Elsevier

NeuroImage

Volume 120, 15 October 2015, Pages 43-54
NeuroImage

Decomposition of brain diffusion imaging data uncovers latent schizophrenias with distinct patterns of white matter anisotropy

https://doi.org/10.1016/j.neuroimage.2015.06.083Get rights and content

Highlights

  • Fractional anisotropy (FA) analysis of DTI images can produce disparate results.

  • Inconsistencies may arise from averaging heterogeneous groups of patients.

  • We developed a method capable of identifying heterogeneous but meaningful patterns.

  • The patterns were significantly associated with clinical symptoms in schizophrenia.

  • Our results suggest that schizophrenia is a heterogeneous group of disorders.

Abstract

Fractional anisotropy (FA) analysis of diffusion tensor-images (DTI) has yielded inconsistent abnormalities in schizophrenia (SZ). Inconsistencies may arise from averaging heterogeneous groups of patients. Here we investigate whether SZ is a heterogeneous group of disorders distinguished by distinct patterns of FA reductions. We developed a Generalized Factorization Method (GFM) to identify biclusters (i.e., subsets of subjects associated with a subset of particular characteristics, such as low FA in specific regions). GFM appropriately assembles a collection of unsupervised techniques with Non-negative Matrix Factorization to generate biclusters, rather than averaging across all subjects and all their characteristics. DTI tract-based spatial statistics images, which output is the locally maximal FA projected onto the group white matter skeleton, were analyzed in 47 SZ and 36 healthy subjects, identifying 8 biclusters. The mean FA of the voxels of each bicluster was significantly different from those of other SZ subjects or 36 healthy controls. The eight biclusters were organized into four more general patterns of low FA in specific regions: 1) genu of corpus callosum (GCC), 2) fornix (FX) + external capsule (EC), 3) splenium of CC (SCC) + retrolenticular limb (RLIC) + posterior limb (PLIC) of the internal capsule, and 4) anterior limb of the internal capsule. These patterns were significantly associated with particular clinical features: Pattern 1 (GCC) with bizarre behavior, pattern 2 (FX + EC) with prominent delusions, and pattern 3 (SCC + RLIC + PLIC) with negative symptoms including disorganized speech. The uncovered patterns suggest that SZ is a heterogeneous group of disorders that can be distinguished by different patterns of FA reductions associated with distinct clinical features.

Introduction

White matter lesions in the brain are generally associated with abnormal connectivity of brain regions in complex diseases such as schizophrenia (SZ) (Lee et al., 2013, Liu et al., 2013, Skudlarski et al., 2013). It has been reported that white matter abnormalities may suggest clues relevant to the neurodevelopmental origin of these diseases (Huang et al., 2011, Keller et al., 2007, Lee et al., 2013). To allow the investigation of white matter structural abnormalities in vivo, diffusion tensor imaging (DTI) has been widely used; this magnetic resonance imaging technique measures localized water diffusivity reflecting the geometric properties and directionality of both the axonal membrane and myelin in large white matter tracts of the brain (Mori, 2007). Fractional anisotropy (FA) is a scalar measure often used in DTI, which describes the directional dependence of diffusion (i.e., anisotropy); it is related to axonal fiber density, axonal diameter and myelination in white matter (Mori, 2007, Song et al., 2002). Unfortunately, DTI findings in SZ have been inconsistent across studies (Alba-Ferrara and de Erausquin, 2013, Kubicki et al., 2007). Specifically, studies have either reported no white matter FA differences between controls and patients with SZ (Foong et al., 2002, Suddath et al., 1990, Wible et al., 1995), minimal regional FA abnormalities (Lee et al., 2013, Liu et al., 2013, Skudlarski et al., 2013), or widespread FA abnormalities (Douaud et al., 2007, Lim et al., 1999, Minami et al., 2003).

The inconsistencies in SZ studies may arise from averaging heterogeneous groups of patients with varying FA abnormalities (Fig. 1). Identification of relevant changes in FA is particularly challenging when subjects vary extensively in their clinical features and severity, which are likely to have complex relations with different brain structures and functions (Blanchard and Cohen, 2006, Holliday et al., 2009). Regional differences can be missed due to the fuzziness of the data, particularly when the abnormality is mild (Alba-Ferrara and de Erausquin, 2013).

To characterize the heterogeneity of structural brain abnormalities in SZ, we devised an unsupervised machine learning approach that decomposes a collection of FA images from patients with SZ into local partitions or biclusters (Cichocki, 2009, Pascual-Montano et al., 2006b, Tamayo et al., 2007). These biclusters are composed of co-differentiated white matter FA sets of voxels (i.e., voxels with low FA values) shared by subsets of subjects. Biclustering captures local and intrinsic relationships between subsets of observations (subjects) sharing subsets of descriptive features (voxels) instead of relationships between all subjects and all their descriptive features. These relationships can be weakened when all features are used in a single global model of data, as is typically done by clustering methods.

