Elsevier

NeuroImage

Volume 32, Issue 2, 15 August 2006, Pages 570-582
NeuroImage

Technical Note
Testing anatomically specified hypotheses in functional imaging using cytoarchitectonic maps

https://doi.org/10.1016/j.neuroimage.2006.04.204Get rights and content

Abstract

The statistical inference on functional imaging data is severely complicated by the embedded multiple testing problem. Defining a region of interest (ROI) where the activation is hypothesized a priori helps to circumvent this problem, since in this case the inference is restricted to fewer simultaneous tests, rendering it more sensitive. Cytoarchitectonic maps obtained from postmortem brains provide objective, a priori ROIs that can be used to test anatomically specified hypotheses about the localization of functional activations. We here analyzed three methods for the definition of ROIs based on probabilistic cytoarchitectonic maps. (1) ROIs defined by the volume assigned to a cytoarchitectonic area in the summary map of all areas (maximum probability map, MPM), (2) ROIs based on thresholding the individual probabilistic maps and (3) spherical ROIs build around the cytoarchitectonic center of gravity. The quality with which the thus defined ROIs represented the respective cytoarchitectonic areas as well as their sensitivity for detecting functional activations was subsequently statistically evaluated. Our data showed that the MPM method yields ROIs, which reflect most adequately the underlying anatomical hypotheses. These maps also show a high degree of sensitivity in the statistical analysis. We thus propose the use of MPMs for the definition of ROIs. In combination with thresholding based on the Gaussian random field theory, these ROIs can then be applied to test anatomically specified hypotheses in functional neuroimaging studies.

Introduction

Functional neuroimaging such as positron emission tomography (PET) or functional magnetic resonance imaging (fMRI) is based on changes in cerebral blood flow or metabolism measured in subjects scanned repeatedly under different experimental conditions. In order to identify brain regions which show significant signal differences between those conditions, a univariate model is fitted independently to each voxel of the volume. Voxels where the subsequently computed test statistic exceeds an defined threshold are then classified as active in this particular contrast. This approach, however, embodies a massive multiple testing problem as up to 100,000 test statistics (corresponding to the analyzed voxels) have to be assessed simultaneously. Using an α-level of 0.05 for these tests implicates a 5% chance of falsely rejecting the null hypothesis. Thus, up to 100,000 * 0.05 = 5000 voxels would be declared active, even if there were no signal in the data. The situation is further complicated by the fact that the individual test statistics are highly correlated. Accordingly, Bonferroni's correction for multiple comparisons would be overly conservative since it would overestimate the true number of independent observations.

This problem has been solved by the introduction of Gaussian random field theory (GRF) into neuroimaging. As described in the seminal paper by Worsley et al. (1996), statistical parametric maps can be interpreted as lattice representation of an underlying random field. Thresholds corrected for the family wise error rate (FWE), i.e., the average chance of any false positive activation, can then be derived from these fields. These thresholds are based on Euler's characteristics of the excursion set, e.g., the expected number of “peaks” in the thresholded random field. The necessary threshold to correct for a specified FWE rate depends on the type and smoothness of the random field and on the assessed region of interest (ROI). Evidently, smaller ROIs comprise fewer multiple tests and require a lower threshold. Thus, if the search volume can be confined to a specific brain region where activation is hypothesized, the sensitivity of the analysis increases. This procedure, known as small volume correction (SVC), can be used to reveal more subtle activations. Since the ROIs for small volume correction have to be specified a priori, they are commonly defined as spheres of user-specified radius around the coordinates of previously reported activations for similar tasks. Often, however, the expected localization of activation also is framed as an anatomical hypothesis, e.g., “We expect area 44 to be activated”. However, testing for activation in, e.g., area 44 using a spherical ROI is not an optimal solution, because this ROI will include neighboring cortical areas and white matter as well. Thus, a significant portion of the activation obtained through small volume correction for “area 44” might actually be located in, e.g., area 6.

Incorporating data from anatomical brain mapping studies may circumvent these problems and yield anatomically more valid ROIs. This idea has motivated the generation of ROIs based on the Talairach and Tournoux atlas (Lancaster et al., 2000, Maldjian et al., 2003, Talairach and Tournoux, 1988). However, the obtained results are limited by the major drawbacks of the Talairach atlas, e.g., the tentative transfer of Brodmann's drawing (Brodmann, 1909) to a dissected and photographed postmortem reference brain, the different reference system used by this atlas as compared to functional neuroimaging or the missing information about the inter-individual variability in size and location of cortical areas (Eickhoff et al., 2005). In contrast to the Talairach atlas, probabilistic cytoarchitectonic maps (Amunts and Zilles, 2001, Amunts et al., 2004, Eickhoff et al., 2005, Zilles et al., 2002, Zilles et al., 2003; Fig. 1) provide stereotaxic information on the location and variability of cortical areas. They are based on an observer-independent cytoarchitectonic analysis in a sample of 10 human postmortem brains which were subsequently spatially normalised to the anatomical MNI reference space. This space differs from the original MNI reference space (Collins et al., 1994, Evans et al., 1992, Holmes et al., 1998), which is a widely used reference system in functional neuroimaging, by an affine translation along the y and z axes of 4 and 5 mm, respectively. This shift was introduced in order to relocate the origin of the coordinate system to the anterior commissure of the T1-weighted MNI single subject template (Eickhoff et al., 2005). The advantage of the anatomical MNI space is the correspondence of its origin with that of the atlas system of Talairach and Tournoux (1988), where the x, y and z coordinates indicate the distance in millimeters from the anterior commissure, a clearly defined anatomical landmark. The MNI space and the anatomical MNI space do not differ with respect to any other linear or non-linear transformation. The anatomical MNI space is thus preferable to the “Talairach space” (Talairach and Tournoux, 1988), which shows significant differences to the MNI space not only in the size but also in the shape of the reference brain, rendering a coordinate-based comparison between them difficult or even impossible (Brett et al., 2002, Chau and McIntosh, 2005).

