Logan Young
Staff Writer
Published March 21, 2020
To be sure, radiology has come a long, long way. Only 10 years ago, the best medical imaging could do for children with autism spectrum disorder (ASD) was to identify key abnormalities in the brains of those already diagnosed—i.e., 1 in 59 children, according to today’s estimates from the Centers for Disease Control and Prevention’s (CDC) Autism and Developmental Disabilities Monitoring Network. A half-decade earlier, cortical gray-matter studies were discovering overall substantially thicker cortex for boys with autism, alongside similar findings in the temporal and parietal lobes, whereas diffusion tensor imaging was being used to illustrate disruption of white-matter tracts between regions implicated in impaired social cognition. Meanwhile, just as early functional MRI (fMRI) studies on ASD were exploring core symptom domains via activation patterns in response to mimesis, facial processing, theory of mind, semantic sentence comprehension, lexical semantic processing, and tasks involving variable imagery content, researchers were also looking to magnetic resonance spectroscopy (MRS) to assess models regarding excitation and inhibition ratios in ASD.
Writing on MRS in the October 2004 issue of the Journal of Neuroscience, Matthew K. Belmonte from the Autism Research Centre at the University of Cambridge duly noted: “It has been said that people with autism suffer from a lack of ‘central coherence,’ the cognitive ability to bind together a jumble of separate features into a single, coherent object or concept. Ironically, the same can be said of the field of autism research, which all too often seems a fragmented tapestry stitched from differing analytical threads and theoretical patterns”.
Fifteen years removed, while ASD remains very much an heterogeneous disorder of multifactorial etiology, evidencing an expansive range of symptoms and severities alike, radiology is in the process of reconciling so many image threads. True, bereft of a priori behavioral phenotyping (e.g., Autism Diagnostic Observation Schedule [ADOS], Social Responsiveness Scale, Kaufman Brief Intelligence Test, composite IQ score), right now, radiology alone still cannot definitively diagnose ASD in anyone, child or adult. There is good news, though. The radiology research paradigm is shifting—away from mere aberration identification to clinical diagnosis.
The sands underneath it all first loosened in 2014, when University of Pittsburgh and Carnegie Mellon researchers utilized machine-learning algorithms to grade 34 young adults as either autistic or control with > 97% accuracy based upon fMRI neurocognitive markers for eight social interaction verbs: compliment, insult, adore, hate, hug, kick, encourage, and humiliate. Moving quickly, one year later, Virginia Tech Carilion Research Institute professor P. Read Montague synthesized nine years’ worth of previous trials to announce in Clinical Psychological Science that his team had developed an even more efficient technique to diagnose children with ASD in under two minutes: single-stimulus fMRI. Subjects were shown 15 images of themselves and 15 images of another child, matched according to age and gender, for four seconds per image in randomized order. Like the control adults in Montague’s earlier experiments with imaging for ASD, when viewing their own pictures, the control children had a high response in the middle cingulate cortex; by contrast, children with ASD showed an appreciably diminished reaction. Notably, Montague et al. could detect this disparity using one, solitary image.
This May, much of Montague’s same colleagues, including principal investigator, Kenneth Kishida of the Wake Forest School of Medicine, made headlines for a Biological Psychology article demonstrating that a single stimulus and < 30 seconds of fMRI data were sufficient to differentiate ASD children from their typically developing (TD) peers. To test a hypothesis that responsiveness of the brain’s ventral medial prefrontal cortex (vmPFC) in children diagnosed with ASD is diminished for visual cues, denoting high-value social interaction, 40 participants (of which 12 had ASD and 28 were TD), aged 6–18 years old, were prompted to observe images of four faces and four objects, which were projected onto a screen and viewed through a mirror during fMRI scanning. With each image characterized as favorite, pleasant, neutral, or unpleasant, the favorite images depicted each of the participants’ self-selected favored face and object, and the remaining images were selected from the International Affective Picture System (IAPS) database. Each of the eight images was then displayed only once for five seconds during a block that repeated six times. Following the completion of 12- to 15-minute MRI scans, participants were shown the identical set of images on a computer screen, ranking them in order, from pleasant to unpleasant, with a self-assessing sliding scale. Results showed that the average response of vmPFC was significantly lower in the ASD cohort, compared to the TD cohort.
“How the brain responded to these pictures is consistent with our hypothesis that the brains of children with autism do not encode the value of social exchange in the same way as typically developing children,” Kishida said in a prepared statement. “Based on our study,” he continued, “we envision a test for autism in which a child could simply get into a scanner, be shown a set of pictures, and within 30 seconds, have an objective measurement that indicates if their brain responds to social stimulus and non-social stimuli.”
There are limitations here. Because these 40 children were permitted to specify favored objects and people, reasonably assuming that there were distinct visual differences between these non-IAPS images and that canonical cache, Kishida conceded the possibility that at least some of the reported response differential could simply be due to known vs. novel. Moreover, since ASD disproportionately affects male patients—i.e., four times more common among boys than girls, the CDC maintains—he acknowledged an optimal design could be updated to investigate the gender divide between the ASD and the TD children more thoroughly.
“Based on our study, we envision a test for autism in which a child could simply get into a scanner, be shown a set of pictures, and within 30 seconds, have an objective measurement that indicates if their brain responds to social stimulus and non-social stimuli.”
