Participants
Participants were enrolled in the University of California (UC) Davis MIND Institute Autism Phenome Project. This study was approved by the UC Davis Institutional Review Board. Informed consent was obtained from the parent or guardian of each participant. Structural and diffusion-weighted images (n = 397) were acquired for at least one time point in 221 children (139 ASD, 82 TD). Of these, 98 children (60 ASD [47 males/13 females], 38 TD [25 males/13 females]) were imaged at one time point and 123 children had longitudinal magnetic resonance imaging (MRI) data available: 70 (46 ASD [35 males/11 females], 24 TD [15 males/9 females]) were imaged at two time points, and 53 (33 ASD [30 males/3 females], 20 TD [13 males/7 females]) were imaged at all three time points. Data from a subset of these participants have been reported previously [14,19].
Diagnostic assessments included the Autism Diagnostic Observation Schedule-Generic (ADOS-G) [20,21] and the Autism Diagnostic Interview-Revised (ADI-R) [22]. All diagnostic assessments were conducted or directly observed by trained, licensed clinical psychologists who specialize in autism and had been trained according to research standards for these tools. Inclusion criteria for ASD were taken from the diagnostic definition of ASD in young children formulated and agreed upon by the Collaborative Programs of Excellence in Autism (CPEA) using DSM-IV criteria. Participants met ADOS cutoff scores for either autism or ASD. In addition, they exceeded the ADI-R cutoff score for autism on either the Social or Communication subscale and within two points of this criterion on the other subscale. An ADOS severity score was calculated ranging from 1 to 10 [23], which allows comparison of autism severity across participants tested with different ADOS-G modules. Overall developmental quotients (DQ) were determined for all participants using the Mullen Scales of Early Development (MSEL) [24].
Typically developing children were screened and excluded for ASD using the Social Communication Questionnaire [25]. Children with typical development were also excluded if they had first-degree relatives (that is, siblings) with ASD. Inclusion criteria included developmental scores within two standard deviations on all scales of the MSEL. All children, both TD controls and children with ASD, were native English speakers, ambulatory, had no contraindications for MRI, no suspected vision or hearing problems or known genetic disorders, or other neurological conditions. In the ASD group, one child was excluded for the presence of fragile X.
Imaging
MRI scans were acquired during natural, nocturnal sleep [26] at the UC Davis Imaging Research Center on a 3T Siemens Trio whole-body MRI system (Siemens Medical Solutions, Erlangen, Germany) using an 8-channel head coil (Invivo Corporation, Gainesville, FL, USA). Images were obtained using a three-dimensional T1-weighted magnetization-prepared rapid acquisition gradient-echo (MPRAGE) sequence (TR 2,170 ms; TE 4.86 ms; matrix 256 × 256; 192 slices in the sagittal direction; 1.0-mm isotropic voxels) and a diffusion-weighted, spin echo, echo planar imaging sequence (‘ep2d_diff’; number of slices: 72; slice thickness: 1.9 mm; slice gap: 0.0; matrix size: 128 × 128; voxel size: 1.9-mm isotropic; phase-encoding direction: anterior to posterior (A>>P); TR: 11,500; TE: 91; scan time: 6 min and 29 s), with an effective b-value of 700 mm2/s, 30 gradient directions, and five interleaved b = 0 images. Thirty-six children (24 ASD [22 males/2 females], 12 TD [10 males/2 females]) were excluded from the study due to waking up prior to the completion of the diffusion sequence.
To accomplish longitudinal imaging at three time points, scans were acquired from October 2007 to October 2012. In August 2009, the Siemens 3T Trio MRI system was upgraded to a Trio Total Imaging Matrix (TIM) MRI System running version VB15A operating system software. All the VA25A sequences were upgraded and mapped to their corresponding VB15A sequences.
For T1-weighted scans, changes in hardware and software over this scanning period were controlled for using a calibration phantom (ADNI MAGPHAM, The Phantom Laboratory, Salem, NY, USA, http://www.phantomlab.com) scanned at the end of each MRI session. Distortion correction was then carried out on each participant’s MPRAGE image (Image Owl, Inc., Greenwich, NY, USA, http://www.imageowl.com/) [19]. This step ensures accuracy in measurements of the midsagittal area of the corpus callosum and total cerebral volume by removing any distortion associated with changes in scanner hardware over time.
For the diffusion-weighted sequence, the spatial resolution, b-value, and gradient directions were preserved following the MRI system upgrade. While the diffusion-weighted parameters were not directly changed, there may be differences in diffusion-weighted measures in regions with reduced geometric distortion. To control for these differences, we include MRI system upgrade status (pre-upgrade vs. post-upgrade) as a nuisance covariate for all statistical analyses involving diffusion tractography or diffusion-weighted measures.
