Covarying Structural Alterations in Laterality of the Temporal Lobe in Schizophrenia: A Case for Source-Based-Laterality

Abstract

The human brain is asymmetrically lateralized for certain functions (such as language processing), to regions in one hemisphere relative to the other. Asymmetries are measured with a laterality index (LI). However, traditional LI measures are limited by a lack of consensus on metrics used for its calculation. To address this limitation, source-based-laterality (SBL) leverages an independent component analysis (ICA) for the identification of laterality specific alterations, identifying covarying components between hemispheres across subjects. SBL is successfully implemented with simulated data with inherent differences in laterality. SBL is then compared to a voxel-wise analysis utilizing structural data from a sample of patients with schizophrenia and controls without schizophrenia. SBL group comparisons identified 3 distinct temporal regions and one cerebellar region with significantly altered laterality in patients with schizophrenia relative to controls. Previous work highlights reductions in laterality (i.e. reduced left gray matter volume) in patients with schizophrenia compared to controls without schizophrenia. Results from this pilot SBL project are the first, to our knowledge, to identify covarying laterality differences within discrete temporal brain regions. The authors argue SBL provides a unique focus to detect covarying laterality differences in patients with schizophrenia, facilitating the discovery of laterality aspects undetected in previous work.

Introduction:

Alterations in gray matter in patients with schizophrenia (SZ) have been frequently reported in studies analyzing functional and structural MRI data and meta-data ^1–10. Many of these alterations are weighted towards the left hemisphere, leading researchers to suggest schizophrenia may produce alterations in brain laterality in SZ compared to controls with no diagnosis of schizophrenia (CON)^10–23. However, many of the findings upon which this hypothesis is based center on the outcome of regional or whole-brain voxel/vertex-based methods, which may not be ideal for assessing laterality-specific conclusions. Traditionally, such changes can be assessed by calculating a single laterality index (LI) for each subject, using the equation 1:

$$LI=f\frac{Q_{LH}-Q_{RH}}{Q_{LH}+Q_{RH}}$$ eq (1)

Where Q is the quantity of the MRI metric of interest (e.g. blood oxygenation level dependence [BOLD] percent signal change, gray matter volume, cerebral blood-flow volume), LH and RH represent Q for the left and right hemisphere respectively, and f represents a scaling factor, usually 1 or 100^24–27, with positive values indicating left-hemispheric dominance and negative values indicating right-hemispheric dominance. Many toolboxes are available to calculate the LI of neuroimaging data^28,29 utilizing this formula. Unfortunately, the laterality calculation itself has limitations.

One of the most notable limitations of the LI calculation is a lack of consensus on what metric should be used for Q, with suggestions including: voxel intensity^24,26, weighted sums³⁰, mean signal change^31,32, average correlation coefficients³³, F-values³⁴, t-values^35,36, and weighted t-values^37–39 alone or those which survive certain thresholds²⁷. The use of t-values is problematic because negative t-values can lead to significant misinterpretations of the output (which can be mitigated by using absolute values)^27,40. Furthermore, whole-brain vs region of interest (ROI) approaches may yield very different results based on ROI selection and thresholding²⁷. Thresholds within specific ROIs may not be appropriate for some populations in which these functions are altered or have shifted (e.g. stroke, tumor), while whole-brain analyses may lack the ability to identify specific brain regions or patterns of variation across multiple brain regions. Additionally, researchers often choose to omit the cerebellum from laterality analyses due to contralateral connections of the cerebellum and cortex biasing LI measures^{27,39,41–43}. Ironically, the advantages of ROI and/or cluster/vertex wise analyses are also their greatest flaws. For example, while ROI-based approaches (such as those implemented in Freesurfer^44–51 and similar packages), produce stable results with larger effect sizes, in populations where differences may be focal or variable (i.e. hand regions of the primary motor cortex in piano players vs non-piano players), differences may vanish when values are averaged over unnecessarily large ROIs^52,53. This risk considerably increased when large smoothing kernels are applied to the data (see Scarpazza et al., [2015] for a brief review)⁵⁴. In addition, such analyses risk considerable reductions in power relative to the number of ROIs, as multiple comparisons or models would need to be performed for each ROI or ROI pair. While Freesurfer^44,46–51 is capable of vertex-wise corrections within masks (see http://freesurfer.net/fswiki/BuildYourOwnMonteCarlo for documentation and Patriquin et al. [2016]⁵⁵ for an example of implementation) as opposed to ROIs from atlases, this approach rarely seems to be used. However, even when voxel/vertex-wise analyses are implemented, the choice of correction and sample size could lead to vastly different results between studies due to the power requirements for voxel/cluster wise corrections^27,56–59. As such, ROI and voxel agnostic methods would significantly improve laterality inferences in MRI data.

