Discrimination of female schizophrenia patients from healthy women using multiple structural brain measures obtained with voxel-based morphometry


Correspondence: Miho Ota, MD, PhD, Department of Mental Disorder Research, National Institute of Neuroscience, National Center of Neurology and Psychiatry, 4-1-1 Ogawa-Higashi, Kodaira, Tokyo 187-8502, Japan. Email: ota@ncnp.go.jp



Although schizophrenia and control subjects differ on a variety of neuroanatomical measures, the specificity and sensitivity of any one measure for differentiating between the two groups are low. To identify the correlative pattern of brain changes that best discriminate schizophrenia patients from healthy subjects, discriminant analysis techniques using voxel-based morphometry were applied.


The first analysis was conducted to obtain a statistical model that classified 105 female healthy subjects and 38 female schizophrenia patients. First, the differences in gray matter and cerebrospinal fluid volume between the patients and healthy subjects were evaluated using optimized voxel-based morphometry. Then, a discriminant analysis reflecting the results of this evaluation was adopted. The second analysis was performed to prospectively validate the statistical model by successfully classifying a new group that consisted of 23 female healthy subjects and 23 female schizophrenia patients.


The use of these variables resulted in correct classification rates of 0.72 in the control subjects and 0.76 in the schizophrenia patients. In the second validation analysis using these variables, correct classification rates of 0.70 in the control subjects and 0.74 in the schizophrenia patients were achieved.


Schizophrenia patients have structural deviations in multiple brain areas, and a combination of structural brain measures can distinguish between patients and controls.

THE OPERATIONAL DIAGNOSTIC criteria in the DSM-IV for schizophrenia are based on clinical manifestations and associated psychosocial impairments. Since the establishment of DSM-IV, a number of neuroimaging studies using magnetic resonance imaging (MRI) have demonstrated subtle but significant structural changes, such as dilatation of the Sylvian fissure and the third ventricle, and reduction of the volume of the frontal and temporal lobes in schizophrenia.[1-3] Some studies have attempted to discriminate between patients with schizophrenia and healthy subjects using brain anatomical structures obtained on MRI.[4-7] Recently, some studies have reported on an unbiased, rater-independent technique known as the voxel-based morphometry (VBM)-based classification approach, and on the Freesurfer image analysis suite, which is documented and freely available for download online (http://surfer.nmr.mgh.harvard.edu/).[8-12] Most previous studies involving volume data for gray matter, however, analyzed only one case–control sample; that is, they did not validate their discriminant function analysis by applying the same parameters to a second case–control sample.[4-9] To our knowledge, only a few studies have validated the method using two independent samples,[10-12] but their samples were relatively small. The aging-related morphometric change of the brain needs to be considered in discriminant function analysis. Some VBM studies with healthy subjects found that increasing age was associated with decreases in gray matter density throughout the brain.[13, 14] Others found a negative correlation of age with volume in the bilateral occipital, and parietal regions, in the post-central gyri, and in the insulae.[15, 16] Another found that the magnitude of the effect greatly differed across regions, and the largest age-related decreases in gray matter density were found in various frontal regions (right frontal pole, left dorsolateral prefrontal cortex, anterior cingulate and the anterior part of the insula bilaterally), in the temporal lobes (left hippocampus and middle and superior temporal gyrus), and in the primary visual cortex.[14] Bonilha et al. found that the age-related brain shrinkage had an asymmetrical pattern.[17]

In the present study, we hypothesized that the characteristic distribution of regional gray matter changes in schizophrenia patients in addition to the age-related changes would have diagnostic value to discriminate such patients from healthy subjects. There have been many reports of gender difference in structural brain abnormalities in schizophrenia.[18] Additionally, it is also known that the duration of schizophrenia illness has a sex-specific pattern in the progression of symptoms.[19] Taking sex-difference into account, we evaluated the characteristic distribution of regional gray matter changes only in female subjects.

The analysis used in the present study consisted of two stages. The first involved production of a statistical model to classify subjects according to the current diagnostic systems, and the second involved prospective validation of the statistical model by classifying a new cohort.



