Structural Associations of Cortical Contrast and Thickness in First Episode Psychosis

Abstract

There is growing evidence that psychosis is characterized by brain network abnormalities. Analyzing morphological abnormalities with T1-weighted structural MRI may be limited in discovering the extent of deviations in cortical associations. We assess whether structural associations of either cortical white–gray contrast (WGC) or cortical thickness (CT) allow for a better understanding of brain structural relationships in first episode of psychosis (FEP) patients. Principal component and structural covariance analyses were applied to WGC and CT derived from T1-weighted MRI for 116 patients and 88 controls, to explore sets of brain regions that showed group differences, and associations with symptom severity and cognitive ability in patients. We focused on 2 principal components: one encompassed primary somatomotor regions, which showed trend-like group differences in WGC, and the second included heteromodal cortices. Patients’ component scores were related to general psychopathology for WGC, but not CT. Structural covariance analyses with WGC revealed group differences in pairwise correlations across widespread brain regions, mirroring areas derived from PCA. More group differences were uncovered with WGC compared with CT. WGC holds potential as a proxy measure of myelin from commonly acquired T1-weighted MRI and may be sensitive in detecting systems-level aberrations in early psychosis, and relationships with clinical/cognitive profiles.

Introduction

The first episode of psychosis (FEP) is a critical stage for potential biomarker discovery in psychotic disorders. This is because of a relative freedom from a large number of confounders that can reduce the reliability of investigations in patients with a longer history of psychosis (e.g., longer exposure to medications, social isolation, etc.). CT has often been chosen to study brain morphology in FEP (Wiegand et al. 2004; Narr et al. 2005; Crespo-Facorro et al. 2011; Scanlon et al. 2014; Pina-Camacho et al. 2015; Gong et al. 2016); however, this metric may be limited in capturing the full extent of pericortical abnormalities in FEP, emphasizing the need for novel approaches to analyze structural magnetic resonance imaging (MRI) in patients.

There has also been a rising interest in studying systems-level brain organization, given the well-supported theory of schizophrenia and related psychoses as disorders of dysconnectivity (Davis et al. 2003; Stephan et al. 2009; Haroutunian et al. 2014; Friston et al. 2016; Kelly et al. 2017). Recent methods using structural covariance have underlined the possibility of looking at brain organization in both normal development/ageing and disease using commonly acquired structural T1-weighted scans (Bullmore and Sporns 2009; Guye et al. 2010; Raznahan et al. 2011; Alexander-Bloch et al. 2013; Evans 2013; Wheeler and Voineskos 2014). Structural covariance is thought to capture similarities between brain regions which may be influenced by common genetic, environmental, and/or other biological factors. Measures of gray matter morphology, such as CT and volumes, have often been used to define cortical structural covariance. There is great interest in extending these methods to better understand differences in brain organization underlying psychotic disorders. A relevant review (Wheeler and Voineskos 2014) identified 16 studies published between 1992 and 2014 using structural covariance to analyze data from patients with schizophrenia; almost all studies used gray matter volumes in covariance analyses, with the exception of one study using CT. Since then, several other studies have incorporated CT in investigations of structural covariance in psychosis (Wheeler et al. 2015; Guo et al. 2016; Buchy et al. 2017; Kuang et al. 2017). Another recent study has extended this method to explore cortical folding connectomics in FEP patients (Palaniyappan et al. 2016). Although such gray matter-based covariance networks have the potential to identify meaningful structural relationships, interpreted as putative “connections” in previous literature, it has also been acknowledged that CT has relatively weaker correspondence with gold standard connectivity data compared with other imaging modalities (Reid et al. 2016).

MRI-derived CT relies on the tissue intensity contrast between gray and white matter, and gray matter and cerebral spinal fluid in T1-weighted structural MRI. However, it has often been overlooked that the placement of the gray–white matter boundary on MRI is driven by the myeloarchitecture both intracortically and within superficial white matter (Rowley et al. 2015). Obtaining a measure of WGC from T1-weighted MRI may provide a meaningful marker of myelin content and other biophysical properties that may complement measures of CT, as well as confer sensitivity in detecting subtle group differences (Bezgin et al. 2018; Lewis et al. 2018).

A few studies have examined WGC in enduring schizophrenia (Bansal et al. 2013; Jorgensen et al. 2016), finding increased contrast in primary sensory regions. Given that primary sensory regions are characterized by high levels of intracortical myelin in HC (Glasser and Van Essen 2011), it has been posited that increased contrast in these regions in schizophrenia may be a result of changes in myelin. It should be noted that alterations in myelin in schizophrenia and the related psychosis have been found both intracortically (Bartzokis et al. 2009; Uranova et al. 2011; Bartzokis 2012; Lake et al. 2017) and within superficial white matter (Nazeri et al. 2013). However, none of these studies examined patients in early stages of psychosis, where patients are less affected by factors underlying chronicity of illness, such as the effects of antipsychotic medication on myelin (Bartzokis et al. 2009, 2011). Given that WGC may capture myelination properties within the cortex and superficial white matter, it is feasible that this metric may be more sensitive in capturing the extent of disruptions in the coordinated development and plasticity of cortico-cortical connections in early psychosis compared with CT.