Our goal in this work was to develop an improved measure of laterality that exhibits 1) agnosticism regarding to the size/shape/cluster extent of a region, 2) laterality-specific pattern identification rather than region-specific identification, 3) robustness to patterns of alteration in the absence of a priori regions. To that end, the authors propose a novel solution to analyzing laterality by directly estimating covarying networks from homotopic mirror images via ICA, called source-based laterality (SBL). SBL is an extension of source-based morphometry (SBM), which utilizes a multivariate, ICA-based approach to identify interrelationships across voxels in anatomical analyses such as gray matter maps in voxel-based-morphometry⁶⁰. In lieu of utilizing a gray matter map, SBL utilizes a homotopic subtraction of right and left voxels from gray matter hemispheres, implemented by flipping the right hemisphere and subtracting it from the left, producing a single, voxel-wise difference image between the two hemispheres. The resulting hemispheric difference image for each participant is then analyzed across participants using spatial ICA. In this manner, ICA solutions are specific to laterality-based components which covary between hemispheres across subjects. The authors argue the SBL approach provides a unique approach to detecting covarying laterality differences between SZ and CON participants, enabling researchers to capture subtle alterations in cortical symmetry that are not identified by traditional voxel or ROI based approaches.

As a proof of concept, a simulation of gray matter difference maps is used to demonstrate the robustness in the identification of spatial covarying patterns using SBL. SBL is then applied to difference images of gray matter volume maps from patients with SZ and CON from the function biomedical informatics research network (FBIRN)^61,62. The authors hypothesize gray matter difference images leveraged by SBL will identify alterations in structural gray matter laterality in SZ compared to CON participants (demonstrated in previous literature^{2–6,8–10,13–23,63–72}), and structural network covariance (SNC) while retaining spatial specificity and statistical power unavailable to voxel or ROI based approaches.

Methods:

Simulation:

Data Generation:

Step 1 included the generation of a 400×400 2D images with 1mm dimensions containing 3 asymmetric ROIs (ROI 1 area = 2188mm², ROI 2 area = 757mm², ROI 3 area = 347mm²) to serve as sources. Step 2 included source generation, with Source 1 (ROI 1) consisting of an inverse hyperbolic tangent transformation of a normally distributed vector of 2188 values centered on a mean of 0 with a standard deviation of 0.27, with 0.5 added as a constant. Source 2 (including ROIs 2 and 3) also consisted of an inverse hyperbolic tangent transformed vector with an added constant of 0.5, but of 1104 values with a mean of 0 and a standard deviation of 0.4, all of which were multiplied by a value of 0.1. Gaussian noise was generated across the remaining values in the image mask using a normal distribution centered at a mean of 0 and a standard deviation 0.15. Example images produced with this process are displayed in Figure 1. Step 3 generated 300 2D images with the specified source parameters, with 150 representing “controls” and 150 representing “patients.” For each simulated control, weights for source 1 were chosen randomly from values between 0.5–0.8, and 0.2–0.4 for source 2. For each simulated patient, weights for source 1 were randomly chosen between 0.2–0.4, and 0.8–1 for source 2. Finally, ground truth weights were recorded for each source to assess the accuracy of the SBL component weightings.

Figure 1:

A) Top Figure: Simulated source 1 (single sphere, red), and source 2 (double sphere, blue). Bottom Two Figures: For controls (middle) source 1 is primarily positive while source 2 is primarily negative. For patients (bottom) intensities are negative for source 1 and positive for source 2. Intensities for raw laterality images are bound between 0.5 and −0.5 for visualization. B) Average difference maps by group are displayed for the left hemisphere (top most figure), source 1 (middle figure), and source 2 (bottom figure). Values above 0 (black line) indicate greater volume in the left hemisphere, while values below 0 indicate greater GM volume in the right hemisphere. C) Breakdown of independent component analysis utilizing the formula X[data] = A[Mixing Matrix] * S[Source Matrix]. Simulated difference images (X) are provided and decomposed into component weights (A) and a source matrix (S).

Source-Based Laterality:

SBL was performed on the 300 simulated subtraction maps implemented within the group ICA of fMRI toolbox (GIFT; http://mialab.gsu.edu/software/gift)⁷³ utilizing group-level PCA, spatial ICA with infomax, and calculation of individual spatial maps using PCA-based back-reconstruction to identify 3 independent components, resulting in loading parameters for each component in the simulated group data⁶⁰.