Subjects were assigned to one of two independent groups according to the timing of their participation. The first sample consisted of 38 patients with schizophrenia who were diagnosed according to DSM-IV and 105 healthy subjects. The second sample for the validation consisted of 23 individuals with schizophrenia and 23 healthy controls. All subjects were Japanese women. There was no significant difference in age between patients and controls in each sample. The demographic and clinical data are listed in Table 1.

Table 1. Subject characteristics
Female (n = 38)Female (n = 105)
Mean ± SDMean ± SD
Original study  
Age (years)46.6 ± 14.042.0 ± 13.0
Education (years)12.6 ± 2.4*15.3 ± 2.8
Whole brain volume (L)1.1 ± 0.11.1 ± 0.1
Age at onset (years)28.6 ± 11.0 
Medication (mg/day)709.4 ± 584.7 
 Female (n = 23)Female (n = 23)
  1. *P < 0.05 compared with control (two-tailed t-test).
Validation study  
Age (years)42.4 ± 13.246.4 ± 9.7
Education (years)13.4 ± 2.513.9 ± 2.6
Whole brain volume (L)1.1 ± 0.11.1 ± 0.1
Age at onset (years)23.7 ± 9.5 
Medication (mg/day)594.9 ± 534.9 

Consensus diagnosis by at least two psychiatrists was made for each patient based on all the available information obtained from interviews and medical records. Healthy controls were interviewed for enrollment by research psychiatrists using the Japanese version of the Mini-International Neuropsychiatric Interview (MINI).[20, 21] Those who had no current or past history of psychiatric illness or contact with psychiatric services were enrolled as controls. Participants were excluded if they had a prior medical history of central nervous system disease or severe head injury. The study protocol was approved by the ethics committee of the National Center of Neurology and Psychiatry, Japan, and written informed consent for participation in the study was obtained from all subjects.

MRI data acquisition and processing

The first series of MRI was performed on a 1.5-T Magnetom Vision Plus system (Siemens, Erlangen, Germany). 3-D volumetric acquisition of a T1-weighted gradient echo sequence produced a gapless series of 144 sagittal sections using a magnetization prepared rapid acquisition with gradient echo (MPRAGE) sequence (echo time [TE]/repetition time [TR], 4.4/11.4 ms; flip angle, 15°; acquisition matrix, 256 × 256; 1NEX; field of view [FOV], 315 × 315 mm2; slice thickness, 1.23 mm). The second MRI series was performed on a 1.5-T Magnetom Symphony (Siemens). The 3-D T1-weighted images were scanned in the sagittal plane (TE/TR, 2.64/1580 ms; flip angle, 15°; slab thickness, 177 mm; matrix, 208 × 256; 1NEX; FOV, 256 × 315 mm2; slice thickness, 1.23 mm) yielding 144 contiguous slices through the head. The raw 3-D T1-weighted volume data were transferred to a workstation, and structural images were analyzed using an optimized VBM technique. Data were analyzed using SPM5 (Welcome Department of Imaging Neuroscience, London, UK) running on MATLAB 7.0 (Math Works, Natick, MA, USA). Images were processed using an optimized VBM script according to Good et al.[15] First, each individual 3-D T1 image was normalized to the standard space of the SPM5 with the optimized VBM method. Normalized segmented images were modulated by multiplication with Jacobian determinants of the spatial normalization function to encode the deformation field for each subject as tissue density changes in normal space. Images were smoothed using a 12-mm full-width at half-maximum of an isotropic Gaussian kernel.

Statistical analysis

First, we evaluated the differences between the patients and healthy subjects in each of the two case–control samples using two-sample t-tests. These tests were performed using SPSS version 11 (SPSS Japan, Tokyo, Japan). There was no significant difference in age and whole brain volume between the patients and controls for either sample.