To explore relationships between brain regions with white–gray matter contrast (WGC) and cortical thickness (CT) in FEP patients, we applied the following: (1) dimension reduction techniques to extract information reflective of known microstructural (e.g., pericortical myelin) composition in the brain (Hopf 1955; Glasser and Van Essen 2011; Pandya et al. 2015; Margulies et al. 2016), and whether such patterns are different between FEP patients and HC; (2) examination of whether these patterns relate to symptoms and cognitive deficits in FEP patients; and (3) comparison of structural covariance measurements between FEP and controls using WGC and CT. We hypothesize that multivariate dimension reduction of WGC and CT data will cluster brain regions that share similar architectonics (Glasser and Van Essen 2011; Rowley et al. 2015; Margulies et al. 2016), and these dimensions will be altered in FEP patients compared with controls, with relationships to psychopathology severity in FEP. We also expect structural covariance analyses with WGC data to yield more distributed brain changes in patients compared with structural covariance alterations obtained with CT.

Materials and Methods

Patients were recruited from the Prevention and Early Intervention Program (PEPP) at the Douglas Institute in Montreal, Canada, and were part of a longitudinal naturalistic outcome study. Details of PEPP are outlined elsewhere (Iyer et al. 2015). Inclusion criteria at PEPP include a diagnosis of affective or non-affective psychosis, IQ > 70, and limited (<1month) to no previous exposure to antipsychotic medication. Patients recruited to PEPP (ages 18–35) were invited to take part in a neuroimaging study, comprising 3 timepoints (baseline, 1-, and 2-year follow-ups) as described in our previous work (Makowski et al. 2016, 2017). Non-clinical healthy controls (HC) were recruited through advertisements within the same local catchment area. All participants provided written informed consent, and the research protocol was approved by the Douglas Institute human ethics review board. Of the 150 FEP patients that were recruited for the study, 116 patients (male, N = 82) and 88 HC (male, N = 57) were included in this analysis. See Methods in the Supplement for more detailed information on the final included sample and demographic/clinical data collection.

MRI Acquisition

Scans were collected at the Montreal Neurological Institute on a 1.5-Tesla Siemens Sonata whole body MRI system. Structural T1-weighted volumes were acquired for each participant using a 3D gradient echo pulse sequence with sagittal volume excitation (resolution = 1 mm³, repetition time = 22 ms, echo time = 9.2 ms, flip angle = 30°, 180 1 mm contiguous sagittal slices). The rectangular field of view (FOV) for the images was 256 mm (AP) by 204 mm (SI).

Post-Processing—CT

Raw T1-weighted images were submitted to the CIVET pipeline (Version 2.1.0: http://www.bic.mni.mcgill.ca/ServicesSoftware/CIVET) (June et al. 2005) for extraction of gray and white matter surfaces. Main processing steps include (1) registration of T1-weighted images to standardized space (Collins et al. 1994) and correction for non-uniformity artefacts (Sled et al. 1998); (2) segmentation of gray, subcortical gray and white matter, and cerebral spinal fluid (Zijdenbos et al. 2002; Tohka et al. 2004); (3) extraction of the white matter surface using a marching-cubes algorithm and extraction of the gray matter surface using the CLASP algorithm (Kim et al. 2005); (4) surface registration to a template for inter-subject correspondence (Lyttelton et al. 2007); (5) reverse transformation (initially done in step 1) to estimate CT in native space for each subject at 81 924 vertices using the t-laplace metric (Lerch and Evans 2005); and (6) smoothing the data in native space with a 20 mm FWHM Gaussian kernel to diminish the impact of noise (Boucher et al. 2009).

Post-Processing—WGC

WGC was estimated using a recently published method from our group (Lewis et al. 2018). In detail, a distance map relative to the white surface provided by CIVET was created at 0.25×0.25×0.25 mm resolution, smoothed with a 0.5-mm FWHM Gaussian kernel, and used to create a gradient vector field of the distance map. A copy of the white surface was moved 1 mm inward along this gradient vector field to produce a sub-white surface, and 1 mm outward to produce a supra-white surface. The intensity values on the T1-weighted image (without non-uniformity correction or normalization) were sampled at each vertex of both the supra-white surface and the sub-white surface, and the resulting surface maps were smoothed with a 20 mm FWHM Gaussian kernel. Smoothing was done prior to calculation of the contrast measure to diminish the impact of noise, which could potentially be amplified if the ratio was calculated first. The WGC measures were then formed by dividing the value at each vertex of the sub-white surface by the value at the corresponding vertex of the supra-white surface. Quality control procedures are outlined in the Supplementary Material. The WGC method is also depicted in Supplementary Figure 1, and the mean distance between the sampled supra- and sub-white surfaces to the gray–white matter boundary across groups is depicted in Supplementary Figure 2.