Real data:

Demographic Information:

T1 weighted images from 326 participants, 167 participants with SZ (M/F: 129/38) and 159 (M/F: 113/46) CON from 7 imaging centers: The Mind Research Network (MRN), Duke University, University of California Los Angeles (UCLA), University of California Irvine (UCIrvine), Neuroimaging Center of University of California Irvine (NICIrvine), University of Iowa (UIowa), and the University of Minnesota Medical School’s Center for Magnetic Resonance Research (UMCMRR). Participants provided informed consent in accordance with institutional review board requirements of each participating university. The Edinburgh Handedness Inventory (EHI)⁷⁴ was collected to assess the dominant hand from each participant as an additional variable of consideration when assessing laterality. Briefly, the EHI scores responses regarding hand use for 10 different tasks, and assigns ambidexterity, left, or right hand dominance based of the decile from the totaled hand use during each task⁷⁴. The scale ranges from −100 (completely left dominant), +100 (completely right dominant), with a median of −28 ≤ LI < 48 indicating ambidexterity⁷⁴. An interactive example of the questionnaire can be found at the Organization for Human Brain mapping at: http://www.brainmapping.org/shared/Edinburgh.php. Detailed descriptions of data processing, storage protocols, and participant anonymization can be accessed at the main FBIRN portal on neuroimaging tools and research collaboratory (NITRC) at https://www.nitrc.org/projects/fbirn/, or queried directly through http://schizconnect.org following registration. BIRN protocols designate identifying information (i.e. gender, age, subject ID, ect) cannot be directly shared without registering with the BIRN initiative.

Patients with SZ did not display significant differences between controls on age (Age Range = 18–68 yrs, Mean Age = 38.19±11.39, W = 14169, p = 0.29, gender composition (χ² = 1.3176, p = 0.25), or handedness (χ² = 1.9454, p = 0.39 [Monte-Carlo w/2k]). These statistics are summarized in Table 1. Site-wise differences in participant sex approached significance (χ² = 11.9, df = 6, p = 0.06), but no significant differences in the representation of diagnostic groups (χ² = 1.07, df = 6, p = 0.98) were present.

Table 1.

Participant demographic information and comparisons

	SZ Patients	Typical Controls	Test Statistic	p-value
Age Range	18–62	19–60	W = 14169	0.29
M/F	129/38	113/46	χ2= 1.32	0.25
Handedness R/L/A	152/11/4	151/6/2	χ2= 1.95	0.39
Median PNASS Negative	14		Range: 7–39
Median PNASS Positive	14		Range: 7–33

Note:

^*Indicates statistical significance at or below the 0.05 level.

W represents the rank-difference statistic from a Wilcoxon rank-sum test.

χ² Represents the goodness-of-fit from a Pearson chi-squared test.

Image Parameters and Quality Analyses:

High-resolution T1 Siemens MP-RAGE were collected from 284 patients on six 3T Siemens Tim® Trio Systems (MRN, UCLA, UCIrvine, NICIrvine, UMCMRR, and UIowa). MP-RAGE parameters were TR/TE/TI=2300/2.94/1100 ms, flip angle=9°, resolution=256×256×160. The remaining 42 T1 images were acquired using a 3T General Electric (GE) Discovery MR750 scanner using a GE IR-SPGR sequence (TR/TE/TI=5.95/1.99/450 ms, flip angle=12°, resolution=256×256×166). Images from all sites covered the entire brain with field of view (FOV)=220 mm², with 324 of the scans utilizing voxel dimensions=0.86×0.86×1.2 mm³, 1 scan using dimensions=0.9×0.9×1.2 mm³, and one using 1mm³ isometric voxels. All scans utilized the sagittal scan plane, GRAPPA/ASSET acceleration factor=2, and number of excitations=1.

CON vs SZ:

Image quality metrics on each participant were calculated using virtual python 3.6 environment of MRIQC 0.11.0⁷⁵. Group comparisons revealed significant differences between patients with SZ and CONs in the coefficient of joint variation (W = 15940, p = 0.002, r = 0.17), higher contrast-to-noise ratio t_(322.62) = −3.7636, p = 0.0002, d = −0.42, higher average data smoothness (FWHM) (W = 11471, p = 0.03, r = 0.11), and greater intensity non-uniformity parameters (INU) t_(318.46) = −3.9682, p = < 0.0001, d = −0.44, CI95%−0.07 −0.02, in controls compared to SZ patients, but higher entropy focus criterion t_(322.9) = 4.39, p = < 0.0001, d = 0.48, CI95%: 0.01 0.03, gray matter residual partial volume error, t₍₃₂₄₎ = 2.69, p = 0.007, d = 0.30, CI95%: 0.15 0.61, white-matter maximum-intensity ratio (W = 11474, p = 0.03, r = 0.12), and initial ICBM asymmetric tissue probability map overlap (W = 7745, p = < 0.001, r = 0.36) were found in SZ patients compared to controls However, no significant differences between patients and controls in gray matter signal-to-noise ratio (W = 12742, p = 0.5) were found. These results are summarized in Supplementary Figure 1.

