Background
Studies of volumetric quantification of magnetic resonance imaging (MRI) data in autism have largely been non-replicative. To illustrate with a single example of considerable theoretical importance to autism, interest in hippocampal volume has been high since the neuropathology studies of Bauman and Kemper [
1] reported evidence of increased cell packing density and reduced cell size in the hippocampus. However, as pointed out in two recent articles [
2,
3], there is little agreement between studies measuring hippocampal volume in individuals with autism. To date, there are 4 studies reporting no differences [
4‐
7], 3 reports of decreased volume in autism relative to controls [
8,
9], and 3 reports of increased hippocampal volume in autism [
2,
3,
10]. The differences among the studies are considerable, including different sample characteristics, MRI scanners, methods for assessing hippocampal volume and statistical approaches for analyzing data (reviewed in part in Rojas et al. [
2]; see also [
11]), and no single, clear candidate emerges to account for the varying findings for the hippocampus.
There are a variety of potentially confounding factors in autism research that challenge the scientific community that are not MRI specific, such as age, gender, IQ, and the inherent heterogeneity of the disorder separate from these other factors (reviewed in Palmen and van Engeland [
12]). For example, Palmen and van Engeland [
12] accept the argument made by Piven and Arndt [
13] that IQ must be strictly matched or statistically factored out in autism studies, and their conclusions on cerebellar studies in autism are based on dismissal of the various studies that do not adhere to this recommendation (e.g., [
14‐
17]). In contrast, Yeung-Courchesne and Courchesne [
18] argued strongly that in attempting to "control" for a domain non-specific construct such as IQ, variability truly associated with autism could be discarded as "non-specific". Indeed, few MRI studies critically examine the effects of using IQ matched, but etiologically heterogeneous samples of developmentally disordered controls. The alternate use of statistical covariates for IQ is also usually performed without consideration for the assumptions underlying such analyses (e.g., correlation between the covariate and dependent measure in the sample being measured).
In commenting on differences in major findings between MRI studies, methodological differences are often appealing candidates. Differences in MRI scanner hardware and acquisition hardware can lead to differences across sites (e.g., see Patwardhan et al. [
19]) that are difficult to assess. Compounding this problem, many, if not most of the MRI-based morphometric literature in autism have employed manual tracing methods. Manual tracing methods suffer from several problems that negatively impact their reliability and validity. First, they rely on sulcal/gyral boundary determinations that are relatively arbitrarily determined and are highly variable from individual brain to brain. Second, those sulcal/gyral boundaries are often not well correlated with histoanatomically defined areas [e.g., [
20‐
22]]. Third, they are susceptible to rater judgments, which necessarily introduce inter-rater, and therefore intra- and inter-laboratory error variance into the comparisons. Individual laboratories develop highly reliable protocols for such boundary determination, but these protocols are not standardized and are challenging to adapt between sites.
Despite the difficulties inherent with manually traced structural MRI studies, a small number of findings appear to replicate across laboratories. The most consistent finding has been that brain volume appears to be larger in autism [
23‐
26], although there is emerging consensus that this finding is restricted to young children with autism [
27,
28]. Differences in cerebellar volume are reported in many studies [
15‐
17,
29‐
34], but the directionality of the finding in autism (larger or smaller relative to comparison group) and region within the cerebellum (vermis or hemisphere) differ greatly between studies. Most studies of the cerebellum, however, report reduced vermal volume in autism, unless IQ is matched or covaried in the analysis (e.g., see [
12] for a review). Some studies have reported enlarged caudate nuclei [
35‐
37], although this finding does not always survive correction for overall brain volume [e.g., [
35]] and another study failed to observe enlargement with or without correction for total brain volume [
38]. Enlargement of the caudate nucleus has been correlated with repetitive/stereotyped behaviors in two studies [
35,
36].