Multivariate Neuroimaging Analyses

Here, we aimed to assess whether group differences exist between patterns of brain regions that share similar features, such as myeloarchitecture, and whether these patterns relate to psychopathology and cognition in FEP. Principal component analysis (PCA) was applied to the WGC and CT data across 81 924 vertices of the cortical surface to reduce data dimensionality. In addition to being a common approach to reduce dimensionality in large datasets, PCA allows for observation of potential patterns among sets of brain regions that may share common information, without fitting a specific model a priori. The “SurfStatPCA” function (http://www.math.mcgill.ca/keith/surfstat/doc/SurfStat/SurfStatPCA.html) was applied to WGC and CT values for the entire baseline sample included in the analysis (116 FEP, 88 Controls) at each of 81 924 vertices per subject, after regressing out age and sex. A scree plot was generated to choose the number of components to interpret. Relationships between selected component loadings and the following behavioral measures in FEP patients were assessed with Pearson r correlations, and corrected for multiple comparisons with a Bonferroni correction: (a) positive and negative symptoms, as assessed by the SAPS and SANS, respectively; (b) general cognitive index (GCI), excluding social cognition; and (c) 3 measures tapping into higher-order cognitive processes: verbal memory, executive function, and working memory. These measures were chosen a posteriori to test specific hypotheses about relationships with sets of brain regions uncovered in PCA components, as discussed in the results. See Supplementary Methods and Supplementary Table 2 for more information of neuropsychological tests used to define cognitive measures listed in (b) and (c), and how non-normality of symptom data was handled. Several supplementary analyses with PCA were run to (a) investigate whether similar results would be obtained when retaining only patients with first-episode schizophrenia spectrum (FES) disorder (i.e., excluding patients with affective disorder, delusional disorder, or psychosis not otherwise specified, leaving N = 81 patients), (b) examine whether additional potentially important clinical factors, namely duration of untreated psychosis and antipsychotic medication exposure, were associated with component scores in FEP patients, and (c) to investigate patterns of brain regions when running PCA on FEP and healthy control groups separately. For analysis (b), 9 patients had missing data assessing duration of untreated psychosis, and one other patient was excluded as a clear outlier (duration of untreated psychosis was >1000 weeks, which is plausible as this reflects approximately 19 years, but reflected a very large deviation from the remainder of the data). The remaining data for duration of untreated psychosis (N = 106) was then log-transformed to more closely approximate a normal distribution. Similarly, for antipsychotic medication exposure, measured as chlorpromazine equivalents, weighted by percentage adherence, the data were non-normally distributed and a square root transformation was applied to normalize the data. Analysis (c) naturally does not allow us to look at group differences, but allows for a visualization of patterns of brain regions that may have similar cyto-/myelo-architectonics and may diverge between patients and controls.

Structural Covariance Neuroimaging Analyses

To test for group differences in structural covariance, we adopted methods used previously by our group (He et al. 2007; Khundrakpam et al. 2017). First, the Automated Anatomical Labeling (AAL) atlas (Tzourio-Mazoyer et al. 2002) was used to parcellate the brain for each subject into 78 regions. Mean WGC and mean CT were calculated for each region per subject. Next, residuals were computed for each region, where a linear regression was applied to remove the effects of age and gender, as well as mean WGC or mean CT, for respective analyses. These residuals were used to compute all possible pairwise correlations, excluding self-correlations (i.e., [78*78]/2–78/2), for FEP patients and HC separately. Correlation coefficients were compared between groups by converting Pearson r correlation coefficients to z-scores, using a Fisher transformation, and comparing the differences to a normal distribution to obtain P-values. P-values were thresholded using a 2-stage false discovery rate (FDR) correction, which limits the false-positive rate for a family of hypothesis tests below 0.05 (Benjamini et al. 2006). This approach was used over the traditional FDR correction (Benjamini and Hochberg 1995), as it has been shown to better account for both independent and positively dependent correlations. Only the lower triangle of the thresholded matrices (i.e., non-duplicate pairs of regions) were considered when reporting significant pairwise correlations. Therefore, for the AAL parcellations, a total of 3003 tests were conducted (i.e., [78*78]/2–78/2). Identical analyses were also performed using the Desikan Killany–Tourville (DKT40) atlas to parcellate the brain into 62 regions (Klein and Tourville 2012). Structural covariance analyses were also run including only first-episode schizophrenia spectrum patients (N = 81), as described in the previous section.

Results

Sample

The final sample comprised 116 FEP and 88 HC. See Table 1 for demographic and clinical information, as well as baseline group differences in cognitive data. Patients did not differ from controls in age or sex ratio. However, patients had significantly less years of education, and worse performance on cognitive domains (general cognitive index, executive function, working memory, verbal memory; P’s < 0.05) compared with controls.

Table 1.

Demographic and clinical information

	FEP N = 116	HC N = 88
	Mean (SD)	Mean (SD)
Age	24.0 (3.9)	24.4 (3.3)
Education in years*	11.9 (2.5)	14.3 (2.4)
General cognitive index*	−0.7 (0.8) [114]	0.03 (0.6) [82]
Executive function*	−0.7 (1.12)	0.04 (0.7)
Working memory*	−0.6 (0.9)	0.002 (0.8)
Verbal memory*	−1.1 (1.2) [115]	−0.03 (0.9) [83]
Positive symptoms	9.9 (12.4)
Negative symptoms	19.3 (12.6)
Duration of untreated psychosis (weeks)	70.3 (130.4) [107]
Duration of untreated illness (years)	6.9 (6.1) [110]
% time in positive symptom remission	40.4 (38.9) [113]
Chlorpromazine equivalents	804.7 (734.5)
Antipsychotic medication adherence	83.2 (27.0)
	N (%)	N (%)
Male	82 (71)	57 (65)
Right handed	96 (83)	80 (91)
Other medication
Antidepressant	20 (17.7) [113]
Benzodiazepine	9 (8.0) [113]
Anticholinergic	11 (9.7) [113]
Mood stabilizer	10 (8.8) [113]
Diagnosis
Schizophrenia/schizophreniform	81 (69.8)
Affective disorder	21 (18.1)
Delusional disorder	3 (2.6)
Psychosis not otherwise specified	10 (8.6)
Not known	1 (0.9)

Mann–Whitney U tests were used to test for significant differences between groups for non-normally distributed variables (i.e., age and education in years). Chi-squared tests were used for categorical variables (i.e., sex and handedness). An ANCOVA was applied to test for group differences between cognitive measures, covarying for neuropsychological test battery. Significant group differences are marked by an asterisk (*), where FEP patients had significantly less years of education compared with HC, and had significantly lower performance on cognitive measures (P < 0.05). Note that cognitive measures are presented as z-scores compared with controls. The mean and standard deviation for controls do not equate exactly to 0 and 1, as the scores were calculated for a larger subset of controls part of a larger study. Positive and negative symptoms reflect the sum of item-level scores (excluding globals) for the Scale for the Assessments of Positive and Negative Symptoms (SAPS, SANS, respectively). Percent of time spent in positive symptom remission reflects the proportion of time the patient spent between entry to the clinic and their first MRI scan without active positive symptoms. All antipsychotic totals are presented as chlorpromazine equivalents in mg, as prescribed by a psychiatrist, and are reported along with a percentage of medication adherence. Note, the “Affective” diagnostic category includes patients with bipolar disorder and major depression with psychotic features. Data are presented for 116 FEP Patients and 88 Controls, unless otherwise indicated by sample size presented in square brackets.