Comparisons by Data Acquisition Site:

Site-wise analyses of image quality metrics identified differences on coefficient of joint variation, Kruskal-Wallis χ² = 91.175, df = 6, p < 0.001, gray matter signal-to-noise ratio F(6, 319) = 39.81, p < 0.001, ώ² = 0.42, entropy focus criterion, F(6, 319) = 39.81, p < 0.001, ώ² = 0.32, INU correction parameters Kruskal-Wallis χ² = 111.46, df = 6, p < 0.001, and average smoothness Kruskal-Wallis χ² = 129.37, df = 6, p < 0.0001. However, no significant differences were found for contrast-to-noise ratio, gray matter residual partial volume error, or overlap with the initial ICBM asymmetric tissue probability map. These comparisons are also summarized in Supplementary Figure 2 a–e.

Image Processing and Analysis:

Gray matter volume maps were computed from all T1 images using the SPM12 voxel-based morphometry “new segmentation” pipeline (build 6906)^76–79. The pipeline includes segmentation based on priors for 6 tissues classes, normalization to the MNI 152 template, and modulation of gray matter maps the Jacobian determinant to preserve tissue volume. Once processed through SPM, gray matter maps were smoothed with a 10mm FWHM Gaussian kernel. Gray matter maps were then re-registered from the asymmetric template in SPM to a symmetric MNI template in FSL ^80,81 version 6.0.0 using ANTs^82,83 with nearest neighbor interpolation. Following smoothing and re-normalization, each map was mirrored using AFNI’s (Version AFNI_18.3.01)⁸⁴ 3dLRflip, and the mirror was subtracted from the original smoothed image, and a mask of all positive values of x (left hemispheric) was applied to the difference image to generate a laterality map.

Voxel-wise Analysis:

The laterality map computed using difference images were analyzed using voxel-based morphometry implemented in SPM12 (version 7487). Ordinary least-squares regression was performed at each voxel with the following factors: Diagnosis (CON vs SZ), Gender (M vs F), Handedness (R, L, Ambidextrous), and Site of data acquisition. Comparisons between CON and SZ were computed on the residuals with all other factors treated as nuisance regressors.

Source-Based Laterality:

Previous SBL work in whole-brain information from the same data utilized Akaike’s information criterion⁸⁵ to suggest a 30-component model for source based morphometry analysis⁶⁰. As such, a similar 30 component solution was computed from the 326 laterality maps utilizing entropy bound minimization (EBM)-ICA⁸⁶, an algorithm which is able to capture more flexible source distributions (e.g. sub-Gaussian, symmetric, or skewed) compared to approaches like infomax^87,88 and selection of the best run from 20 for component stabilization.

Multiple regression:

Component weights which are visually identified to be local to gray matter were regressed onto diagnostic status (SZ vs CON), handedness, and data acquisition site. The 30 regressions were multiple comparison corrected using a 5% false discovery rate (FDR)⁸⁹.

Clinical Correlations:

Weights from SZ patients were extracted from components associated with diagnostic status. Spearman’s ρ partial correlations across diagnosis-specific components were performed with positive, negative, and total positive and negative syndrome Scale (PANSS; ⁹⁰ to probe brain/behavior relationships in patients with SZ.

SNC Calculations:

Mixing matrices for components associated with diagnostic status totaling 6 comparisons, were correlated with one another to measure the inter-relationship between covarying networks, or SNC, using a partial correlation to account for data acquisition site. Contrasts were adjusted for multiple comparisons using a Holm correction⁹¹, as an FDR of 5% for 6 comparisons is not an integer (0.03) to examine relationships between components.