We then evaluated the differences in the regional gray matter and cerebrospinal fluid (CSF) volumes between the patients and controls from the first cohort. Statistical analyses were performed using SPM5 (Welcome Department of Imaging Neuroscience). We examined the differences in regional gray matter volume by analysis of covariance (ANCOVA), controlling for age. Only the associations that met the following criteria were deemed statistically significant for post-hoc analysis: a seed level of P < 0.001 (uncorrected) and a cluster level of P < 0.05 (uncorrected).

Second, discriminant function analyses were conducted to assess the ability of a combination of brain anatomical variables to distinguish between schizophrenia patients and controls. The independent variable was the volume of a particular tissue (gray matter or CSF) derived from the normalized individual image using the regions of interest (ROI) method. To begin with, according to the anatomical location, rectangle ROI were manually inserted on the regions of ‘avg152T1.nii’, the standard brain space of SPM5, where the aforementioned analysis indicated a significant volume difference between the patients and controls; and on the inferior occipital and superior parietal cortices, where previous studies and the present one found no difference in volume between schizophrenia patients and controls.[1-3] The particular tissue volumes of individual normalized image surrounded by the ROI were extracted using the software MarsBar (a toolbox for SPM that provides routines for ROI analysis, such as extraction of data for ROI with and without SPM preprocessing).[22] Discriminant functions were derived by stepwise methods after Wilks' method. The stepwise selection criteria were determined by the overall multivariate F of each variable to test the differences between the patients and controls and to maximize the discriminant function between the groups. An enter-criterion was F = 1, and a remove-criterion was F = 0.5. Third, an analysis was performed to prospectively validate the statistical model, the linear discriminant function analysis, by successfully classifying the second case–control sample.


Volume differences between the schizophrenia patients and healthy subjects

In the first case–control sample, there was statistically significant gray matter volume reduction in the bilateral insulae–superior temporal regions and medial frontal gyrus in female schizophrenia patients compared with healthy women. No increase was detected in patients with schizophrenia. Significant CSF dilatations were found in the third ventricle, bilateral Sylvian fissures, and around the prefrontal regions in the patient group (Fig. 1). There were no volume differences in the superior parietal and occipital regions between schizophrenia patients and controls.

Figure 1.

Volume differences between the healthy subjects and schizophrenia patients. (a) Schizophrenia patients had gray matter volume reduction in the bilateral insulae and medial frontal cortex compared with healthy women according to the analysis of covariance controlling for age (SPM5; healthy subjects > schizophrenia patients). (b) There were statistically significant differences in the cerebrospinal fluid volume in the third ventricle, bilateral Sylvian fissures, and around the prefrontal regions between healthy women and female schizophrenia patients (healthy subjects < schizophrenia patients) according to the analysis of covariance controlling for age (SPM5).

Discriminant analysis

Based on these results, we then performed discriminant analysis. We located the rectangular manual ROI in (i) the parietal and (ii) occipital regions bilaterally for the negative loadings, and in (iii) the insulae, and (iv) medial frontal cortices, and (v) the third ventricle on the normalized individual CSF images (Fig. 2) for the positive loadings on ‘avg152T1.nii’, the standard space, according to the anatomical location. We then applied these ROI to the individual normalized images.

Figure 2.

Regions of interest. (a) Parietal cortices and (b) occipital cortices were selected for the negative loadings. (c) Insulae, (d) medial frontal cortices and (e) the third ventricle were chosen for the positive loadings, reflected in the first analysis (Fig. 1).

The following four variables were entered in a stepwise manner: right insula, medial frontal gyrus, right parietal gyrus and the third ventricle. The coefficients of the discriminant analysis were −8.7, −10.3, 19.7, and 4.9 in the right insula, medial frontal gyrus, right parietal gyrus and the third ventricle, respectively, and the constant was 1.3.

The use of these variables resulted in correct classification rates of 0.76 (sensitivity) in the schizophrenia patients and 0.72 (specificity) in the control subjects in the first sample (F = 8.40; d.f. = 4, 138; P < 0.001; Wilks' lambda = 0.80; Fig. 3a).

Figure 3.