In addition, we evaluated the proportion of participants (by diagnosis and sex) scanned pre- vs. post-upgrade. Prior to the upgrade, 126 (74 ASD [59 males/15 females], 52 TD [37 males/15 females]) scans were acquired. After the upgrade, 271 (177 ASD [148 males/29 females], 94 TD [57 males/37 females]) scans were acquired. Importantly, there were no differences across scanner upgrade status for diagnostic group (chi-square = 1.6, P = 0.21) or sex (chi-square = 0.01, P = 0.91). Within each diagnostic group, there was also no difference in observed frequencies between males and females (ASD: chi-square = 0.55, P = 0.46, TD: chi-square = 1.6, P = 0.20).
For participants who were scanned at multiple time points, we also evaluated diagnoses and sex of participants whose longitudinal scanning took place entirely pre-scanner upgrade, those that spanned the upgrade point, and those whose scanning was entirely post-upgrade. Of the 123 participants with longitudinal data, 9% (7 ASD [5 males/2 females], 4 TD [3 males/1 female]) have complete pre-scanner upgrade data, 43% (31 ASD [26 males/5 females], 22 TD [13 males/9 females]) span the upgrade point, and 48% (41 ASD [34 males/7 females], 18 TD [12 males/6 females]) have complete post-scanner upgrade data. There were no differences in the proportion of participants scanned either pre-, post-, or spanning upgrade status for diagnostic group (chi-square = 1.47, P = 0.48) or sex (chi-square = 0.35, P = 0.84). Within each diagnostic group, there were also no differences across sexes (ASD: chi-square = 0.63, P = 0.73, TD: chi-square = 0.491, P = 0.78).
DTI image processing
Raw diffusion images were checked for the presence of motion artifacts prior to preprocessing. Each image was visually inspected, and volumes were excluded if any signal dropout was detected. The number of volumes excluded was recorded, and if the number of diffusion directions excluded was greater than or equal to six (20% of total diffusion directions), the entire scan was excluded. By these criteria, 14 scans (4 ASD [4 males/0 female], 10 TD [3 males/7 females] were excluded for too much motion. The remaining 397 scans were included in the analysis. Of these, 289 (73%) contained no artifacts - all diffusion directions were included. In 47 scans (12%), one diffusion direction was excluded. This most frequently occurred at the beginning of the sequence - some children would startle in their sleep at the onset of the noises. Two to three diffusion directions (volumes) were excluded in an additional 50 scans (12.6%), and four to six diffusion directions (volumes) were excluded in 11 scans (3%). Additional file 1: Table S1 provides details about the number of volumes excluded for each diagnostic group and across sexes. Importantly, the number of excluded volumes (0 to 6) did not differ by diagnostic group (Fisher’s exact test, P = 0.13) or sex (Fisher’s exact test, P = 0.16). Within each diagnostic group, ASD or typical development, Fisher’s exact test revealed no differences by sex (ASD: P = 0.16, TD: P = 0.44).
Diffusion tensor imaging (DTI) data were preprocessed and analyzed using mrDiffusion, a custom, freely available software package developed by the Vision, Imaging Science and Technology Activities (VISTA) lab, Stanford, CA, USA (http://vistalab.stanford.edu/newlm/index.php/Software). DTI preprocessing included removal of eddy current distortion effects [27], alignment to the T1 image in AC/PC space, and calculation of diffusion tensors. Artifacts were removed using the robust estimation of tensors by outlier rejection (RESTORE) algorithm [28].