Results:

Toy example:

The GIFT ICA pipeline generated three components as anticipated. These included: component 1 (which was resembled source 2), component 2 (which resembled source 1), and a third component labelled as noise due to its appearance (see Figure 2). Images displaying the components from the simulated data are summarized in Figure 2. Comparisons of loading parameters for the simulated CON vs SZ groups identified a significantly greater (W = >0, p < 0.0001, r = 0.865) component weightings for component 1 (source 2) in simulated CONs (median = 0.98) compared to simulated SZ (median = −0.995). The opposite relationship was apparent in component 2 (source 1), with simulated CON exhibiting a significantly greater (W = 22500, p < 0.0001, r = 0.865) median component weights (median = 0.99) compared to simulated CONs (median = −0.99).

Figure 2:

Component estimations resolved from SBL on simulated data. The estimation of source 1 (left) has a distribution and appearance and distribution similar to source 1. The component estimated for source 2 (middle) similarly has a distribution and appearance nearly identical to source 2, also occurring across multiple regions. The third estimation (right) represents a component estimated from sections of Gaussian noise generated as part of the simulated data.

Voxel-wise analysis on laterality maps:

No voxels survived a 5% voxel-wise FDR⁸⁹ correction (t₍₃₁₆₎=4.814, k=14.22). If the cluster extend threshold is removed, a small number of subcortical and frontal regions are present, but these regions appear in regions more associated with dura and CSF and are likely artifact. See Figure 3 for uncorrected image results.

Figure 3:

Statistically significant voxels after a 5% FDR correction (t₍₃₁₄₎ = 4.861), with no cluster thresholding (k = 0). Warm colors indicate CON > SZ, while cool colors indicate SZ > CON. The model used for this analysis treats age, gender, handedness, and data acquisition site as nuisance variables. Results are flagged with red circles for visualization.

SBM component analysis on laterality maps:

Each of the 30 components from the SBM were visually inspected to flag components as “possible source” or “possible artifact.” The former category included regions within the cortical mask, while the latter included ringing around the mask and the inclusion of non-cortical regions. No components exhibited patters defined as “possible artifact,” so loadings from all 30 components were retained for subsequent analyses. An interactive image in html constructed with papaya (build 782a193, https://rii-mango.github.io/Papaya/) may be found in Supplementary Material 1. Of the 30 performed regressions, 10 of the models reached statistical significance after a 5% FDR correction for multiple comparisons. However, only 4 of the component weights (component 7, 16, 28, and 30 were significantly influenced by participant diagnostic status (CON vs SZ). Components significantly related to diagnosis are displayed in Figure 4A, while contrasts are displayed in Figure 4B. F-values, p and corrected p values, adjusted R², and partial omega squared values for each are summarized in Table 2. Post-hoc Wilcoxon rank-sum tests⁹² identified significantly greater cerebellar component weights in CON compared to patients with SZ (W = 16216, p < 0.001, r = 0.19, CI95% 0.17 0.57), but significantly reduced weights in CON participants compared to patients with SZ in superior temporal component, (W = 11586, p < 0.05, r = 0.11 CI95% −0.45 −0.003), STG/MTG/Postcentral component (W = 11416, p < 0.03, r = 0.12, CI95% −0.44 −0.03), and middle temporal gyrus (W = 10357, p < 0.001, r = 0.19, CI95% −0.61 −0.17). Component spatial maps were thresholded at a Z score of 3, (displayed in Figure 4A), and average difference scores for laterality maps for each subject were compared. Differences reflected those displayed by component weights, with CON participants displaying right laterality, and SZ participants displaying left laterality. It should be noted that the weights from component 7 suggest an increase in CON and a reduction in SZ, but the degree to which right lateralization is present in CON participants is greater than the degree of left lateralization in SZ patients. Weights for components 16, 28, and 30 similarly mirror mean distributions in which CON participants display reduced weights compared to SZ patients, but this is relative to right-lateralization (expressed as negative numbers). The ICA based approach highlights the relationship between the two groups agnostic to the direction.

Figure 4:

A) Components significantly associated with CON (Blue) vs SZ. These include component 7 (red/cerebellum), 16 (blue/superior temporal), 28 (green/postcentral, superior, and middle temporal), and 30 (yellow/middle temporal). B) Participants with SZ (red violin) displayed significantly lower weightings in component 7 compared to CON participants (blue violin), but participants with SZ displayed significantly greater weightings in components 16, 28, and 30 compared to CON individuals. C) Average difference score for thresholded (Z = 3) components. Zero on the graph is indicated by a dotted line. Positive values indicate greater gray matter volume in left hemisphere (L), while negative values indicate greater gray matter volume in the right (R) hemisphere. For component 7, participants with SZ exhibit greater left hemispheric gray matter, while CON participants exhibit right laterality. For components 16, 28, and 30, gray matter is larger in the right hemisphere, but the volume is significantly increased in the patients with SZ compared to CON participants. Hemispheric displays are organized relative to participants with SZ.