Several automated approaches to MRI morphometry have been developed over the past decade, including template-based, region of interest (ROI) labeling procedures (e.g., see [
39‐
42]), cortical thickness approaches [
43,
44] and voxel based morphometry (VBM [
45‐
47]). Some of these approaches have been applied to studies of autism spectrum disorders [
44,
48‐
55], particularly using VBM methods. In the first VBM study of 15 people with autism, Abell et al. [
48] identified abnormalities in the frontal cortex, parahippocampal gyrus, fusiform gyrus, occipital-temporal junction, and the cerebellum [
48]. Waiter et al. [
54] obtained significant findings in similar regions, although the directionality of the findings (increase or decrease) for some structures (e.g., left inferior frontal gyrus) was opposite to Abell et al. [
48]. McAlonen et al. [
51] identified reductions in fronto-striatal regions. Two additional VBM studies have limited their analyses to regions associated with social cognition. Boddaert et al. [
49] reported bilateral gray matter (GM) decreases in the superior temporal sulcus, a region whose activity has been reported to be linked with biological motion perception and theory of mind [
56]. Salmond et al. [
55] identified abnormalities in the amygdala and hippocampal region, the orbital frontal cortex, the superior temporal gyri and the cerebellum. Using a related technique, but focusing on GM thickness rather than volume, Hadjikhani et al. [
57] found GM thinning in autism in regions associated with the so-called mirror-neuron system, including pre- and post-central gyri, inferior frontal gyrus, medial frontal gyrus and middle temporal gyrus, all of which were significantly correlated with social and communication deficits in the autism participants. In summary, studies using automated volumetric and/or thickness measurements suggest that autism may be associated with GM differences in distributed cortical and subcortical regions that are important in social cognitive and/or motor processes.
VBM studies in autism have not, however, capitalized on the ability of the method to do whole-brain correlations between GM and symptom measures. In the current study, we applied VBM methods to a well defined sample of patients with autistic disorder with the goal of replicating prior whole brain findings in the literature reported for both gray and white matter. In addition, we examined correlations between GM volume, IQ and autism symptom measures derived from diagnostic interviews. We hypothesized that caudate nucleus volume would correlate with repetitive behaviors in autism, replicating previous hand tracing results. Finally, we hypothesized that social communication deficits would correlate with volumes of structures known to participate in social cognitive processes such as the amygdala, medial frontal gyri and superior temporal sulcus.
Discussion and conclusion
Most, if not all, of the regions in which we report differences between the autism and control groups have been previously associated with social cognitive processes in functional imaging and lesion studies. For example, functional MRI studies of Theory of Mind (TOM: i.e., the ability to make mental state inferences about others) have consistently reported activation of the medial aspect of the superior frontal gyrus, anterior to the region commonly referred to as the pre-supplementary motor cortex, despite using a wide variety of TOM tasks [
72‐
76]. The precuneus region of the medial parietal lobe is also observed to be active in many TOM experiments, may be particularly important in self awareness and is heavily interconnected to the medial prefrontal regions and the dorsal striatum [
77]. The posterior middle temporal gyrus/sulcus region has a role in biologically relevant motion perception, intentionality of eye gaze direction and inferring intentionality from stories, and is often active in functional imaging of TOM tasks [
72,
75,
78‐
80]. Many of these same regions have been reported to be activated differently in people with autism and their first degree relatives while performing TOM tasks, relative to the activation seen in matched controls [
72,
81].
We found no significant correlations between full scale IQ and regional GM volume in our sample. Although this was a study of GM volume, rather than thickness, our findings our consistent with Hadjikhani et al. [
57], who used the latter methodology in a recent study on autism. Taken together, these findings suggest that caution is warranted in using IQ as a covariate in anatomical studies of autism (e.g., see [
18]). However, age was highly correlated with the volume of the hippocampus and surrounding region in our sample, which is consistent with the finding of Good et al. [
67], who first described the optimized VBM approach in the literature and found positive correlations between age and medial temporal lobe volume in large sample of adults when controlling for total brain volume. In our sample, this might be driven by the lower end of the age range. However, only the linear component for age was significant. A study including a larger number of older adults might be able to demonstrate an asymptote on aging effects on the hippocampal region (i.e., a significant quadratic term).