PCA Results with WGC

The first 3 components from PCA collectively explained 37.27% of the variance in WGC data; from component 4 onwards, each component contributed relatively little and was not further interpreted (see Supplementary Figure 3A for scree plot). Cortical surface maps of the PCA loadings for the 3 selected components are depicted in Figure 1. As can be expected from PCA, the first component, explaining 25.2% of variance, reflected mean WGC signal, with weak positive vertex-wise loadings across widespread brain regions (Supplementary Figure 4), and was significantly and positively associated with the general cognitive index in FEP patients (r = 0.27, P = 0.004). The second component, explaining 9% of the variance, captured positive loadings within unimodal sensory and motor regions (i.e., primary visual and sensorimotor cortices). The third component, explaining 3.1% of the variance, encompassed the superior temporal sulcus (including the auditory cortex, with stronger loadings on the left hemisphere) and heteromodal cortical regions, with negative loadings of prefrontal cortices, and positive loadings of posterior temporal/lateral occipital cortices, fusiform gyri, and precuneus bilaterally. There were no significant group differences when comparing component scores between FEP and controls for any of these 3 components, although component 2 showed a trend-like difference (F₁₂₀₂ = 3.58, P = 0.060), where patients had higher component scores than controls. This suggests that patients had marginally higher contrast in primary sensory (i.e., visual, somatosensory) and motor regions compared with controls.

Figure 1.

Top 3 components from PCA with WGC data and related variables. Cortical surface maps reflect component loadings for each principal component (PC). PC1 reflects mean WGC signal (see Supplementary Fig. 3). PC2 captures primary sensory and motor regions, that are associated with positive and negative symptoms (Plot A) and a general cognitive index (GCI; Plot B) in FEP patients. PC3 captures bilateral superior temporal sulci (albeit with stronger loadings in the left hemisphere) and heteromodal cortical regions, associated with verbal memory performance in FEP patients (Plot C). Pearson r-correlations are presented for FEP patients and controls for significant findings, where applicable. Association of principal component 2 with both positive and negative symptoms are presented together in Plot A, with lighter orange reflecting association with positive symptoms (note they are square root transformed) and dark orange for negative symptoms (not transformed). For all other plots, relationships are represented in orange for patients, and compared with controls in green. A solid line is drawn for significant associations that remain significant after correction for multiple comparisons, and a dashed line represents nominally significant associations for P < 0.05, uncorrected. The general cognitive index is significantly and negatively associated with FEP patients’ component scores on principal component 2, whereas positive and negative symptoms are nominally and positively associated with this component. Finally, verbal memory is significantly and negative associated with FEP patient’s component scores on principal component 3. This latter relationship also holds when covarying for antipsychotic medication (r = −0.26, P = 0.0056). Note, Plot A reflects data from full sample of 116 FEP patients; for Plot B (general cognitive index), 2 patients and 6 controls had missing cognitive data from 1 or more of the 6 domains used to calculate GCI. For verbal memory (Plot C), 1 patient and 5 controls were missing data.

Given that component 2 captured a “primary sensory and motor module” that seems to capture group differences in WGC, both as mentioned above and in other work (Jorgensen et al. 2016), and following the line of thought that brain dysfunction of primary sensory regions may subserve general psychopathology in psychosis (Javitt 2009a, 2009b), we tested whether scores for this component were related to positive/negative symptoms and general cognitive deficits in patients. With a statistical threshold of P = 0.017 (Bonferroni correction for 3 tests: 0.05/3), the general cognitive index was significantly associated with scores on component 2 in FEP (r = −0.26, P = 0.0048). Nominal associations were found with positive (r = 0.20, P = 0.031) and negative symptoms (r = 0.18, P = 0.031), although not significant after correcting for multiple comparisons.

Component 3 captured heteromodal regions strongly associated with higher-order cognitive processing; thus, we tested whether 3 higher-order cognitive domains (verbal memory, executive function, and working memory) were associated with component 3 scores in patients. Verbal memory was found to be significantly and negatively related to component 3 scores (r = −0.25, P = 0.0065). No significant relationships were found with the other 2 domains (executive function: r = −0.078, P = 0.41; working memory: r = −0.066, P = 0.48). Results were largely similar when including only first-episode schizophrenia patients (Supplementary Fig. 6), although the group difference in component 2 scores was weaker (F₁₁₆₇ = 1.89, P = 0.17), albeit in the same general direction (i.e., patients had higher component scores compared with controls).

No significant associations were found between duration of untreated psychosis and any of the 3 principal component scores derived from WGC data within the patient group. However, a nominally significant association was found between patient component scores on principal component 3 and antipsychotic medication exposure (r = 0.19, P = 0.042). Given this association, we re-examined the association between principal component 3 and verbal memory performance in patients, covarying for antipsychotic medication, and found that this relationship still held (r = −0.26, P = 0.0056).

