Main

Autism spectrum disorder (ASD) (1,2) is a heterogeneous disorder (or collection of related disorders) of multifactorial etiology, with great range in severity of symptoms. ASD affects ∼1 in 166 children and is four times more prevalent in boys than in girls. Core features of individuals with autism include the following: 1) impairment in reciprocal social interactions; 2) verbal and nonverbal communication deficits; 3) repetitive and ritualized behaviors; and 4) a narrow range of interests. Approximately 30% of individuals with ASD manifest some degree of mental retardation (3).

Magnetic resonance (MR) examination allows researchers and clinicians to noninvasively examine brain anatomy in vivo. Structural MR examination is widely used to investigate brain morphology because of its high-contrast sensitivity and spatial resolution and because it entails no radiation exposure; the last feature is particularly important for children and adolescents (4). MR diffusion-tensor imaging (DTI) assesses the direction of water diffusion for each voxel; this technique is sensitive to loss of white-matter (WM) integrity and to differences in connectivity.

Structural MR examination provides several means by which researchers can delineate structural changes in the brains of individuals with ASD. We will review recent advances in our understanding of the structural MR correlates of ASD. We will review analytic methods that researchers have applied to structural MR data, such as voxel-based morphometry (VBM), surface-based morphometry (SBM), and tensor-based morphometry (TBM), longitudinal MRI studies, and DTI studies of ASD. Finally, we will review efforts to generate predictive models of ASD based on MR-derived features.

Structural MR in ASD

There exists an extensive literature centered on the use of structural MRI to identify abnormalities in patients with ASD. Based on the image-analysis methods applied to the structural (i.e. image) data, these studies can be classified into region-of-interest (ROI)-based, VBM-based, SBM-based, or TBM-based.

ROI-based analysis requires experts to manually or semi-manually delineate brain regions. Depending on the degree of automation, this process may be labor intensive and time consuming. An inherent bias in ROI-based analysis is that such studies can analyze only a limited number of brain regions, and the results clearly depend on the ROIs chosen by investigators. However, to the extent that the voxels within a region manifest similar structural abnormalities common to patients with ASD, the use of ROIs will increase statistical power. In contrast, voxel-wise approaches assess structural changes throughout the brain and may be more appropriate when no consensus exists regarding brain regions thought to be central to ASD. VBM, SBM, and TBM are examples of voxel-wise morphometric analysis methods.

ROI-based volumetry.

MR studies of total brain volume demonstrate that young children with ASD (age, 18 mo-4 y) have 5–10% abnormal enlargement in brain volumes compared with those of normal controls (5–7). This finding seems to be related to increases in both gray-matter (GM) and WM volumes, but not to ventricular volumes (7). However, whether this abnormal enlargement persists or not into later childhood and adolescence is not as clear (5,8).

Researchers have focused particular attention on the corpus callosum: reduced volumes in the anterior (genu and rostrum), middle (body), and posterior (isthmus and splenium) callosal subregions have been reported in juveniles and adults with ASD (9–12). For example, Piven et al. (12) used T1-weighted MRI to examine the size of the anterior, body, and posterior subregions of the corpus callosum in autistic individuals {26 males and 9 females, mean age = 18 y [SD 4.5] and normal controls [20 male and 16 female, mean age = 20.2 y (SD 3.8)]}. They found a significantly smaller average size of the body and posterior subregions of the corpus callosum in the autistic individuals.

Increased amygdala volumes have been reported in children with ASD (6,13). In a meta-analysis (14), Stanfield et al. reported that enlargement of amygdala present in children with ASD was not found in older subjects. They found significant relationships between age and effect size for the left and right amygdala. As age increases amygdala volume in autistic subjects decreases relative to controls.

Voxel-based morphometry.

In VBM studies, there are two principal features: tissue density and tissue volume (15). For density-based analysis, researchers have focused on the relative concentrations of GM and WM structures in spatially normalized images (i.e. the proportions of GM and WM to all tissue types within a region). However, many studies instead aim to detect regional differences in volumes of a particular tissue (GM or WM). These studies are volume-based.

Table 1 summarizes VBM studies of ASD. Ideally, we would aggregate the results of these VBM studies at the voxel level; however, there are two major barriers to this approach. First, we cannot compare the coordinates of voxels across studies that use different brain templates. Second, different studies use different image-processing pipelines. There is significant variability in choices of registration methods and smoothing-kernel parameters, among other aspects of image processing. Therefore, even for studies using the same brain template, the chance that voxel coordinates would remain consistent across studies is low.

Table 1 VBM studies of ASD
Full size table

Given this difficulty with aggregating results at the voxel level, we have summarized VBM findings at the lobar level. When we pooled density-based and volume-based studies together, we found that there was no consistent pattern of regional specificity with respect to GM and WM differences. However, interesting patterns emerged when we analyzed density-based and volume-based studies separately. For volume-based studies, the mean age for the ASD group ranged from 8.9 to 31 y. For density-based studies, the mean age for the ASD group ranged from 9.3 to 32 y. We compared the mean ages for the ASD group between volume-based studies and density-based studies using the two-sample t test and found no significant difference (p = 0.807).