Our approach combines the advantages of a number of complementary clustering strategies into a Generalized Factorization Method (GFM, Supplementary Fig. 1) and has been previously widely applied in different biomedical problems (Harari et al., 2010, Romero-Zaliz et al., 2008b, Zwir et al., 2005b). Non-negative Matrix Factorization (NMF) algorithms have also been utilized in facial recognition (Lee and Seung, 1999), gene expression (Tamayo et al., 2007), and several other biomedical problems (Cichocki, 2009). More recently, we successfully applied a composite GFM–NMF to uncover eight different subtypes of SZ by dissecting genome wide association studies into biclusters composed of distinct sets of genetic variants and clinical symptoms of SZ patients (Arnedo et al., 2013, Arnedo et al., 2014).

In the current study, we applied the GFM–NMF methodology to examine a sample of SZ patients and healthy controls whose diffusion-weighted brain images were processed using Tract Based Spatial Statistics (TBSS) (Smith et al., 2006). We refer to the output of TBSS, which is the locally maximal FA projected onto the group white matter skeleton, as an FA-TBSS image. We searched for biclusters reflecting different FA patterns that can be shared by distinct subsets of SZ patients. Then we evaluated the significance of each bicluster by comparing the differential FA within a bicluster with that exhibited by healthy controls, as well as with that shown in other individuals with SZ who were not present in the bicluster. We then analyzed the anatomical location of DTI abnormalities identified in the biclusters. In the final step, we cross-correlated the uncovered biclusters with collected descriptions of clinical features of the patients including positive and negative symptoms scores as defined by the Scale for the Assessment of Positive Symptoms (SAPS) and the Scale for the Assessment of Negative Symptoms (SANS) (Andreasen, 1984).

The method illustrated here is able to agnostically decompose FA-TBSS images and to distinguish subsets of SZ patients using white matter FA patterns, just as we decomposed SZ into subtypes with complex relations between sets of genotypes and sets of clinical phenotypes (Arnedo et al., 2014). These abnormalities may suggest distinct etiologies in patients diagnosed with SZ, characterized by different brain areas leading to distinct symptoms and clinical outcomes. The software is available upon request from the authors.

Section snippets

Study sample

The participants included in the study were drawn from a population of volunteers for studies of brain structure and function at the Conte Center for the Neuroscience of Mental Disorders at Washington University Medical School, St. Louis. All participants gave written informed consent for participation following a complete description of the risks and benefits of the study.

Participants consisted of 47 individuals (mean age = 37.2 yrs; SD = 8.5) who met DSM-IV (American_Psychiatric_Association, 1994)

Statistical analysis

Identification of biclusters was unbiased without a prior knowledge of the anatomical location of voxels and/or the clinical symptoms of the subjects. Using internal criteria in cluster evaluation is biased towards algorithms that use the same cluster model (Bezdek, 1998). Therefore, external evaluations based on criteria that were not used for clustering, such as tests based on ANOVA and its F-statistic, are often added to the cluster evaluation (Färber et al., 2010). To assess the statistical

Biclusters encode sets of voxels with FA reductions shared by subsets of subjects with SZ

We first investigated decreased FA regions in a sample composed of TBSS images from 47 SZ patients, which together did not exhibit significant differences from a similar sample composed of 36 images of healthy controls (see Statistical analysis, Methods, p-value > 0.81). A partition of the images corresponding to the patients with SZ using GFM–NMF uncovered eight optimal local partitions or biclusters, where each bicluster encoded a subset of subjects characterized by a similar degree of FA

Discussion

Studies of white matter using diffusion weighted MRI in SZ have reported decreased FA in many regions, but increased FA in some tracts has also been reported (Douaud et al., 2007, Kubicki et al., 2007). Findings about the degrees of FA in the literature have also been inconsistent. Recently, both a meta-review (Shepherd et al., 2012) and a large meta-analysis (Haijma et al., 2013) failed to find consistent reductions in white matter. White matter abnormalities in SZ may be progressive (Whitford

Acknowledgments

This work was supported in part by the Spanish Ministry of Science and Technology TIN2009-13950, TIN2012-38805 including FEDER funds, the R.L. Kirschstein National Research Award 5 T32 DA 7261-23 to I.Z.; the National Institutes of Health including grant 5K08MH077220 to G.AdeE; K08MH085948 to D.M., and the National Institute of Mental Health MH066031 to D.M.B. G.A.deE is a Stephen and Constance Lieber Inverstigator, and Sidney R. Baier Jr. Investigator, as well as Roksamp Chair of Biological

References (68)

  • S.M. Smith et al.

    Tract-based spatial statistics: voxelwise analysis of multi-subject diffusion data

    NeuroImage

    (2006)
  • S.K. Song et al.