PCA Results with CT

Similar to results with WGC, only the first 3 components derived from the CT data were interpreted (see scree plot in Supplementary Fig. 3C), collectively explaining 50.75% of the variance in CT. Cortical surface maps of the PCA loadings for the 3 selected components are depicted in Figure 2. The first component (explaining 41.4% of the variance) reflected mean CT, and was significantly and negatively associated with positive symptom severity (r = −0.22, P = 0.016), as well as significantly positively associated with GCI (r = 0.22, P = 0.017) (Supplementary Fig. 4). The second component, explaining 5.9% of the variance, showed positive loadings on primary motor cortices bilaterally, sharing similarities to the second component derived with WGC, as well as weak negative loadings in fronto-insular-temporal regions. The third component, explaining 3.5% of the variance, captured positive loadings in right prefrontal cortex and negative loadings in occipital cortex bilaterally. Again, no significant group differences (FEP vs. control) were found in component scores for these 3 components, although component 3 showed a trend-like difference (F₁₂₀₂ = 3.61, P = 0.059), where patients had lower component scores than controls; in other words, patients tend to have marginally lower CT in right prefrontal cortex and higher thickness in occipital cortex bilaterally compared with controls.

Figure 2.

Top 3 components from PCA with CT data and related variables. Cortical surface maps reflect component loadings for each principal component (PC). PC1 reflects mean CT signal (see Supplementary Fig. 3). PC2 captures positive loadings on dorsal primary motor regions, and weak negative loadings in frontal-insular-temporal regions. For direct comparison to results presented for WGC in Figure 1, we show here that positive/negative symptoms were not associated with component 2 derived from CT data (Plot A), but a nominal association was found between component loadings and a general cognitive index (GCI; Plot B) in FEP patients. PC3 captures positive loadings on right prefrontal cortex and negative loadings on occipital cortices bilaterally. For direct comparison with results presented for WGC in Figure 1, a plot depicting the non-significant relationship between verbal memory and component 3 scores is presented (Plot C). Pearson r-correlations are presented for FEP patients and controls for significant findings, where applicable. Association of principal component 2 with both positive and negative symptoms are presented together in Plot A, with lighter orange reflecting association with positive symptoms (note they are square root transformed) and dark orange for negative symptoms (not transformed). For all other plots, relationships are represented in orange for patients, and compared with controls in green. A dashed line represents nominally significant associations for P < 0.05, uncorrected. There was only a significant relationship between FEP patients’ GCI and component scores on principal component 2. Note, Plot A reflects data from full sample of 116 FEP patients; for Plot B (general cognitive index), 4 patients and 6 controls had missing cognitive data from one or more of the 6 domains used to calculate GCI. For verbal memory (Plot C), 1 patient and 5 controls were missing data.

To maintain consistency in the analysis of WGC data, we assessed brain–behavior relationships with general psychopathology and cognitive deficits as described above for WGC for components 2 and 3. For component 2, a significant positive association was found between component scores and the general cognitive index in FEP patients (r = 0.24, P = 0.0092). No notable associations were found with component 2 scores and positive/negative symptoms (r = −0.13, P = 0.15; r = −0.06, P = 0.54, respectively), or component 3 scores and “higher-order” cognitive abilities (verbal memory: r = 0.015, P = 0.87; executive function: r = −0.14, P = 0.13; working memory: r = −0.15, P = 0.12). For direct comparison, similar plots as those presented in Figure 1 for WGC data are included in Figure 2 for CT data. Results were largely similar when retaining only first-episode schizophrenia patients (Supplementary Fig. 7) in the PCA with CT data, and similarly, a nominal group difference was found for component 3 scores (F₁₁₆₇ = 3.95, P = 0.049), where patients had lower scores compared with controls. No significant associations were found between duration of untreated psychosis or antipsychotic medication exposure and any of the component scores derived with CT data within the patient sample.

Finally, PCA was applied separately for FEP patients and the healthy control group. Patterns of brain regions uncovered for the top 3 principal components were largely similar to the combined sample. However, for component 3, similar anatomical patterns were found but with reverse loadings in some regions when comparing groups. For instance, for component 3 of WGC data, patients have negative loadings on prefrontal cortex and positive loadings around the temporoparietal junction, whereas controls have positive loadings on prefrontal cortex and negative loadings on temporoparietal regions. A similar pattern emerged with CT data for principal component 3; for instance, negative loadings on this component for patients in occipital cortex contrasted with positive loadings on occipital cortex in HC. See Supplementary Figures 10 and 11 for PCA derived from WGC and CT data, respectively.

Structural Covariance

Structural covariance analyses for both WGC and CT revealed pairwise correlations that were both increased and decreased in FEP patients compared with HC, both within and across hemispheres. See Supplementary Table 1 for a list of AAL and DKT regions. For structural covariance analyses with WGC, 28 of 3003 unique pairwise correlations using the AAL atlas were found to be significantly different between FEP and HC, with the majority of differences in the direction FEP < HC (Fig. 3). Differences were highly widespread across brain regions, with the highest positive z-score (i.e., FEP > HC; z = 4.88) between right anterior cingulate and left medial superior frontal gyrus, and the strongest negative z-score (i.e., FEP < HC; z = −5.16) between fusiform and medial superior frontal gyri of the right hemisphere. Additionally, a significant pairwise difference between proximal regions within the left occipital cortex emerged, complementing the vertex-wise results presented earlier. For CT analyses, out of 3003 tests, 3 were found to be significant, with one significant positive z-score (i.e., FEP > HC; z = 4.57) found between left superior orbitofrontal gyrus and right superior parietal lobule, and the strongest negative score (i.e., FEP < HC; z = −4.56) between right fusiform and left superior medial orbitofrontal gyrus (Fig. 3).

Figure 3.

