Elsevier

NeuroImage

Volume 28, Issue 3, 15 November 2005, Pages 544-552
NeuroImage

Empirical comparison of maximal voxel and non-isotropic adjusted cluster extent results in a voxel-based morphometry study of comorbid learning disability with schizophrenia

https://doi.org/10.1016/j.neuroimage.2005.04.045Get rights and content

Abstract

We present an empirical comparison of cluster extent and maximal voxel results in a voxel-based morphometry (VBM) study of brain structure. The cluster extents are adjusted for underlying deviation from uniform smoothness. We implement this comparison on a four-group cohort that has previously shown evidence of a neuro-developmental component in schizophrenia (Moorhead, T.W.J., Job, D.E., Whalley, H.C., Sanderson, T.L., Johnstone, E.C. and Lawrie, S.M. 2004. Voxel-based morphometry of comorbid schizophrenia and learning disability: analyses in normalized and native spaces using parametric and nonparametric statistical methods. NeuroImage 22: 188–202.). We find that adjusted cluster extent results provide information on the nature of deficits that occur in the schizophrenia affected groups, and these important structural differences are not all shown in maximal voxel results. The maximal voxel and cluster extent results are corrected for multiple comparisons using Random Fields (RF) methods. In order to apply the cluster extent measures, we propose a post-hoc method for determining the primary threshold in the analysis. Unadjusted cluster extent results are reported, for these, no allowance is made for non-isotropic smoothness, and comparison with the adjusted extent results shows that the unadjusted results can be either conservative or anti-conservative depending upon the underlying tissue distributions.

Introduction

Our understanding of brain abnormalities in schizophrenia has been greatly advanced through the use of neuroimaging techniques. Since the first computed tomography study (Johnstone et al., 1976) demonstrated enlarged ventricular size, the ability of neuroimaging to inform researchers in this field has been transformed by the development of magnetic resonance imaging (MRI) with its excellent spatial resolution and, more recently, the ability to combine and contrast large numbers of images through analysis techniques such as voxel-based morphometry (VBM).

Under Statistical Parametric Map (SPM) tests with Random Fields (RF) correction for multiple comparisons, we have the facility to quote maximal voxel, cluster extent and set level corrected P values. The set level results have the most power but minimal localization and are limited to reporting whole brain hypothesis tests. The maximal voxel results provide voxel-wise localization with minimal power for rejection of the null hypothesis. The cluster extent results provide regional rejection of the null hypothesis and reporting power less than that expected for the set level results but greater than that expected for maximal voxel results (Friston et al., 1994, Friston et al., 1995). In structural VBM analysis, it is the accepted practice to limit the reported results to maximal voxel returns, and it is taken that these maximal voxels point to regions where differences exist in the contrasted groups (Job et al., 2002, Keller et al., 2002, Suzuki et al., 2002, Woermann et al., 1999).

The assumption of uniform smoothness in the SPM{t} image, as required for RF multiple comparison correction, cannot be assured in structural VBM analysis. Thus, without adjustment for non-isotropic smoothness, these cluster localization tests are not used (Ashburner and Friston, 2000). It has been shown that the deviation from uniform smoothness in the SPM{t} can be corrected for by evaluating cluster extents in terms of an adjusted cluster size. The adjusted size in resels is evaluated from the cluster coverage in the resel per voxel (RPV) image (Worsley et al., 1999, Cao, 1999, Hayasaka et al., 2004). A comparison of adjusted cluster results RF corrected for multiple comparisons and permutation methods (Nichols and Holmes, 2001) was made by Hayasaka et al. (2004). In this, it was shown that adjusted cluster results are valid under RF correction when the full width half maximum (FWHM) smoothness estimate of the SPM{t} is greater than three times the voxel sampling resolution, and the degrees of freedom (df) exceeds 30. We report on cluster extents from a VBM study of comorbid learning disability and schizophrenia, previously reported on the basis of maximal voxel VBM (Moorhead et al., 2004). In this empirical report, we compare the maximal voxel results with adjusted cluster extent results, and we contrast these adjusted cluster extent results with those given when no correction is made for the underlying non-isotropic smoothness. In this, the FWHM is more than seven times the sampling resolution and df > 42. Thus, the requirements for the use of RF correction of multiple comparisons are met (Hayasaka et al., 2004).

Section snippets

Participants

The MRI scans for this report were obtained from individuals who took part in a clinical comparison of four matched subject groups. The subject groups were comorbid learning disability and schizophrenia, schizophrenia alone, learning disability alone and a normal control group (Moorhead et al., 2004, Sanderson et al., 1999, Sanderson et al., 2001). The patients with comorbid learning disability and schizophrenia meet research diagnostic criteria for schizophrenia (Spitzer et al., 1975) and have

Results

The contrasts between the normal controls and the schizophrenia, comorbid and learning disability groups were examined to assess the specificity that could be obtained in the cluster results for differing primary thresholds. The primary threshold was set to uncorrected t values of 3.3, 3.7, 4.07, 4.45 and 4.79. At each of these settings, the number of significant suprathreshold clusters (unadjusted) was noted and the results tabulated in Table 1. These results show that the maximum specificity

Discussion

Structural VBM implemented under the SPM99 package involves the formation of statistical parametric maps that give a voxel-wise perspective from which rejection of the null hypothesis can be assessed. After corrections for multiple comparisons, localized rejection is made on the basis of maximal voxel significance. Corrections for multiple comparisons in the statistical map are determined in random fields analysis by computing the likely number of resolution elements given the smoothness of

Conclusion

We have implemented an empirical comparison of unadjusted and adjusted cluster extent, tested for null hypothesis rejection and corrected for multiple comparisons by random fields methods in SPM structural analysis. Due to non-isotropic smoothness in the SPM{t}, the validity of unadjusted cluster results is uncertain. Adjusting the cluster size to take account of deviation in the underlying SPM{t} from the image mean allows valid corrected P values to be computed for cluster extents. Our

Acknowledgment

This study was funded by a program grant from the Medical Research Council of Great Britain.

References (29)

Cited by (44)

  • Altered cortical structure and psychiatric symptom risk in adolescents exposed to maternal stress in utero: A retrospective investigation

    2019, Behavioural Brain Research
    Citation Excerpt :

    Clusters were defined using an uncorrected threshold of p < 0.001, k = 10. Corrections for multiple comparisons were made using non-stationary cluster correction (p < 0.05) [87,88], a method which avoids known problems in the use of cluster p-values in VBM data when not accounting for non-stationarity [72,89]. Groups were similar in demographic (SES, race, and sex) and adolescent characteristics (age at MRI, IQ, pubertal development score, BMI, total intracranial volume, total GMV) (Table 2).

View all citing articles on Scopus
View full text