Table 2.

Linear regression results by component

									Assumption Tests p-values
Component	F(9,317)	F p-value	FDR p-value	Adj. R2	Significant Predictor(s)	Predictor t-value	Predictor p-value	pώ2	Heteroscedasticity	Non-normal Residuals	Autocorrelated Residuals
1	1.79	0.070	0.175	0.021					0.528	0.701	0.996
2	0.95	0.484	0.660	−0.001					0.355	0.355	0.452
3	0.77	0.649	0.778	−0.007					0.724	0.723	0.358
4	1.40	0.187	0.318	0.011					*<0.001	*<0.001	0.986
5	3.51	*<0.001	*0.003	0.065					0.639	0.026	0.928
					Site ID/(Duke)	4.31	<0.001	0.071
6	2.33	*0.015	*0.045	0.036					*0.03	0.705	*0.006
					Site ID/(Duke)	2.35	<0.001	0.036
7	2.80	*0.004	*0.012	0.047					0.375	*<0.001	0.534
					Diagnostic Status (SZ vs CON)	3.57	<0.001	0.033
					Handedness (Left)	2.06	0.040	0.015
8	0.55	0.834	0.863	−0.013					0.765	*0.006	0.836
9	1.35	0.210	0.332	0.010					0.338	0.096	*0.032
10	0.91	0.519	0.677	−0.003					0.741	0.893	0.24
11	7.90	*<0.001	*<0.001	0.161					0.438	*0.031	*0.004
					Site ID/(Duke)	−5.82	<0.001	0.158
					Site ID/(MRN)	1.97	0.050	0.158
12	1.39	0.191	0.318	0.011					0.932	0.562	0.432
13	1.22	0.284	0.426	0.006					0.623	*<0.001	0.892
14	4.34	*<0.001	*<0.001	0.085					0.818	0.136	0.546
					Site ID/(Duke)	4.34	<0.001	0.086
					Site ID/(UCLA)	2.27	0.024	0.086
15	1.40	0.189	0.318	0.011					0.98	0.178	0.608
16	8.42	*<0.001	*<0.001	0.171					0.657	0.22	0.968
					Diagnostic Status (SZ vs CON)	−2.27	0.024	0.012
					Handedness (Ambidextrous)	−1.99	0.047	0.015
					Site ID/(Duke)	−5.84	<0.001	0.151
17	2.98	*0.002	*0.009	0.052					*0.003	*0.027	0.618
					Site ID/(MRN)	3.80	<0.001	0.053
					Site ID/(NICIrvine)	2.28	0.024	0.053
					Site ID/(UCIrvine)	2.04	0.042	0.053
					Site ID/(UCLA)	2.94	0.004	0.053
					Site ID/(UI)	2.85	0.005	0.053
18	0.87	0.554	0.693	−0.004					0.261	0.278	0.384
19	2.03	0.036	0.098	0.028					0.51	0.505	0.272
20	0.68	0.731	0.812	−0.009					0.349	0.12	0.926
21	0.32	0.970	0.970	−0.019					0.762	*0.004	0.642
22	0.61	0.785	0.841	−0.011					0.297	0.113	0.636
23	2.87	*0.003	*0.011	0.049					0.445	0.496	0.17
					Site ID/(UCLA)	2.60	0.010	0.038
24	1.62	0.109	0.219	0.017					0.421	0.121	0.404
25	1.62	0.110	0.219	0.017					0.886	*0.041	0.078
26	0.73	0.677	0.781	−0.007					0.746	0.16	0.232
27	1.75	0.078	0.179	0.020					0.603	0.096	0.734
28	3.04	*0.002	*0.008	0.053					0.309	*<0.001	0.35
					Diagnostic Status (SZ vs CON)	−2.31	0.021	0.014
					Site ID/(MRN)	2.92	0.004	0.046
					Site ID/(UCIrvine)	2.23	0.026	0.046
29	1.05	0.398	0.568	0.001					0.829	0.218	*0.008
30	3.29	*<0.001	*0.005	0.060					0.418	0.257	*0.05
					Diagnostic Status (SZ vs CON)	−3.58	<0.001	0.035
					Site ID/(Duke)	3.14	0.002	0.032
					Site ID/(MRN)	2.24	0.026	0.032

Note:

^*Indicates statistical significance at or below the 0.05 level. p^ώ2 values are calculated by independent variable (e.g. diagnosis, data collection site), not by factor in which dummy variables for each category has been coded. We present the p^ώ2 for the entire independent variable with the factor of interest for ease of visualization.