For each combination of study type (volume-based or density based) and anatomic structure (frontal, temporal, parietal, and occipital lobes and the limbic system), we identified the most frequently occurring pattern, which we call prototypical. For example, we identified seven studies investigating GM density changes in the frontal lobe; among them, four reported decreased GM density, two reported no change, and one reported increased GM density; thus, the prototypical pattern was decreased GM density.

For volume-based studies, there is evidence for increased GM volumes in the frontal, temporal, and parietal lobes and in the limbic system, and decreased WM volumes in the frontal and temporal lobes and in the limbic system. Similarly, there is evidence for decreased GM density in the frontal and temporal lobes, and decreased WM density in the temporal lobe.

Both density-based and volume-based studies reported GM differences in the frontal and temporal regions. The most striking difference between volume-based and density-based studies is the direction of the abnormality. The majority of volume-based studies reported increased GM volume in these regions, whereas most density-based studies reported decreased GM density. For WM, both volume and density tend to decrease in the temporal lobe.

Surface-based morphometry.

The intrinsic topology of the cerebral cortex is that of a 2-D sheet with a highly folded and curved geometry; VBM cannot directly measure this topology. In contrast, SBM centers on cortical topographic measurements, and thereby provides information complementary to that provided by VBM.

Table 2 summarizes several studies of cortical thickness changes in patients with ASD (16–19). Among them, two (16,19) studies are voxel-based, and two (17,18) are atlas-based. The mean age for the ASD group ranges from 9.2 to 33 y. The changes occur primarily in the frontal, temporal, and parietal lobes but not in the occipital lobes. However, the direction of change (cortical thinning versus increased cortical thickness) is not consistent across studies for the frontal and temporal lobes. The majority of studies reported increased cortical thickness in the parietal lobes (17–19).

Table 2 Cortical-thickness-based studies for ASD
Full size table

Hardan et al. (20) investigated gyrification patterns in autism. They found in the autistic group, left frontal gyrification index was higher in children and adolescents but not in adults. Cortical folding was decreased bilaterally with age in the total ASD sample but not in controls.

Tensor-based morphometry.

Both TBM and VBM measure volume changes. The advantage of TBM over VBM is that false-positive findings due to systematic group differences in registration errors are less likely. This is because in TBM, the signals analyzed are generated based on the registration of the images rather than on aligned segmented GM. That is, TBM does not require GM be perfectly registered across subjects. Brun et al. (21) used TBM to analyze T1-weighted MR images of 24 male children with ASD [mean age, 9.5 y (SD 3.2)] and 26 age-matched controls. They found significantly decreased GM volumes in the parietal, left temporal, and left occipital lobes, bilaterally.

Longitudinal MRI studies.

The studies in previous sections are cross-sectional; that is, they aim to detect differences in structural MRI features between individuals with ASD and age-matched controls. When researchers are interested in the effects of age on ASD, i.e. the trajectory of brain growth in ASD, they perform longitudinal studies (22).

Schumann et al. (23) performed a longitudinal study of brain growth in young children with ASD. For each child, a structural MRI scan was acquired at multiple time points beginning at 1.5 y up to 5 y of age. They collected 193 MRI scans on 41 toddlers who received a confirmed diagnosis of autistic disorder at about 48 mo of age and 44 typically developing controls. Volumes of eight ROIs (frontal GM, temporal GM, parietal GM, occipital GM, cingulate GM, total GM, total WM, and total cerebral volume) were calculated using FreeSurfer. They found that all regions except occipital GM undergo an abnormal growth trajectory. The cerebrum and several of its subdivisions in most children with ASD was enlarged by 2.5 y of age.

Hardan et al. (24) examined developmental changes in brain volume and cortical thickness in children with ASD, using structural MRI. Their study included 18 children with autism and 16 healthy age- and gender-matched controls. MRI scans were obtained at baseline and at follow-up. The mean time difference between the two scans was 2.1 y. The mean ages of ASD and control group were 10.9 y (SD 1.2) and 10.7 y (SD 1.2), respectively. They found differences in total GM volume over time, with significantly greater decreases in the autism group compared with controls. Differences in cortical thickness were also observed over time, with greater decreases in the autism group compared with controls in the frontal, temporal, and occipital lobes.

Diffusion Tensor Imaging in ASD

DTI (25) generates quantitative measures of WM tract integrity by providing detailed information about how water molecules diffuse within each voxel. Two widely used DTI-derived features are fractional anisotropy (FA), which quantifies the spatial symmetry of diffusion, and the apparent diffusion coefficient (ADC), which quantifies the degree of restriction of diffusion. Other, less commonly examined, DTI-derived features are mean diffusivity and radial diffusivity. Abnormalities in myelination, axonal number, diameter, and orientation can lead to changes in FA and ADC (26).

