Elsevier

NeuroImage

Volume 34, Issue 4, 15 February 2007, Pages 1405-1415
NeuroImage

Pattern classification using principal components of cortical thickness and its discriminative pattern in schizophrenia

https://doi.org/10.1016/j.neuroimage.2006.11.021Get rights and content

Abstract

We proposed pattern classification based on principal components of cortical thickness between schizophrenic patients and healthy controls, which was trained using a leave-one-out cross-validation. The cortical thickness was measured by calculating the Euclidean distance between linked vertices on the inner and outer cortical surfaces. Principal component analysis was applied to each lobe for practical computational issues and stability of principal components. And, discriminative patterns derived at every vertex in the original feature space with respect to support vector machine were analyzed with definitive findings of brain abnormalities in schizophrenia for establishing practical confidence. It was simulated with 50 randomly selected validation set for the generalization and the average accuracy of classification was reported. This study showed that some principal components might be more useful than others for classification, but not necessarily matching the ordering of the variance amounts they explained. In particular, 40–70 principal components rearranged by a simple two-sample t-test which ranked the effectiveness of features were used for the best mean accuracy of simulated classification (frontal: (left(%)|right(%)) = 91.07|88.80, parietal: 91.40|91.53, temporal: 93.60|91.47, occipital: 88.80|91.60). And, discriminative power appeared more spatially diffused bilaterally in the several regions, especially precentral, postcentral, superior frontal and temporal, cingulate and parahippocampal gyri. Since our results of discriminative patterns derived from classifier were consistent with a previous morphological analysis of schizophrenia, it can be said that the cortical thickness is a reliable feature for pattern classification and the potential benefits of such diagnostic tools are enhanced by our finding.

Introduction

Recent advances in magnetic resonance image acquisition and processing have allowed for the morphometric analysis of the cerebral cortex at a macroscopic level and have allowed an investigation of normal and abnormal changes. Much structural magnetic resonance imaging (MRI) of the brain in schizophrenia has indicated subtle cortical abnormalities compared with healthy controls. Specifically, gray matter (GM) deficits in superior and medial temporal cortices have been widely reported, but there is moderate evidence of focal GM volume reductions in frontal, parietal, and occipital neocortices, and subcortical abnormalities (Shenton et al., 2001). Although voxel-based morphometry (VBM), which involves a voxel-wise comparison of local GM concentration, is the most widely used approach, it might in part reflect differences in the surrounding tissue and might be influenced by the sulcal widening in schizophrenia (Ashburner and Friston, 2000, Narr et al., 2005a, Shenton et al., 2001). Although GM concentration reflects the proportion of GM within cortical mantle with respect to other tissue types, and cortical thickness represents the distance across the cortex according to some geometric definition, it was reported that these measures are highly correlated, but none of several potential defects in GM concentration were associated with the cortical thickness (Narr et al., 2005a). Therefore, it can be said that measuring cortical thickness provides a closer approximation to the underlying anatomical reality and a direct quantitative index of cortical morphology.

Several postmortem studies have assessed cortical thickness in schizophrenia, but they are limited by labor-intensive procedures, making it impractical to measure cellular density and thickness in all cortical regions (Selemon, 2004, Selemon et al., 1995). However, in vivo data with the latest computational neuroimage analysis methods may allow differences in cortical thickness to be estimated from the nodes of a 3D polygonal mesh rather than from a 3D voxel grid. The surface-based approach has the following additional advantages over image- or voxel-based approaches. First, it can be applied in more general situations where a surface is not embedded in an image, but is defined in another way such as segmented boundaries or triangulations. Second, unless the appearance inside the object is also the focus of interest, it may be more appropriate for shape analysis, as the boundary or surface of a volumetric object actually defines the shape. Finally, some noise generated from resampling in the voxel-based analysis can be avoided. Few prior surface-based approaches have examined cortical thickness in schizophrenia (Kuperberg et al., 2003, Narr et al., 2005a, Narr et al., 2005b, White et al., 2003, Wiegand et al., 2004). Kuperberg et al. (2003) assessed thinning across the entire cortex and showed widespread significant thinning that particularly affected the prefrontal and temporal cortices in chronic schizophrenia. White et al. (2003) reported significant cortical thinning in cortex underlying the sulci in frontal, temporal and parietal regions and beneath the gyri in the temporal lobe in patients with childhood and adolescent onset schizophrenia. Another study examining cortical thickness averaged across the entire prefrontal cortex failed to detect significant cortical thinning in first-episode schizophrenia (Wiegand et al., 2004). Narr et al. (2005a) revealed significant regional GM thinning in the frontal, temporal and parietal heteromodal association cortices bilaterally in first-episode schizophrenia. Therefore, it could be expected that these findings related to differences of cortical thickness as a quantitative index of cortical morphology contribute affirmatively toward a categorization between schizophrenia and healthy control.