    Dysmyelination revealed through MRI as increased radial (but unchanged axial) diffusion of water

    NeuroImage

    (2002)
  • G. Spalletta et al.

    Brain white matter microstructure in deficit and non-deficit subtypes of schizophrenia

    Psychiatry Res.

    (2015)
  • L.M. Alba-Ferrara et al.

    What does anisotropy measure? Insights from increased and decreased anisotropy in selective fiber tracts in schizophrenia

    Front. Integr. Neurosci.

    (2013)
  • American_Psychiatric_Association

    Diagnostic and Statistical Manual of Mental Disorders

    (1994)
  • N.C. Andreasen

    Scale for the Assessment of Positive Symptoms (SAPS)

    (1984)
  • J. Arnedo et al.

    PGMRA: a web server for (phenotype × genotype) many-to-many relation analysis in GWAS

    Nucleic Acids Res.

    (2013)
  • J. Arnedo et al.

    A multiobjective method for robust identification of bacterial small non-coding RNAs

    Bioinformatics

    (2014)
  • J.C. Bezdek

    Pattern Recognition With Fuzzy Objective Function Algorithms

    (1981)
  • J.C. Bezdek

    Pattern analysis

  • T. Bittner et al.

    A theory of granular partitions

    Foundations of Geographic Information Science

    (2003)
  • J.J. Blanchard et al.

    The structure of negative symptoms within schizophrenia: implications for assessment

    Schizophr. Bull.

    (2006)
  • A. Cichocki

    Nonnegative Matrix and Tensor Factorizations: Applications to Exploratory Multi-way Data Analysis and Blind Source Separation

    (2009)
  • O. Cordon et al.

    Linguistic modeling by hierarchical systems of linguistic rules

    IEEE Trans. Fuzzy Syst.

    (2002)
  • K. Deb

    Multi-objective Optimization Using Evolutionary Algorithms

    (2001)
  • K. Deb

    Nonlinear goal programming using multi-objective genetic algorithms

    J. Oper. Res. Soc.

    (2001)
  • G. Douaud et al.

    Anatomically related grey and white matter abnormalities in adolescent-onset schizophrenia

    Brain

    (2007)
  • F. Duan et al.

    Test–retest reliability of diffusion measures in cerebral white matter: a multiband diffusion MRI study

    J. Magn. Reson. Imaging

    (2015)
  • I. Färber et al.

    On using class-labels in evaluation of clusterings

  • M.B. First et al.

    Structured Clinical Interview for DSM-IV Axis I Disorders, Research Version, Patient Edition (SCID-I/P)

    (2002)
  • J. Foong et al.

    Investigating regional white matter in schizophrenia using diffusion tensor imaging

    Neuroreport

    (2002)
  • C. Fraley et al.

    How many clusters? Which clustering method? Answers via model-based cluster analysis

    Comput. J.

    (1998)
  • A.L. Fred et al.

    Combining multiple clusterings using evidence accumulation

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2005)
  • A.A. Freitas

    A critical review of multi-objective optimization in data mining: a position paper

    SIGKDD Explor.

    (2004)
  • Cited by (37)

    • White matter integrity in schizophrenia and bipolar disorder: Tract- and voxel-based analyses of diffusion data from the Connectom scanner

      2019, NeuroImage: Clinical
      Citation Excerpt :

      The mean FA for each subject's tracts was calculated by TRACULA based on the probabilistic fibers that were at least 20% of the maximum probability fiber for each pathway (Yendiki et al., 2011). Tract-based spatial statistics (TBSS) was used to perform voxel-wise analyses of all white matter tracts, as previously described (Arnedo et al., 2015; Smith et al., 2006). FA images were calculated and projected onto the mean FA skeleton, which represents the center of white matter tracts, and thresholded at FA = 0.2.

    • A systematic review of task-based functional neuroimaging studies investigating language, semantic and executive processes in thought disorder

      2018, Neuroscience and Biobehavioral Reviews
      Citation Excerpt :

      Additional functional correlates identified in the current review encompassed the cerebellum, caudate nucleus, cingulate cortex, and fusiform gyrus, amongst others. Moreover, structural correlates of TD include the orbitofrontal cortex (Gur et al., 2004; Horn et al., 2010; Nakamura et al., 2008; Sans-Sansa et al., 2013), corpus callosum (Arnedo et al., 2015; Kubicki et al., 2008), amygdala and hippocampus (Bogerts et al., 1993; Rajarethinam et al., 2001; Sallet et al., 2003; Spalletta et al., 2010), nucleus accumbens (Ballmaier et al., 2004; Spalletta et al., 2010), and cerebellum (Sandyk et al., 1991); and there is evidence of reduced striatal dopamine D2 receptor availability in patients with TD, though this may not be specific to TD (Schmitt et al., 2008). Therefore, the neural substrate of TD is not constrained to superior and middle temporal, and inferior frontal gyri.

    View all citing articles on Scopus
    View full text