Duke: Duke University, MRN: Mind Research Network, UCIrvine: University of California Irvine, NIUCIrvine: Neuroimaging Center of University of California Irvine, UCLA: University of California Los Angeles, UI: University of Iowa, CON: Control participant without a diagnosis of schizophrenia, SZ: Participant with a diagnosis of schizophrenia.

Correlations with Clinical Measures:

Partial Spearman ρ correlations accounting for site of data acquisition did not reveal any significant relationship between components for total positive or general total PANSS scores. One statistically significant negative correlation between total negative PANSS symptoms and component 7 was identified which survived a Holm ⁹¹ correction for multiple comparisons (ρ = −0.20, p = 0.009, Holm-p = 0.03). When extracting average gray matter volumes relative to the component, this translates to an increase in right lateralization (negative values) associated with less negative symptoms in patients with SZ. No other statistically significant relationships were identified for components 16, 28, and 30 with PANSS negative symptoms scores. The relationship is described in a spearman plot generated using the fifer (v1.1) package in R⁹³ (see Figure 5).

Figure 5:

Spearman Rank correlation between the residuals of PNASS total negative symptom scores and laterality component 7 weights when accounting for differences in acquisition sites. The line is the linear correlation between ranks for residuals of component 7 weights and total negative symptoms from the PNASS. The histograms on the x and y axis illustrate the (skewed) distributions of the residuals of the variables accounting for data acquisition site.

Structural SNC:

Pearson correlations between the mixing matrices from components 7, 16, 28, and 30 identified a significant negative relationship between component 7 and component 30 which survived a Holm correction⁹¹ for multiple comparisons (r = −0.2, Holm-p = 0.002). Significant uncorrected correlations between component 16 and 28 with component 30 were identified (r = 0.12, p = 0.03 and r = −0.17 p= 0.03 respectively), but these results did not survive correction for multiple comparisons (Holm-p = 0.13). A nodal illustration of the structural SNC matrix with nodal connections are displayed in Figure 6.

Figure 6:

Structural FNC matrix of components 7, 16, 28, and 30 (left). Nodal illustration of all 4 components, with structural connectivity between component 7 and 30 (* on matrix and cyan on nodal illustration) statistically significant following a Holm correction for multiple comparisons.

Discussion:

We demonstrate a novel approach to assess structural laterality within the brain. SBL provides advantages over traditional approaches to studying brain laterality in that it does not require apriori selection of an ROI, preserves spatial information within the data, and identifies covarying patterns of laterality alterations in lieu of potentially skewed voxelwise quantifications which identify differences irrespective of covariation in results. As a demonstration of the utility of SBL, the technique was applied to a cohort of participants with SZ relative to controls. This comparison revealed alterations in multiple covarying networks including left middle and superior temporal regions, and the planum temporale consistent with extensive SZ literature^{2–6,8–10,13–23,63–72}. An exception to this trend is altered subcortical asymmetry in schizophrenia from meta-analyses such and Okada et al., (2016)⁹⁴. Okada et al. (2016)’s meta-analysis of structural laterality of subcortical regions from 2564 data sets across 78 studies identified significant differences in LI specific to the globus pallidus and no other subcortical regions. However, it should be noted that even when effect-size weighted meta-analytic methods are considered, meta-analyses may be influenced by publication bias, where effects within the literature which do not fail-to-reject the null hypothesis are not reported^95–97. While not 100% reflective of the literature, tools are available to provide estimates of such biases^97–101. In addition, it is important to keep in mind that atlas-based ROI and data-driven ROIs have different advantages. To the degree an entire region is changing in volume, an ROI approach may provide enhanced sensitivity. However, if changes are variable across the ROI, more focal, or vary across individuals, a data-driven approach will likely perform better under these circumstances. Given these possibilities and the relatively small effect sizes reported for CON vs SZ individuals across studies (≤ 0.3, see Okada (2016) supplementary table 10), subcortical alterations in laterality may have been negligible to a degree they were not detected in SBL. It will be important to leverage multiple analytic approaches to capture changes via a pluralistic lens.

The results of this pilot SBL approach are the first, to our knowledge, to identify laterality differences within paired, but discrete regions of the temporal lobe. The covariation of separate subregions within the temporal lobe in the real data may reflect separate sources discrete to the numerous functions recruited by the temporal lobes, and relationships these regions have with laterality alterations present in patients with SZ compared to TD participants. While additional work will be required to validate such brain/behavior relationships, this highlights the unique advantage of SBL over traditional approaches in that discrete covarying networks are identified.