Most DTI studies of ASD have focused on ROI measurements, voxel-wise measurements, or fiber tracking (27–34). In a DTI study of the corpus callosum including 43 individuals with ASD [mean age, 16.2 y (SD 6.7)] and 34 age-matched controls, Alexander et al. found reduced FA, increased mean diffusivity, and increased radial diffusivity in the ASD group (29). These results suggested that the microstructure of the corpus callosum was affected in autism. Barnea-Goraly et al. (27) used DTI to investigate WM structural integrity in seven male children and adolescents with autism [mean age, 14.6 y (SD 3.4)] and nine gender- and IQ-matched control subjects. Based on voxel-wise analysis, they found reduced FA values in brain regions that were implicated in theory-of-mind tasks (ventromedial prefrontal cortex, anterior cingulate, temporoparietal regions, and amygdala) and in social cognition (fusiform gyrus and adjacent to superior temporal sulcus). Additional clusters of reduced FA values were seen in occipitotemporal regions and in the corpus callosum. Sundaram et al. (33) used tractography to investigate frontal lobe DTI changes in 50 children with ASD [mean age, 4.75 y (SD 2.43)] and 16 typically developing controls. There was a trend toward statistical significance in the FA of whole frontal lobe fibers (p = 0.1). They found that FA was significantly lower in the ASD group for short-range fibers (p = 0.003) but not for long-range fibers. There was no significant difference in the number of frontal-lobe fibers across groups.

Collectively, DTI-based ASD studies have consistently reported abnormalities of the corpus callosum across a broad age range (from young children to adults) (27–29,32,34). These studies have also consistently reported differences in prefrontal WM (27,33,34), cingulate gyrus (27,30,34), and internal capsule (27,28,32,34).

Diagnostic Models of ASD Based on MR-Derived Features

MR-based diagnostic models hold the promise of complementing standard behavioral assessment of patients with ASD. Studies centering on MR-based diagnostic models usually involve three steps. First, investigators extract features from MR images; then they build a diagnostic model (a classifier) using machine-learning algorithms or statistical models; finally, investigators evaluate and validate the resulting diagnostic model.

Several studies centered on MR-based diagnostic models of ASD (18,35–37). For these studies, the mean age for the ASD group ranged from 6 to 33 y. Different features (volume, surface, and thickness) have been used to build diagnostic models of ASD. The sensitivity of these diagnostic models is in the range of (0.77, 0.95), the specificity is in (0.75, 0.92), and the accuracy is in (0.81, 0.87). Overall, MR-based diagnostic models can accurately differentiate individuals with ASD from normal controls.

The performance of a diagnostic model is critically affected by the feature types selected as components of that model. For example, Jiao et al. (18) compared thickness-based diagnostic models to those based on structure volumes. They used four machine-learning techniques to generate diagnostic models, and found thickness-based classification was superior to volume-based classification, for each combination of classifier and performance metric.

Summary and Conclusion

In reviewing recent advances in our understanding of structural MRI correlates of ASD, we have found several consistent patterns. Figure 1 illustrates these findings:

  • ROI-based volumetry reveals that young children with ASD have abnormally increased total brain volume. In addition, juveniles and adults with ASD have reduced corpus callosum volume, and children with ASD have increased amygdala volume.

  • VBM studies of ASD differ depending on whether a density-based or volume-based analysis was performed. The majority of volume-based studies found increased GM volume in the frontal and temporal lobes, whereas most density-based studies found decreased GM density in these regions. For WM, both volume and density tend to decrease in the temporal lobe.

  • The majority of SBM studies of ASD reported increased cortical thickness in the parietal lobes.

  • Longitudinal MRI studies of ASD reported abnormal growth trajectories in the frontal and temporal lobes.

  • DTI studies of ASD consistently reported corpus-callosum abnormality across a wide age range. Differences in prefrontal WM, cingulate gyrus, and internal capsule were also consistently reported.

  • MR-based diagnostic models can differentiate individuals with ASD from controls with high sensitivity and specificity. However, the performance of diagnostic models was critically affected by the feature types that investigators selected.

Figure 1
figure 1

Findings of structural MRI studies in ASD. (A) ROI-based volumetry reveals that young children with ASD have increased total brain volume, juveniles and adults with ASD have reduced corpus-callosum (green) volume, and children with ASD have increased amygdala (red) volume. (B) VBM studies of ASD. The majority of volume-based studies found increased GM volume in the frontal (green) and temporal (orange) lobes; whereas most density-based studies found decreased GM density in these regions. (C) SBM studies of ASD reported increased cortical thickness in the parietal lobes (green). (D) Longitudinal MRI studies of ASD reported abnormal growth trajectories in the frontal (green) and temporal (orange) lobes. (E) DTI studies of ASD reported abnormalities in the corpus callosum (red), prefrontal WM (blue), cingulate gyrus (gold), and internal capsule (not shown).

Full size image

Finally, we note that there were many conflicting findings regarding MR abnormalities in individuals with ASD. Many factors, such as inclusion and exclusion criteria, population age, MR acquisition parameters, details of the image-processing pipeline, feature extraction procedures, analytic methods used to detect group differences, and sample sizes, may have contributed to these disparities.