Discriminant scores for (a) 38 schizophrenia patients and 105 healthy comparison subjects, and (b) the second cohort of 23 healthy subjects and 23 patients with schizophrenia, based on linear discriminant function analysis between the original comparison subjects and patients. A demarcation line is given at 0.1. image Schizophrenia patients; image healthy subjects.

When discriminant analysis using the same variables was performed for the second sample, the correct classification rates were 0.74 (sensitivity) in the schizophrenia patients and 0.70 (specificity) in the control subjects (Fig. 3b).


We found that stepwise discriminant function analysis identified the combinations of ROI that characterized brain anatomical features that distinguished female schizophrenia patients from healthy women with good sensitivity and specificity. It seems that the automated method for segmentation and parcellation is as sensitive as the ROI approach to neuroanatomic changes associated with schizophrenia. To our knowledge, this is one of the largest studies in terms of sample size among the discriminant function analysis studies in schizophrenia.

Significant enlargements of the third ventricle, interhemispheric fissure, and the bilateral Sylvian fissure and volume reduction of the medial frontal region, insula and superior temporal region were observed in female patients. These results are consistent with a number of previous studies.[2, 5, 23-25]

Eight ROI were located in the medial frontal region, insulae, and third ventricle region as positive loadings, and in the inferior occipital and superior parietal cortices as negative loadings. Four ROI, including those in the right insula and parietal region, medial frontal region, and third ventricle were regarded as statistically useful as determined by the stepwise method. As for the correct classification rates, the present results showed good sensitivity, although the rates were not as high as those in previous studies.[4-12] This discrepancy may have resulted from the fact that the present participants' age range was relatively wide. Intensive age-related brain shrinkage has been reported in the parietal regions.[15] In contrast, there is little evidence of altered volume in the superior parietal region in schizophrenia.[1-3] We speculated that we could distinguish the age-related brain change from the disease-related change by adding the negative loadings derived from the parietal region. It has also been found that age-related loss of gray matter is more pronounced in the non-dominant hemisphere.[17] The fact that negative loadings derived from the right parietal region had a high coefficient, and the positive loadings from the right insula had a relatively low coefficient may be due to asymmetrical age-related change.

All VBM analysis methods are susceptible to the effects of spatial normalization transformation, by which images from different individuals are registered. Regions in which this spatial transformation has relatively lower accuracy will tend to display artificially higher variability, which will adversely affect statistical significance. For example, the registration and measurement of highly variable prefrontal cortical regions is less accurate than the measurement of less variable structures, such as the precentral gyrus. A recent registration method, known as diffeomorphic anatomical registration using exponentiated Lie algebra (DARTEL) analysis, being more deformable, notably improves the realignment of small inner structures.[26, 27] This new method may solve the problem of the spatial normalization transformation. But this method creates a set of group-specific templates, so it is not applicable to studies involving two different case–control samples. Further work with a refined normalization method is required for the development of better discriminant capability.

Regarding the subcortical regions, an automated approach did not detect the reduced thalamic volume that was previously reported using ROI methods[28] and automated methods.[29-31] Measurements of subcortical structures may require a different tissue classification technique because subcortical structures, especially the thalamus, have MRI signal characteristics that place them between gray matter and white matter. The incorporation of fuzzy classification methods may help to address this problem.

In this study we used different MRI scanners for the first and the second cohorts, and we did not control the differences in quality (e.g. distortion and intensity inhomogeneity) or possible volume inconsistency between the images. We could, however, discriminate between patients and controls in the first and the second cohorts equally using the same discriminant method. This may indicate that correlative pattern of regional brain volumes rather than exact volume or shape matters in volumetry-based discrimination.

The present methods of automated analysis of morphometric data largely support the findings of earlier studies using expert-guided ROI or automated procedures. The discriminant function analysis used in the present study may provide objective biological information adjunct to the clinical diagnosis of schizophrenia.


This study was supported by Health and Labour Sciences Research Grants (Comprehensive Research on Disability, Health, and Welfare, H21-kokoro-001 and H23-seisin-young scientist 013) and Intramural Research Grants for Neurological and Psychiatric Disorders of NCNP (20-3; 21-9). There is no conflict of interest for this research.