The increased left hippocampal volume in the autism group, relative to the control group, is consistent with the findings reported in Rojas et al. [
2]. This is expected given that almost 60 percent of the autism subjects from this study also participated in that paper, which used hand-tracing methods to assess the hippocampus. As discussed in the introduction, however, both increases and decreases in hippocampal volume have been reported in autism by hand tracing methods. To our knowledge, this is the first report of hippocampal volume change, increase or decrease, in autism using VBM methodology. A recent paper by Testa et al. [
82] comparing hand-tracing to VBM for detecting hippocampal atrophy in Alzheimer's Disease provided data supporting the comparability of the two methods for hippocampal volumetry.
To our knowledge, this is the first VBM study in autism to utilize symptom measures from the ADI-R and ADOS-G to examine correlations with both cortical and subcortical GM. Hadjikhani et al. [
57] performed such correlations, but only with regional cortical thickness, not with subcortical structural measures. Our significant positive correlation between caudate nucleus volume and the ADI-R Repetitive and Stereotyped Behavior domain is supportive of the same association reported in the manual tracing study of Hollander et al. [
36]. Caudate enlargement has also been reported in Obsessive-Compulsive Disorder [OCD: 83]. Interestingly, caudate nucleus activity in OCD has been reported to be higher in both resting and symptom provocation conditions in functional neuroimaging studies [
84,
85], and treatment with either serotonin agonist medications or behavior therapy have been shown to reduce this over-activity [
86,
87].
An alternative, but not necessarily mutually exclusive, interpretation of the caudate nucleus correlation is that it simply reflects disorder severity generally. In support of this notion is that the caudate nucleus volume was significantly associated with both with the ADI-R Social and Communication total score and with the Repetitive and Stereotyped Behaviors score. Arguing against this interpretation is the lack of correlation observed in these same structures with the ADOS-G Severity measure. These other structures, along with the superior temporal gyrus and cerebellum, have been heavily implicated in social cognitive and motor abilities. Taken together, the GM findings and the symptom correlations suggest that the regional GM differences are highly relevant to the behavioral phenotype in autism.
Some caution is warranted with respect to the caudate nucleus findings, however. Three of our autism participants had a past history of neuroleptic medication treatment, use of which has been previously associated with caudate volume increases in schizophrenia [e.g., [
88,
89]]. However, all of three of these were taking atypical antipsychotic agents and discontinued use of the drugs at least 2 years prior to participation in the study. Studies have suggested that caudate volume increases to neuroleptic treatment are associated more strongly with typical neuroleptics and that volumes decrease rapidly after neuroleptic treatment is withdrawn [
90,
91]. It therefore seems unlikely that the caudate volume enlargement in our sample could be explained by treatment by neuroleptic medications.
There are several limitations to the results of this study to consider. Among them, the non-isotropic voxel dimensions of the MRI scans may have introduced partial volume effects that influenced the tissue segmentation step of the VBM procedure. Additionally, the large age range of participants could have introduced variability that reduced the effect sizes for some structures we hypothesized would differ between groups (e.g., the amygdala), but for which we did not find group differences. We attempted to control for age effects statistically, by treating the effects of age as covariates in the main analysis. This may not be the most optimal design, however, and further studies that more tightly control the age range of the samples are warranted. Third, segmentation algorithms such as the one employed here on T1 weighted MRI data generally do not discriminate well between deep cerebellar nuclei and the surrounding cerebellar white matter (most of the nuclei are classified as white matter). Fourth, the lack of inclusion of female participants, while advantageous in terms of sample homogeneity, limits the generalizability to males with autism. Finally, the IQ of the control participants (mean = 119), while not strictly a limitation, should be mentioned, because it is worth noting that previous population-based studies have indicated a higher mean IQ level in the Denver metropolitan region than in the national standardization sample for the Wechsler scales [
92].
Competing interests
The author(s) declare that they have no competing interests.
Authors' contributions
DR conceived of the study, made the main contribution to the design and analysis, performed all VBM analyses and drafted the manuscript. EP, EW, MR and JT participated in the design of the study. SR made a substantial contribution to the selection and use of the clinical measures reported in the study.