By Lauren Sakowski
The development of neuroimaging-based biomarkers that enable clinicians to more accurately diagnose psychiatric disorders – and to tailor individual treatment regimens accordingly – is one of the most urgent and elusive goals facing neuroscience researchers today. A recent post on this site discussed the promise and challenges of these efforts.
Some previous attempts to use neuroimaging for psychiatric diagnosis have been insufficient for clinical applications, because they relied only on imprecise measures such as the overall volume of a particular brain region.
More recent investigations, including the 2012 study published by Dr. Bradley Peterson’s group, Anatomical Brain Images Alone Can Accurately Diagnose Chronic Neuropsychiatric Illnesses, have employed MRI-based algorithms, which use local variation in different morphological brain features that are sensitive to differences between healthy individuals and patients with diagnosed psychiatric illnesses. Peterson and colleagues found that this technique can be used to diagnose chronic psychiatric illness in adults and children with Tourette Syndrome, ADHD, Bipolar Disorder, Schizophrenia, and high or low risk for Major Depressive Disorder when compared to healthy individuals.
In this post, Lauren Sakowski discusses these findings and talks to Dr. Peterson about their implications and developments in the field since 2012. – PLOS Neuro
Two 2010 studies reported in the Journal of Adolescent and Child Psychiatry can be seen as laying important groundwork for Peterson’s findings. One study examined the volume of the amygdalae and hippocampi in adolescents with autism using segmentation in MRI and detected abnormal enlargement in patients. Another study manually delineated the basal ganglia region to correlate changes in shape with motor and praxis deficits in boys with autism spectrum disorder. These attempts resulted in poor sensitivity and specificity, which caused significant overlap between diagnostic groups.
Study design and findings
In their 2012 study, Ravi Bansal, Dr. Peterson, and colleagues employed surface morphometry, a technique that constructs and analyzes surfaces and structural boundaries in the brain, to delineate a region of the brain independently from other regions. This technique then registers points on the surface of each region to corresponding points of a template brain. Figure 1 illustrates deformations in the dorsolateral prefrontal cortex (DLPFC) or occipital cortex (OC) to the template brain, which was generated using synthetic data, first from a single brain, then from 20 individuals.
Deformed brains were normalized to undeformed templates, allowing for classification of each brain into four groups: protrusions in DLPFC, indentations in DLPFC, protrusions in OC, or indentations in OC. Transformations were applied to compute scaling coefficients and groupings were identified using hierarchical clustering, which identifies natural structure within a dataset. Natural groupings were identified in a dataset generated from identical brains with different deformations as well as in brains from different individuals. Synthetic datasets were also constructed for application with subsequent human data. Scaling coefficients at higher spatial resolutions were more accurate at encoding deformations and minimizing variation surface morphology than at lower spatial resolutions, and were therefore better suited to use in combination with the cluster groups to analyze human datasets.
Imaging datasets were acquired from healthy participants as well as adults and children with Tourette Syndrome (TS), ADHD, Bipolar Disorder I (BD), Schizophrenia (SZ), and high or low risk for Major Depressive Disorder (MDD) using structural MRI. The classification system was used to differentiate between healthy adults or children and adults or children diagnosed with the aforementioned psychiatric disorders, as well as between individuals with different disorders. Figure 2 provides a select example of the human datasets and delineates the patterns of surface features (grey regions) of dorso-anterior vs. ventro-posterior regions of the right hippocampus in healthy adults vs. those with TS, and the right globus pallidus and hippocampus in healthy children vs. those with TS. Additionally, sensitivity and specificity between disorders are illustrated in the clustered dendrograms at the bottom of Figure 2. These data illustrate the fact that sensitivities tend to be higher than specificities when classifying groups, potentially due to brain features associated with a disorder that never result in illness. Discrimination between healthy individuals and those with disorders, as well as between different disorders, showed a positive predictive value of close to 1.0 in most cases, indicating the classification system’s efficacy in correctly identifying the illness. While several different regions of the brain were affected in different disorders, the measures most commonly affected by disease were amygdala and hippocampus morphology, as well as cortical thickness.
This study is the first of its kind to use neuroimaging for diagnosis of psychiatric illness with such a high degree of accuracy. A major advantage is the use of a technique that examines each surface independently from remote surfaces. Because region delineation is done manually, as opposed to automated methods used in previous studies, this technique requires a great deal of time for training and validation, which is a limit for clinical applicability. Misclassification rates were higher when attempting three-way discrimination, as in healthy versus SZ versus BD adults. To maintain accuracy, physicians would therefore need to use the classification algorithm to differentiate between two disorders, rather than attempting to determine from a list.
In a recent interview, I addressed some of the study’s implications and limitations with Dr. Bradley Peterson.
LS: How has the ability to diagnose psychiatric illnesses using only imaging techniques progressed since you published this paper ?
“We have been assessing whether the addition of information from other MRI modalities, such as functional MRI, diffusion tensor imaging, and MR spectroscopy, can improve performance of our diagnostic algorithms. Our findings suggest that information from these other modalities can improve diagnostic performance, although the improvement is modest because of the already-excellent performance that anatomical MRI using accurately defined brain regions provides.”
LS: Has a more automated tool been developed for brain region delineation in the last two years?
“Numerous software platforms exist that delineate brain regions in an automated way, but whether those platforms produce delineations that are sufficiently accurate to be useful in our classification algorithms is unknown. We have a funding application under review that proposes, among other things, to answer that question. Our current belief based on our considerable experience with those platforms, however, is that the accuracy of delineations for particularly challenging regions, such as the basal ganglia, thalamus, amygdala, and hippocampus, currently is insufficient for diagnosis using algorithms that relay information from those regions.”
Any views expressed are those of the author, and do not necessarily reflect those of PLOS.
Groen, W. et al (2010). “Amygdala and hippocampus enlargement during adolescence in autism,” Journal of the American Academy of Child and Adolescent Psychiatry 49 (6): 552-560. doi:10.1016/j.jaac.2009.12.023
Qui, A. et al (2010). “Basal Ganglia Shapes Predict Social, Communication, and Motor Dysfunctions in Boys With Autism Spectrum Disorder,” Journal of the American Academy of Child and Adolescent Psychiatry 49 (6): 539-551. doi:10.1016/j.jaac.2010.02.012
Bansal, R. et al (2012). “Anatomical brain images alone can accurately diagnose chronic neuropsychiatric illnesses,” PLoS ONE 7 (12): e50698. DOI: 10.1371/journal.pone.0050698
Lauren Sakowski is a molecular biologist/neuroscientist at the Nemours/A.I. duPont Hospital for Children studying the role of inflammation in neurodegeneration in a leukodystrophy mouse model. https://twitter.com/LaSaks87