The earliest work by Wernicke implied that these regions were critical to language, which has since expanded to include many aspects of social cognition^102,103. Previous work has suggested left anterior portions of the STG/MTG are key for language integration, while posterior STG/MTG are key to speech production and other cognitive functions in a graded manner^102,103. It is entirely possible that components 16, 28, and 30 were separated based on this graded topology, though future studies will be needed to evaluate this hypothesis. Additional behavioral data specific to each language domain (e.g. word generation, integration, phoneme clustering) would be necessary to confirm this hypothesis. Previous literature and the results presented here would suggest alterations in left-lateralized regions specific to language may explain some of the negative symptoms of schizophrenia, however, the significant relationship between negative PANSS scores and the cerebellum may hint towards a rarely explored avenue of research in SZ.

The cerebellum has similarly been linked to language and social cognition. Right cerebellar resection from cranial fossa tumors produce deficits in complex language tasks and speech alterations ranging from mutism to dysarthria^104–106. These symptoms are not unlike some of the negative symptoms described in the PANSS⁹⁰. Previous work has suggested alterations within Purkinje cells and associated proteins within the cerebellum are altered in individuals with schizophrenia, but the impact is not fully understood^107,108. Furthermore, most of these experiments have focused on motor function, a fraction of which is measured by PANSS negative symptom scores. Diffusion weighted imaging tractography studies in humans suggest physical connections are present between the cerebellum and contralateral MTG¹⁰⁹. Additionally, previous work has also found VBM based reductions in white matter within temporal regions in patients with SZ¹¹⁰. Gray matter volume SNC results between the cerebellum and temporal regions may reflect alterations within this pathway.

While this theory is bolstered by the significant negative relationship between negative PANSS symptom scores and the cerebellar component, the relatively small effect sizes for diagnosis for the components (_pώ² = 0.012–0.035), suggests the effect may not be very strong within the modality of structural MRI. We argue that future work using multimodal difference maps (e.g. diffusion anisotropy, fMRI) with fusion approaches such as joint ICA or independent vector analysis (IVA) may be able to capture a structural and functional interaction(s) within patients with SZ and other populations.

The SBL approach has advantages over traditional ROI and voxel-based approaches in that it is can identify covarying spatial laterality patterns, possesses much greater statistical power regarding the number of multiple comparison corrections both for ICA components and a reduction in the total voxels tested, and is relatively easy to implement within the framework of existing ICA tools. SBL component weights appear to reflect differences in hemispheric differences scores (translating to laterality), which can serve as a useful tool for post-hoc volumetric analyses at the ROI or voxel-level. Further work will be required to evaluate the SBL approach for assessing laterality, but these factors suggest the SBL approach is a promising direction studying brain laterality.

Limitations:

While the SBL approach does provide the advantage of localization to covarying regions across hemispheres, interpreting the direction of alterations requires more refinement. In the mirror image, negative values are indicative of greater volume in the right hemisphere, positive values in the left hemisphere, and relative differences of 0 indicate a lack of laterality. The ICA implementation used for the SBL approach will flip signs when estimating components as demonstrated in Figure 4A & 3B. As such, component weights (positive or negative) must be carefully calibrated to represent the direction of lateralization. This is especially important regarding the cerebellum, as the contralateral connections with the cortex may affect the interpretation of the results. As stated in the discussion, the authors have suggested that multi-modal data fusion analyses ¹¹¹ may be able to provide additional information regarding the validity/utility of these effects, and this research venue is recommended prior to any definitive conclusions.

Abbreviations:

SZ

Participants with schizophrenia

CON

participants without schizophrenia

VBM

Voxel-based morphometry

LI

Laterality index

SBL

Source-based laterality

BOLD

Blood-oxygen-level-dependence

BIRN

Biomedical informatics research network

MRN

Mind Research Network

UCLA

University of California Los Angeles

UCIrvine

University of California Irvine

NICIrvine

Neuroimaging Center of University of California Irvine

UIowa

University of Iowa

UMCMRR

University of Minnesota Medical School’s Center for Magnetic Resonance

Research NITRC

Neuroimaging tools and research collaboratory

FWHM

Full width half-maximum

INU

Intensity non-uniformity

SNC

Structural network connectivity

FDR

False discovery rate

PANSS

positive and negative syndrome scale for schizophrenia

EBM

Entropy bound minimization

STG

Superior temporal gyrus

MTG

Middle temporal gyrus

IVA

Independent vector analysis