atlas for early childhood (33). Label maps were created for the following fiber tracts: genu, body, and splenium of the corpus callosum; fornix; inferior longitudinal fasciculus; uncinate fasciculus; anterior thalamic radiation; and anterior and posterior limbs of the internal capsule. Label maps were created bilaterally for all tracts except the corpus callosum. Fiber tracts generated in 3D Slicer were processed for spurious or incomplete streamlines by means of open source software developed in-house (FiberViewer; http://www.ia.unc.edu/dev/). Fractional anisotropy values were generated for each fiber tract. Fractional anisotropy is an index measuring the degree of anisotropy of local diffusivity, ranging from 0, for isotropic diffusion in fluid, to 1, for strongly directional diffusivity in highly structured axonal bundles (18, 19). Axial (λ1) and radial [(λ2+λ3)/2] diffusivity values, which represent diffusion parallel and transverse to axonal directions, were also produced.
Demographic characteristics were recorded at 6, 12, and 24 months. Potential group differences between ASD-positive and ASD-negative groups were tested for age, sex, and Mullen Scales of Early Learning composite score at each time point by using either t tests (age and early learning composite score) or Fisher's exact test (sex).
Longitudinal trajectories of mean fractional anisotropy values for fiber tracts in the ASD-positive and -negative groups were compared by using random coefficient linear growth curve models. The random coefficient model fits a group developmental trend while accounting for variability in individual growth trajectories. Among the 92 high-risk subjects included in the analysis, 14 subjects had data from one visit, 40 had data from two visits, and 38 had data from three visits. The mixed-model framework accommodates different patterns of missing data and unbalanced designs.
Although white matter is known to develop more rapidly in the first year of life than in the second, a linear model was determined to best fit the change in fractional anisotropy across time given the limited number of available time points. Separate random coefficient growth curve models were fit for each fiber tract with age, group, and group-by-age interaction as fixed effects and with intercept and age slope as random effects. All growth curve models included the 24-month early learning composite score as a covariate. The primary hypothesis of different age trajectories in the two groups was tested in the interaction of group and age for each tract. With parameters from the random coefficient models, we were able to estimate separate growth slopes for the ASD-positive and ASD-negative groups. Least-squares mean fractional anisotropy values were estimated for each group and contrasted at 6, 12, and 24 months to highlight differences between groups at each individual time point. For the primary analysis of fractional anisotropy trajectories, the false discovery rate was used to correct for multiple comparisons across multiple tracts. Axial and radial diffusivity were examined in a secondary analysis to fully characterize fractional anisotropy results by means of the growth model described in the preceding.
Though the sex ratios in the two groups were not significantly different, to study for potential confounding effects of sex, we ran sensitivity analyses by adding sex and sex-by-age interaction in the preceding models. The p values for the primary hypothesis of age-by-group interactions decreased slightly with no change to the conclusions (results not presented).
All analyses were done by means of SAS software, version 9.2 (SAS Institute, Cary, N.C.).