Group differences in structural covariance networks across 78 AAL regions. The first row shows results with WGC, whereas the second row shows results with CT. Panel A: Matrix shows the differences in pairwise correlation coefficients between FEP and HC that survived correction for multiple comparisons with FDR. See color bar for z-scores. Panel B: A “glass brain” schematic was generated to better visualize the brain topology of significant inter-regional group differences uncovered in the corresponding matrices. Open circles represent center coordinate of each AAL region, and lines connecting each region represent significant pairwise regional differences, as depicted in the left-hand matrix. For both matrices and glass brains, red represents stronger, more positive correlation coefficients in FEP compared with HC, whereas blue represents stronger, more positive correlation coefficients in HC compared with FEP. For WGC, 28 of 3003 unique pairwise correlations using the AAL atlas were found to be significantly different between FEP and HC, with the majority of differences in the direction FEP < HC. As can be seen in the figure with CT, substantially fewer comparisons survived correction for multiple comparison when comparing groups, i.e., 3 of 3003 unique pairwise correlations were found to be significant.

Supplementary Structural Covariance Results

Results were comparable when using the DKT atlas, with a few key differences highlighted in the Supplement. For WGC, 25 of 1891 tests were found to be significant, with the strongest positive z-score (z = 5.15) found between right calcarine gyrus and left lateral orbitofrontal gyrus, and the strongest negative z-score (z = −5.44) between caudal anterior cingulate and inferior temporal gyrus of the left hemisphere. For CT, 8 of 1891 tests were found to be significant, with the strongest positive z-score (z = 5.24) found between right entorhinal and superior frontal regions, and the strongest negative z-score (z = −4.46) between right superior temporal gyrus and left precuneus (Supplementary Fig. 5). Intriguingly, more group differences in structural covariance across both WGC and CT analyses were uncovered when including only first-episode schizophrenia patients (Supplementary Figs 8 and 9) for both atlases. Notably, a similar pattern of results was found when retaining this subsample of patients, such that more pairwise group differences were uncovered with WGC as compared with CT. Finally, a summary of all pairwise differences found with all structural covariance analyses, and the proportion of inter- to intra-hemispheric pairwise differences uncovered can be found in Supplementary Table 3. In summary, there was a higher proportion of inter-hemispheric pairwise differences.

Discussion

Our findings provide additional and novel insight into the neuroanatomical differences characterizing FEP patients. Utilizing WGC at the inner edge of the cortex, as well as CT, we highlight differences in brain structural patterns in FEP patients compared with controls, which may be driven by altered network connectivity. Data-driven analysis of anatomical patterns with WGC highlighted 2 separate components of (1) primary sensory (somatosensory and visual) and motor regions, and (2) heteromodal cortices that were related to measures of general psychopathology and verbal memory deficits in FEP patients, respectively. Analysis of group differences in structural covariance suggested stronger and more widespread abnormalities in FEP captured predominantly by WGC. Putative network abnormalities revealed with PCA were complemented by those uncovered with structural covariance analyses; for instance, both approaches pinpointed primary somatosensory-motor regions, where patterns of stronger loadings across these areas derived with PCA were also found to have altered structural covariance between groups, particularly for WGC. Notably, primary visual cortex was clustered with the somatosensory-motor regions in the WGC analysis, but not in the CT data, suggesting that architectonic (e.g., cyto-, myeloarchitecture) features may be better captured by WGC. Indeed, a microstructural gradient spanning from sensorimotor to transmodal regions has previously been described (Huntenburg et al. 2018), where primary sensory and motor regions captured by the first PCA component would be expected to share microstructural features under such a framework. These patterns derived with our WGC data are also supported by recent analyses with neurite density maps obtained from diffusion imaging and comparisons to T1w/T2w ratio distributions, previously interpreted as putative “myelin” maps (Glasser and Van Essen 2011), and to granular cortex described by neuroanatomists (Fukutomi et al. 2018). However, it should be noted that group differences in principal component scores did not reach significance, which contrasts with the number of significant group differences uncovered between pairwise cortical regions for both WGC and CT with structural covariance. It is possible that the cortical patterns captured by PCA, specifically for WGC, are capturing individual-level differences in clinical symptoms and cognition, rather than summarizing group differences, which structural covariance may tap into. Finally, the derived component scores from PCA were largely independent from potentially confounding clinical factors such as duration of untreated psychosis and antipsychotic medication exposure, with the exception of principal component 3 derived from WGC data, where scores were found to be significantly associated with antipsychotic medication exposure; this is not an entirely unexpected result, as the prefrontal cortex was a prominent feature of this component; a region that has been linked to alterations in intracortical myelin as a function of antipsychotic medication exposure previously (Bartzokis et al. 2009).

It is important to recall that differences in contrast can be caused by 2 very different plausible mechanisms: one relating to the gray matter intensity or intracortical myelin and the other related to superficial white matter. Taking into consideration the former, it is well established that primary sensory and motor cortices are typically characterized by high levels of intracortical myelin content in HC (Glasser and Van Essen 2011; Zilles et al. 2015; Marques et al. 2017). Thus, the observed increased contrast in FEP patients and reduced covariance with both proximal and distal brain regions suggests that regions within primary sensory and motor cortices may have fewer myelinated axons terminating within deep layers of the cortex. These changes in contrast could also be interpreted in terms of the white matter composition of cortico-cortical or cortico-subcortical connections. Notably, thalamocortical projections are most prominent within deep layers of primary sensory cortices compared with other higher order brain regions (Jones 2002; Lewis et al. 2018). In psychosis, a sizeable number of studies have suggested that thalamocortical connectivity plays an important role in the neuropathology of psychosis (Murray et al. 2016; Giraldo-Chica and Woodward 2017), where hyperconnectivity between the thalamus and primary sensory and motor regions may be predictive of conversion to psychosis and severity of symptoms (Anticevic et al. 2014, 2015). It may be fruitful for future investigations to specifically test whether thalamocortical connectivity could mediate the reported changes in contrast and covariance within visual and primary somatosensory-motor cortices shortly after a FEP.