Because schizophrenia is a large-scale disorder of neurocognitive networks rather than confined to specific regions, and structural changes are present in multiple brain regions, it can be anatomically characterized by abnormality at a supra-regional level of brain organization (Burns et al., 2003, Lawrie and Abukmeil, 1998, Shenton et al., 2001, Wright et al., 1999b). Based on previous findings, Davatzikos et al. (2005) performed whole-brain analysis of structural differences between schizophrenic patients and healthy controls, and applied a high-dimensional nonlinear pattern classification technique to quantify the degree of separation between patients and controls. They were able to classify new individuals as schizophrenic or healthy with 81% accuracy and suggested the potential utility of MRI as a diagnostic aid. Principal component analysis (PCA) is a multivariate method identifying correlation among a set of measurements or variables so that it obtains a set of basis vectors whose linear combination can optimally represent the measured data. And it may also be used to obtain a low-dimensional representation of the measurements themselves. Even if the results may be unstable where the number of subjects is much smaller than the number of variables, dimensionality reduction of PCA is very effective in classification because a higher number of features will easily lead the classifier into the problem of overfitting. Initially, Olson and Miller (1958) proposed that anatomical structures could be recomposed into supra-regional systems by PCA of the covariance or correlations between regional elements (Wright et al., 1999b). Since then, regional elements of supra-regional systems defined by PCA have often been found to share developmental influences or to have a common function (Cheverud, 1982). Narr et al. (2005a) used PCA to reduce cortical thickness measures into principal components, but they examined global effect of cortical GM. Although several shape classification studies have been conducted for discovering hippocampal shape abnormality in schizophrenia, to our knowledge this is the first classification study based on principal components of cortical thickness measured at the spatially homologous cortical surface locations in each individual.

The aim of classification is to instruct the classifier using a training set or set of labeled examples representing different classes, and then use the classifier to predict the class of any new example. This constitutes the final goal of the learning stage in many application domains, including character recognition and text classification. In medical image analysis, however, it has been much more useful in understanding the nature of the differences captured by the classifier than in using it for labeling new examples. These differences, expressed in terms of the original images or shapes, can provide an insight into the anatomical implications of shape differences detected by the learning algorithm. Furthermore, it could be argued that studying the structure of the data captured by the classifier is important in any application, because it puts an emphasis on the nature of the differences between the classes and can potentially help improve the technique. Golland et al. (2005) introduced the notion of the discriminative direction at every point in the feature space with respect to a given classifier, which corresponds to the maximum changes in the classifier's response while minimizing irrelevant changes in the input. It allows to characterize shape differences between the two classes captured by the classifier and to express them as deformations of the original shape. Shen et al. (2004) proposed discriminative patterns that shared a similar idea of discriminative direction: for a linear classifier, the deformation representing class differences could be visualized using the normal to the separating hyperplane. In this study, we proposed pattern classification based on principal components of cortical thickness between schizophrenic patients and healthy controls, and validated its accuracy using a leave-one-out cross-validation (LOOCV) method. Discriminative patterns derived at every vertex in the original feature space with respect to a given classifier were analyzed with definitive findings of brain abnormalities in schizophrenia for establishing practical confidence. The purpose of this study is identification of representative regions contributing to the classification through a discriminative pattern. Although it is less conservative than statistical t-test, our method would be useful to grasp the trend of difference between healthy control and schizophrenia in cortical thickness.