—Kenneth Kishida
Another Wake Forest faculty member, Christopher T. Whitlow, has been presenting related research on ASD imaging since 2014. As his studies have surveyed patterns of joint variability in severely preterm infants, might we see an eventual diagnostic environment where Whitlow’s voxel-based morphometry informs Kishida and Montague’s single-stimulus exemplar to evidence brain dysfunction in patients younger than the age-six threshold?
Although reproductive stoppage (i.e., the tendency for arrested propagation after diagnosis of an affected child) can lead to underestimates of sibling recurrence risk for ASD, with ascertainment biases and overreporting often pointing to its inflation, we should focus on the family first. In 2011, the multisite international network, Baby Siblings Research Consortium, conducted a prospective longitudinal study of 664 infants who had an older biological sibling with ASD, monitoring them from early life to 36 months, when they were classified as having or not having ASD—an ASD taxonomy requiring exceeding the ADOS cut-off, as well as an expert’s diagnosis. In total, 18.7% of infants developed ASD. Whereas infant age at enrollment, gender and functioning level of the infant’s older sibling, and other demographic circumstances did not predict ASD outcome, infant gender and the presence of > 1 older affected sibling were significant forecasters. Again, there was a nearly threefold risk escalation for male subjects and an additional twofold increase in risk if there was > 1 older affected sibling.
Family history, meet deep learning. Recent findings published in Science Translational Medicine by University of North Carolina at Chapel Hill researchers revealed that when applied to functional connectivity MRI (fcMRI) data at six months of age in infants with high familial risk for ASD, a nested, cross-validated machine-learning algorithm predicted an ASD diagnosis with > 96% accuracy at 24 months. Citing several brain variances—both morphological and electro-physiological—members of his team had documented as early as six months in infants later diagnosed with ASD, “Given the complexity and heterogeneity of ASD,” lead author Robert W. Emerson surmised, “methods for the early detection of ASD using brain metrics will likely require information that is multivariate, complex, and developmentally sensitive.” Apropos, Emerson et al. employed an array of 230 regions of interest (ROI) previously defined across the entire brain to create functional connectivity matrices from the fMRI scans of 59 at-risk infants (11 diagnosed with ASD at 24 months, 48 who did not have ASD at 24 months) during natural sleep without sedation at their six-month visit. “Our logic was that these regions would be the most likely to contribute to the discrimination between groups in the 59 separate support vector machine models,” wrote Emerson. With data collection resulting in 26,335 usable ROI pairs exemplifying each infant’s whole-brain functional constitution by training MATLAB’s Statistics and Machine Learning Toolbox (Mathworks, Inc.) to ascertain the causal patterns of individual separation, the probability that infants with a positive classification truly had ASD (positive predictive value) at 24 months was 100% (95% CI, 62.9–100). Negative predictive value at 24 months was 96% (95% CI, 85.1–99.3).
A first-of-its-kind study from November 2018 that leveraged the imaging archive of Geisinger Health System in Danville, Pennsylvania, takes us back to the future, examining early brain markers in ASD to further the promise of artificial intelligence for earlier detection. Renewing his dissertation research, Gajendra J. Katuwal and colleagues applied random forest ensemble learning to models trained on 687 brain features of Freesurfer v 5.3.0 (Martinos Center for Biomedical Imaging) to compare cortical and sub-cortical morphometric features for ASD vs. non-ASD classification. Their query of head MR images from Geisinger’s institutional tranche, after removing those with artifacts, motion, lesions, abnormally large ventricles, and neurodevelopmental disorders as identified by International Classification of Diseases code, yielded 112 non-ASD and 115 ASD subjects. Eschewing gender confounds, 20 non-ASD and 34 ASD scans of female subjects were excluded. Although total intracranial volume (TIV) of ASD measured 5.5% larger than the control, brain volumes of other ROI, when calculated as TIV percentage, measured smaller in ASD—partially due to larger (> 10%) ventricles in ASD. ASD’s larger TIV exhibited correlates with greater surface area and aggregate cortical folding, yet not with cortical thickness. ASD frontal and temporal white-matter tracts evidenced less image intensity, seemingly suggesting myelination deficit. Ultimately, Katuwal’s methodology was able to achieve 95% AUC for ASD vs. non-ASD classification using all brain features. When stochastic discrimination was discrete for each feature type, image intensity yielded the highest predictive power (95% AUC), followed by cortical folding index (69%), cortical and subcortical volume (69%), and surface area (68%).
According to Katuwal, “the most important classification feature was white matter intensity surrounding the rostral middle frontal gyrus,” which measured lower (d = 0.77, p = 0.04) in ASD.
Because medical technology also rises, medical imaging, itself, is sure to manifest a more prominent role over time among allied sciences with regards to forthcoming ASD diagnoses and concomitant, personalized care. To that end, in order to fully apprehend the neuroanatomical foundations of ASD, a comprehensive, multimodal surveillance of early brain alterations would seem to light the best forward path. Progress isn’t always a straight line, of course, so radiology has places yet to go, indeed.
The opinions expressed in InPractice magazine are those of the author(s); they do not necessarily reflect the viewpoint or position of the editors, reviewers, or publisher.