Tractography of callosal fibers and segmentation by cortical projection zone
For fiber tractography, an ROI was defined manually in mrDiffusion by tracing the corpus callosum on a single slice in the midsagittal plane. Fiber tracts in the left and right hemisphere were then estimated separately using a deterministic streamlined tracking algorithm [29-31] with a fourth-order Runge-Kutta path integration method. Step size was fixed at 1 mm and path tracing proceeded using a fractional anisotropy (FA) threshold of 0.15 and a path angle threshold of 30°. The subset of fibers in each hemisphere intersecting the corpus callosum ROI was identified (Figure 1A). Using these sets of fibers, the callosum was segmented for each hemisphere separately according to the fiber projection zone using the method introduced by Huang et al. [17]. In brief, fibers were visualized using Quench (http://white.stanford.edu/newlm/index.php/QUENCH), and a series of planes were used to define anatomical targets of the callosal fibers [16]. A total of 397 scans from 221 participants were analyzed. Five trained raters manually segmented the callosal fibers. Intraclass correlation coefficients (ICCs) were calculated for each fiber region and ranged from 0.80 to 0.99. Mean ICCs for the left and right hemispheres were 0.96 and 0.93. In addition, a single expert rater (CWN) reviewed and edited segmentations for all 794 hemispheres. Defined projection zones included orbitofrontal, anterior frontal, lateral frontal, superior frontal, superior parietal, posterior parietal, occipital, and temporal regions (see Figure 1A,B,C). The cross-sectional area of each cortical projection zone fiber subdivision was determined on the midsagittal plane (Figure 1D). To evaluate diffusion properties, fibers from the right and left hemispheres were merged and cropped to the high coherence zone of 1 cm within the midsagittal plane (Figure 1E). Mean diffusivity (MD), radial diffusivity (RD), axial diffusivity (AD), and FA were measured for each fiber subdivision. In brief, AD describes diffusion parallel to the principle diffusion direction (that is, along the long axis of an axonal bundle), and RD describes diffusion perpendicular to the principle diffusion direction. MD describes the average total diffusion, and FA is a scalar value of the normalized standard deviation of the three diffusion directions.
Corpus callosum midsagittal area and Witelson subdivisions
Distortion-corrected T1-weighted images were preprocessed to remove non-brain tissue and to correct for field inhomogeneity [32]. Total cerebral volume (TCV) was derived as described previously [14,19]. For midsagittal corpus callosum measurements, images were aligned along the axis of the anterior and posterior commissures (AC/PC) and resampled to yield 0.5-mm3 voxels using Analyze 11.0 [33]. The midsagittal slice was defined using the central fissure and the aqueduct of sylvius. The midsagittal area of the corpus callosum was manually delineated by two expert raters. ICCs for subdivisions ranged from 0.83 to 0.97. ICC for the total corpus callosum was 0.98. After the total midsagittal area of the corpus callosum was defined, seven subdivisions were segmented according to the procedure described by Witelson (1989) [18]. Subdivisions included rostrum, genu, rostral body, anterior midbody, posterior midbody, isthmus, and splenium [18].
Analytic plan
We used mixed-effect regression models for repeated measures [34] to characterize the longitudinal changes in the corpus callosum and to examine the association of sex, diagnosis, and different callosal subregions/subdivisions with overall levels and rates of change in callosal size, while accounting for the effect of other variables such as TCV or scanner upgrade. The models are flexible and allow children to have different numbers of scans and different lag times between the scans. This approach allowed us to treat subregion/subdivision as a repeated effect within the mixed-effect models for the corpus callosum. The core model used for the cortical projection zone subregions had fixed effects for subregion (orbital, anterior frontal, lateral frontal, superior frontal, superior parietal, posterior parietal, temporal, occipital), diagnosis, sex, upgrade status (pre- or post-upgrade), age, and TCV. Both age and TCV were centered at the time 1 averages for the TD control subjects. In this way, the intercept in the model can be interpreted as the average occipital subregion area (the reference region) for a TD female with average age and TCV at time 1. Individuals were permitted to have differing overall and subregion sizes, by including random effects for intercept and subregions that were assumed to follow a multivariate normal distribution. We allowed the variance of the residuals to differ across cortical projection zone subregions.
This core model allowed us to describe the overall pattern of differences across regions, diagnosis, and sex and assess maturation (age) effects. We then built a hierarchy of questions by adding and testing all two-way interactions between age, subregion/subdivision, sex, and diagnosis in the model. This allowed us to assess whether the maturation effects differed by subregion, sex, or diagnosis, whether the pattern of regional differences differed by sex or diagnosis, and whether there was a sex by diagnosis interaction. These interactions were not retained in the final model when they failed to add significantly to the model. Higher-order three-way or four-way interaction effects were likewise tested against simpler models including all relevant significant lower-order interaction effects. Any significant interaction effects including diagnosis by sex were further examined for a subset of specific simple comparisons of interest involving simple comparisons of diagnosis within levels of sex (males: ASD vs. TD, females: ASD vs. TD) and simple comparisons of sex within levels of diagnosis (ASD: males vs. females, TD: males vs. females). As such, alpha levels were not adjusted for such simple effect testing given the limited number of comparisons that were considered of interest a priori.
Similar mixed-effect models were used to model diffusion-weighted measures (FA, MD, RD, and AD) and Witelson subdivisions. The models for diffusion measures were adjusted for scanner upgrade status (pre- or post-upgrade) but not for TCV. The model for Witelson subdivisions included TCV.
Secondary analyses examined whether the results of the primary analyses could be accounted for by baseline DQ. All models were implemented using PROC MIXED in SAS 9.4 [35].