Section snippets

Subjects

A group of right-handed schizophrenic patients was recruited from the inpatient unit and the outpatient clinic at Seoul National University Hospital, Seoul, Korea. Fifty-three patients (32 men, 21 women) were interviewed using a structured clinical interview based on DSM-IV (SCID-IV) and met those criteria for schizophrenia. The Structured Clinical Interview (SCID) was first devised with the DSM-III-R (APA, 1987), and it was modified in 1997 to conform to the new DSM-IV system of diagnosis (

Classification results using different feature selection schemes

Classification can be performed using just the first few principal components, which account for significant amount of data variance, based on the hypothesis that this information is crucial for classification and the rest noisy. However, judging from the following results, it seems not to be acceptable. Fig. 2(a) shows the classification results on each lobe using different feature selection schemes according to the ordering of principal components: decreasing arrangement by the variance

Principal component of cortical thickness as a feature for pattern classification

The original data are represented by fewer variables with minimal mean square error as a result of PCA, which reduces the dimensionality of the dataset. However, one of the limitations is that PCA only defines a single projection of the data. For more complex data like in this study, different clusters may require different projection directions. In addition, even if uncorrelated, the principal components might be highly statistically dependent. The other limitation of PCA is that the original

Acknowledgments

This work was supported by the research fund of Hanyang University (HY-2004-N).

References (59)

  • J.P. Lerch et al.

    Mapping anatomical correlations across cerebral cortex (MACACC) using cortical thickness from MRI

    NeuroImage

    (2006)
  • D. MacDonald et al.

    Automated 3-D extraction of inner and outer surfaces of cerebral cortex from MRI

    NeuroImage

    (2000)
  • R.W. McCarley et al.

    MRI anatomy of schizophrenia

    Biol. Psychiatry

    (1999)
  • K.L. Narr et al.

    Cortical thinning in cingulate and occipital cortices in first episode schizophrenia

    Biol. Psychiatry

    (2005)
  • S. Robbins et al.

    Tuning and comparing spatial normalization methods

    Med. Image Anal.

    (2004)
  • M.E. Shenton et al.

    A review of MRI findings in schizophrenia

    Schizophr. Res.

    (2001)
  • T. White et al.

    Gyrification abnormalities in childhood- and adolescent-onset schizophrenia

    Biol. Psychiatry

    (2003)
  • L.C. Wiegand et al.

    Prefrontal cortical thickness in first-episode psychosis: a magnetic resonance imaging study

    Biol. Psychiatry

    (2004)
  • I.C. Wright et al.

    Mapping of grey matter changes in schizophrenia

    Schizophr. Res.

    (1999)
  • U. Yoon et al.

    Quantitative analysis of group-specific brain tissue probability map for schizophrenic patients

    NeuroImage

    (2005)
  • C. Ambroise et al.

    Selection bias in gene extraction on the basis of microarray gene-expression data

    Proc. Natl. Acad. Sci. U. S. A.

    (2002)
  • APA

    Diagnostic and Statistical Manual of Mental Disorders

    (1987)
  • R.W. Buchanan et al.

    Structural evaluation of the prefrontal cortex in schizophrenia

    Am. J. Psychiatry

    (1998)
  • C.J.C. Burges

    A tutorial on support vector machines for pattern recognition

    Data Mining Knowl. Discov.

    (1998)
  • J. Burns et al.

    Structural disconnectivity in schizophrenia: a diffusion tensor magnetic resonance imaging study

    Br. J. Psychiatry

    (2003)
  • J.M. Cheverud

    Phenotypic, genetic, and environmental morphological integration in the cranium

    Evolution

    (1982)
  • M.K. Chung et al.

    Tensor-based Surface Morphometry

    (2002)
  • D.L. Collins et al.

    Automatic 3D intersubject registration of MR volumetric data in standardized Talairach space

    J. Comput. Assist. Tomogr.

    (1994)
  • R.W. Cox et al.

    Software tools for analysis and visualization of fMRI data

    NMR Biomed.

    (1997)
  • Cited by (0)

    Pattern classification using cortical thickness.

    View full text