Anatomical ROIs based on probabilistic cytoarchitectonic maps can therefore provide a priori information for the assessment of anatomically specified hypotheses by functional imaging studies. This has been shown in the last few years by a rapidly growing number of studies successfully using probabilistic cytoarchitectonic maps in combination with fMRI and PET data (e.g., Amunts et al., 2004, Bodegard et al., 2000, Eickhoff et al., in press, Grol et al., 2006, Heim et al., 2005, Horwitz et al., 2003, Hurlemann et al., 2005, Kell et al., 2005, Naito et al., 1999, Naito et al., 2005, Noppeney et al., 2005, Young et al., 2004, Larsson et al., 1999, Binkofski et al., 2000). It seems to be intuitive that ROIs based on cytoarchitectonic probabilistic maps of cortical areas are superior to spherical ROIs or macroanatomical landmarks such as gyri and sulci, in the representation of areal specific hypothesis. However, the influence of different methods for the definition of ROIs has not yet been evaluated. Therefore, the aim of this paper was the comparison and evaluation of different procedures for the definition of binary ROIs from probabilistic cytoarchitectonic maps. Procedures were evaluated both with respect to the quality with which the defined ROIs reflect the underlying anatomical hypothesis, and with respect to their sensitivity for detecting functional activations.

Section snippets

Materials and methods

A maximum probability map (MPM, Eickhoff et al., 2005) is a summary map of different probabilistic cytoarchitectonic maps. It is based on the idea of attributing each voxel of the reference space to the most likely cytoarchitectonic area at this position (Fig. 2). MPMs thus allow the definition of non-overlapping representations of all areas from a set of inevitably overlapping probabilistic maps (cf. Fig. 1B). ROIs can be defined from the MPM by a simple binarization: All voxels which are

Exemplary analysis of three cortical areas

The mean probabilities for areas 3b, TE 1.1 and 45 within the anatomically defined ROIs (MPM, 40% map, 50% map) ranged from 52% to 69%. For all three areas, the mean probabilities within the MPM were similar to those within the 40% maps (Table 2). The mean probabilities in the 50% maps, however, were 4 to 9% higher than those within the MPM or the 40% maps. The mean probabilities within the 5 mm spheres (56%, 53% and 65% respectively) were within the range of the values obtained for

Discussion

In this paper, we evaluated different methods for defining ROIs that allow testing anatomical hypothesis in functional neuroimaging based on a priori anatomical information. In particular, three methods for defining anatomical ROIs (MPM, 40% maps and 50% maps) and three spherical ROI definitions (5, 10 and 15 mm radius) were compared with respect to their anatomical specificity and functional sensitivity.

The comprehensive analysis of ROIs for three exemplary areas created by the different

Conclusion and implementation

The use of anatomical ROIs has major advantages over the traditional approaches used to define ROIs for small volume corrections. (I) A definition based on probabilistic anatomical maps is completely independent from the functional data analyzed and thus represents a genuine a priori hypothesis. (II) In contrast to the traditional ROIs (i.e., spheres, boxes), which require interaction and decisions by the investigator, this algorithmic approach is completely objective. This may improve the

Acknowledgments

This Human Brain Project/Neuroinformatics research was funded by the National Institute of Biomedical Imaging and Bioengeneering, the National Institute of Neurological Disorders and Stroke and the National Institute of Mental Health. K.Z. acknowledges funding by the Deutsche Forschungsgemeinschaft (KFO-112) and the Volkswagenstiftung.

References (47)

  • J.A. Maldjian et al.

    An automated method for neuroanatomic and cytoarchitectonic atlas-based interrogation of fMRI data sets

    NeuroImage

    (2003)
  • P. Morosan et al.

    Human primary auditory cortex: cytoarchitectonic subdivisions and mapping into a spatial reference system

    NeuroImage

    (2001)
  • U. Noppeney et al.

    Early visual deprivation induces structural plasticity in gray and white matter

    Curr. Biol.

    (2005)
  • J. Rademacher et al.

    Probabilistic mapping and volume measurement of human primary auditory cortex

    NeuroImage

    (2001)
  • K. Zilles et al.

    Quantitative analysis of cyto- and receptor architecture of the human brain

  • K. Amunts et al.

    Advances in cytoarchitectonic mapping of the human cerebral cortex

    Neuroimaging Clin. N. Am.

    (2001)
  • K. Amunts et al.

    Broca's region revisited: cytoarchitecture and intersubject variability

    J. Comp. Neurol.

    (1999)
  • F. Binkofski et al.

    Broca's region subserves imagery of motion: a combined cytoarchitectonic and fMRI study

    Hum Brain Mapp.

    (2000)
  • A. Bodegard et al.

    Object shape differences reflected by somatosensory cortical activation

    J. Neurosci.

    (2000)
  • M. Brett et al.

    The problem of functional localization in the human brain

    Nat Rev Neurosci

    (2002)
  • K. Brodmann

    Vergleichende Lokalisationslehre der Groβhirnrinde

    (1909)
  • D.L. Collins et al.

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

    J. Comput. Assist. Tomogr.

    (1994)
  • S.B. Eickhoff et al.

    The human parietal operculum: I. Cytoarchitectonic mapping of subdivisions

    Cereb. Cortex

    (2006)
  • Cited by (528)

    • Network nodes in the brain

      2023, Connectome Analysis: Characterization, Methods, and Analysis
    View all citing articles on Scopus
    View full text