Although no group differences were statistically significant when comparing component scores derived from PCA for both WGC and CT, the data pinpoint patterns of cortical associations that distinguish FEP patients from controls, namely increased contrast in primary sensory cortical regions, increased thickness in occipital cortex, and decreased thickness in right prefrontal cortex. Findings of prefrontal cortical thinning in early psychosis have previously been reported (Voets et al. 2008; Janssen et al. 2009; Hollis and Palaniyappan 2015; Bartholomeusz et al. 2017). Intriguingly, a recent report in schizophrenia has also reported a concomitant decrease in prefrontal CT alongside increases in thickness within occipital cortex (Guo et al. 2016). Further, increased contrast within primary sensory-motor cortices has been shown in later phases of the disorder (Jorgensen et al. 2016), and are likely to already be present at the onset of psychosis, as shown by our findings. Extending these findings to earlier phases of psychosis, a recent study also showed increased contrast in visual cortex, alongside decreased contrast in primary somatosensory and motor cortices, in relation to prodromal psychosis symptoms (Norbom et al. 2018). Altogether, these findings give rise to the possibility that aberrancies in cortical myelination may arise in primary visual cortices, and may be followed by further progression of abnormal pericortical white matter development in somatosensory-motor regions in patients that transition to psychosis. A longitudinal study investigating WGC trajectories from the prodrome to psychosis onset would be invaluable in testing this hypothesis.

Examination of PCA components derived with WGC data further corroborates the idea that WGC may indeed be tapping into a measure of myelin. For instance, the primary sensory and motor regions uncovered with component 2 versus the heteromodal cortical regions contributing to component 3 are characterized by different levels of cortical myelin (Glasser and Van Essen 2011; Rowley et al. 2015). Additionally, these components were meaningfully related to core features and deficits of psychosis, beyond what was found with CT. As alluded to, primary sensory (i.e., somatosensory and visual) and motor regions with higher contrast in patients (potentially due to lower myelin compared with controls) were related to 3 general domains of psychopathology (i.e., positive/negative symptoms and cognitive deficits). This supports a well-cited theory of deficits in bottom-up processing stemming from these primary cortical regions, which serve as a catalyst for further disruptions in higher-order cognitive processes in psychosis (Javitt 2009a, 2009b). In line with this, one of the most disrupted cognitive processes in psychosis, verbal memory, was found to be significantly related to differences in contrast within left superior temporal sulcus and heteromodal association cortices uncovered in component 3. Others have endorsed this notion that degree of myelination within higher-order cortical regions is associated with cognitive performance in healthy individuals (Blackmon et al. 2011; Grydeland et al. 2013, 2015) and in patients with enduring schizophrenia (Bartzokis et al. 2011). The mapping of verbal memory ability specifically (and not visual or working memory) to this component is also intriguing given the strong positive loadings of left superior temporal sulcus/auditory cortex in component 3, which is highly related to language and speech (Devlin et al. 2003). Many studies have also pinpointed this region as a significant neural correlate of communication impairments and auditory verbal hallucinations in schizophrenia (McGuire et al. 1996; Pearlson 1997; Allen et al. 2008, 2012).

There was also some complementarity in components derived with CT data (dorsal motor cortex in component 2, right prefrontal cortex in component 3); these mutual findings are not unexpected, as the placement of the gray–white matter boundary for CT measurements largely depends on tissue contrast. Specifically, relationships were uncovered between general cognitive abilities and the second PCA component, which highlighted positive loadings in dorsal primary motor cortex for both WGC and CT. Our data suggest that greater magnitude of general cognitive deficits may be linked not only to higher contrast as discussed above, but also to lower CT in somatosensory-motor regions. It is feasible that the mechanism underlying this relationship is related to the integrity of white matter connections invading the cortex. However, the differences between components extracted from WGC versus CT data are also worth discussing; for instance, positive loadings of primary visual cortices were included for component 2 with WGC data, but not for CT. Given that the myeloarchitecture underlying somatosensory-motor regions shares features with that of the primary visual cortex (Hopf 1955; Glasser and Van Essen 2011; Huntenburg et al. 2017), this provides further evidence that our WGC metric may indeed be tapping into features of myeloarchitecture and white matter connectivity, more so than CT.

The resemblance of the brain patterns uncovered in components 2 and 3 with WGC data to levels of cortical myelin also suggests that WGC is positioned well to capture relationships between brain regions with microstructural similarities, which has been posited to be one of the basic wiring principles of the mammalian brain (Barbas 2015; Pandya et al. 2015; Huntenburg et al. 2017; Lariviere et al. 2018). This was supported by our structural covariance results, where WGC painted a striking picture of widespread pairwise brain regions with both increased and decreased covariance in FEP compared with controls over and above what CT yielded, predominantly between inter-hemispheric regions. These results lend support to the long-standing theory and accompanying evidence that psychotic disorders may reflect an aberrant state of connectivity (Friston and Frith 1995; Davis et al. 2003; Friston et al. 2016). The greater proportion of altered inter-hemispheric relationships found in our work is also supported by previous studies showing long-range dysconnectivity, and abnormalities in commissural fibers in psychotic disorders (Keshavan et al. 2002; Whitford et al. 2010; Guo et al. 2014).

Although structural covariance cannot be equated with connectivity, the visibly larger degree of differences in structural covariance networks with WGC, as opposed to CT, suggests that WGC may be tapping into a compartment of brain structure (e.g., pericortical myelin) that is strongly and dynamically altered in early psychosis. Several recent studies using magnetic transfer ratio (MTR) imaging provide insight into our described findings. One of these studies defined covariance networks of intracortical myelin with MTR in HC (Melie-Garcia et al. 2018). That study empirically demonstrated that myelination processes within gray matter are synchronized across various regions of the brain, which is also reflected in the widespread differences in pairwise relationships that we uncovered in our FEP sample using WGC. Another recent investigation applied MTR connectomics to a sample of schizophrenia patients and also found diffuse microstructural abnormalities in patients, particularly within the prefrontal cortex (Wei et al. 2017). Together, this evidence corroborates our conjecture presented earlier that metrics related to white matter compartments of the cortex may have closer correspondence to measures of “connectivity” as defined by other brain imaging modalities, and in turn, white matter-related indices of covariance are likely to be more fruitful in uncovering differences pointing towards “dysconnectivity” in FEP patients among cortico-cortical connections. It is acknowledged that this interpretation is still speculative, as further investigation of the precise mechanisms and neurobiology underlying microstructural measures such as WGC and MTR is needed. Our group is currently investigating such questions with the use of quantitative T1 mapping, and future work will further explore correspondence of such maps with multi-shell diffusion mapping techniques.

Our findings should be interpreted in the context of a few limitations. As alluded to, our measure of WGC does not allow us to disentangle the primary source of change in white matter alterations (i.e., within intracortical layers or superficial white matter). Standard T1-weighted acquisitions suffer from strong scanner field biases, which makes it very difficult to interpret gray and white matter intensities separately. Future work should utilize quantitative MRI methods to disentangle the true source of microstructural changes. Bias field inhomogeneities also pose a concern in the extraction of cortical metrics. Our calculation of local contrast at the inner edge of the cortex is largely immune to such inhomogeneities, as our measurement of contrast uses measures from points that are proximal to each other, and only relies on the placement of the gray–white matter boundary, which is based on the intensity gradient between cortical gray and white matter. The same cannot be stated for CT, a measure which depends on a non-local tissue segmentation process and relies on the placement of 2 boundaries spanning a larger area, and thus is much more susceptible to intensity inhomogeneities present in the T1-weighted image (Sled et al. 1998; Bezgin et al. 2018). We also acknowledge that data from a 1.5 T scanner limits our ability to measure WGC; for instance, it has been shown that WGC measured on 3 T is able to predict age with higher accuracy (Lewis et al. 2018). Although we carried out a stringent quality control procedure, it is acknowledged that even subtle motion can also have an impact on cortical measures, particularly in psychiatric patient samples (Reuter et al. 2015; Pardoe et al. 2016; Yao et al. 2017; Makowski et al. 2019). Thus, it may not be surprising that more scans belonging to FEP patients (14) were excluded compared with HC (2). We also collapsed our analyses across diagnostic categories, including patients with both affective and non-affective psychoses in our FEP patient sample. We re-ran results with first-episode schizophrenia patients only, and largely found the same results. Notably, patients with first-episode schizophrenia showed an even greater proportion of pairwise differences, particularly with WGC, compared with controls, which may reflect the idea that patients with schizophrenia tend to exhibit more severe abnormalities compared with patients with affective psychosis, such as those with bipolar disorder (Murray et al. 2004). Finally, we generated structural covariance results using 2 commonly used atlases and found some differences in the derived results. An atlas defined by cytoarchitectonic boundaries would be beneficial in future studies using WGC.

Conclusion

This study extends the notion that psychosis is a disorder of dysconnectivity, by showing that systems-level aberrations are already present in early psychosis. Further, our results with WGC suggest that pericortical myelin may be a sensitive marker of abnormalities in cortico-cortical connections, and such network-level abnormalities are meaningfully related to symptom severity and cognitive ability. Our findings highlight the need for future studies to elucidate the mechanisms underlying aberrant patterns of architectonics, particularly within myeloarchitecture, that contribute to the dysconnection syndrome of psychosis.

Funding

The study was supported by operating grants from the Canadian Institutes of Health Research (CIHR, 68961, MCT-94189); the Fonds de Recherche du Quebec—Santé (FRSQ); Sackler Foundation; and from the Azrieli Neurodevelopmental Research Program (grant ANRP-MIRI13-3388) in partnership with the Brain Canada Multi-Investigator Research Initiative (to A.C.E.). It also benefited from computational resources provided by Compute Canada (www.computecanada.ca) and Calcul Quebec (www.calculquebec.ca). Salary awards include Canadian Institutes for Health Research (CIHR; C.M., R.J., M.L., A.C.E.), Fonds de la Recherche en Santé du Québec (FRSQ; C.M., M.L., and R.J.), funding from the Canada First Research Excellence Fund, awarded to McGill University for the Healthy Brains for Healthy Lives Initiative (C.M. and A.C.E.), James McGill Professorship (M.L. and A.C.E.) and Canada Research Chairs Programme (A.K.M.).

Disclosures

A.C.E. has received consulting fees from Biospective. M.L. reports grants from Otsuka Lundbeck Alliance, personal fees from Otsuka Canada, personal fees from Lundbeck Canada, grants and personal fees from Janssen, and personal fees from MedAvante-Prophase, outside the submitted work. R.J. reports receipt of grants, speaker’s fees, or honoraria from AstraZeneca, BMS, Janssen, Lundbeck, Otsuka, Pfizer Canada, Shire, and Sunovion and royalties from Henry Stewart Talks. A.M. reports receipt of grants, fees, or honoraria from BMS, Lundbeck, and Otsuka. The authors have no other competing interests to disclose. The funding organizations had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.