Elsevier

NeuroImage

Volume 59, Issue 2, 16 January 2012, Pages 1369-1381
NeuroImage

Measuring structural–functional correspondence: Spatial variability of specialised brain regions after macro-anatomical alignment

https://doi.org/10.1016/j.neuroimage.2011.08.035Get rights and content

Abstract

The central question of the relationship between structure and function in the human brain is still not well understood. In order to investigate this fundamental relationship we create functional probabilistic maps from a large set of mapping experiments and compare the location of functionally localised regions across subjects using different whole-brain alignment schemes. To avoid the major problems associated with meta-analysis approaches, all subjects are scanned using the same paradigms, the same scanner and the same analysis pipeline. We show that an advanced, curvature driven cortex based alignment (CBA) scheme largely removes macro-anatomical variability across subjects. Remaining variability in the observed spatial location of functional regions, thus, reflects the “true” functional variability, i.e. the quantified variability is a good estimator of the underlying structural–functional correspondence. After localising 13 widely studied functional areas, we found a large variability in the degree to which functional areas respect macro-anatomical boundaries across the cortex. Some areas, such as the frontal eye fields (FEF) are strongly bound to a macro-anatomical location. Fusiform face area (FFA) on the other hand, varies in its location along the length of the fusiform gyrus even though the gyri themselves are well aligned across subjects. Language areas were found to vary greatly across subjects whilst a high degree of overlap was observed in sensory and motor areas. The observed differences in functional variability for different specialised areas suggest that a more complete estimation of the structure–function relationship across the whole cortex requires further empirical studies with an expanded test battery.

Highlights

► Investigation of the structural–functional relationship in the human brain. ► Curvature driven cortex based alignment accurately aligns cortical macro-anatomy. ► Variability across subjects is calculated by functional probabilistic maps. ► Large variability in structural-functional relationship across functional areas.

Introduction

Despite the advances in functional neuroimaging of the last two decades a question which is central to cognitive neuroscience has remained largely unanswered; that is, to what extent do functional areas respect anatomical landmarks, or more generally, what is the relationship between structure and function in the human cerebral cortex? The creation of probabilistic maps based on the data from a large cohort of brains is probably the most powerful way to approach this question. One way to create probabilistic maps from a large sample of subjects is to utilise the already abundant data obtained in mapping experiments in different labs to carry out a meta-analysis. A meta-analytic approach however, would be sub-optimal as it would introduce unwanted sources of variability. Those sources are, amongst others, differences in the reporting of locus of activation due to different alignment and normalisation schemes, variability arising from different experimental paradigms and differences related to inconsistencies in analysis methods and thresholds used across studies (Thirion et al., 2007). In order to avoid these issues, we suggest the creation of data sets from a large cohort from which rich functional information is acquired using the same test battery and a consistent analysis pipeline.

Spatial correspondence between subject's brains is a mandatory condition for meaningful whole brain group analysis as without accurate spatial correspondence, different cortical locations are compared across subjects as if they were the same area. Not accounting for inter-subject macro-anatomical variability is a major source of loss of statistical power in group statistics.

The most widely used normalisation procedures automatically morph an individual subject to fit a template brain. The two most commonly used volumetric normalisation procedures are a) transformation to Talairach space, based on the classic stereotactic atlas created by Talairach and Tournoux from one post-mortem examination (Talairach and Tournoux, 1980) or b) transformation to the MNI template space (Evans et al., 1992, Evans et al., 1993).

Both the Talairach co-ordinate space and MNI template system have been extremely important tools to create spatial correspondences between subjects for group analysis. These normalisation techniques provide a means by which one can compare activated brain regions across studies as results can be reported in terms of a three dimensional co-ordinate system (x, y, and z position in mm). Furthermore the widespread use of standardised co-ordinate frameworks allows for an efficient and practical system of reporting findings to the neuroimaging community and affords the possibility of creating vast databases (e.g. BrainMap.org). However a universal co-ordinate framework still does not exist making the creation of databases for the use of data-mining and meta-analysis difficult (Derrfuss and Mar, 2009).

Despite the many benefits of Talairach and MNI registration these normalisation procedures have severe drawbacks. The major problem being that the same coordinate can refer to different anatomical areas across subjects. Fig. 1 shows two brains warped into Talairach space. One can see that, in the subject on the left, the co-ordinates x =  30 y =  25 z = 54 refer to a point on the anterior bank of the central sulcus whereas, in the second subject on the right, this same co-ordinate refers to an area on the posterior bank of the central sulcus. Even on visual inspection one can see that, despite being normalised to Talairach space, the brains still have very different cortical folding patterns. It is clear that even the most distinct macro-anatomical landmarks in the brain are often not brought into alignment after this normalisation procedure. This in turn has a detrimental effect on group statistics.

The problem of macro-anatomical alignment lies in the fact that the human neocortex is essentially a two dimensional sheet, which is only a few millimetres thick. This cortical surface is highly folded to allow it to squeeze into the restricted cranial cavity and, although there are many consistencies across subjects, each individual has a unique sulcal and gyral folding pattern. The normalisation techniques described above that align the whole brain volume by keeping it as a three dimensional box and perform affine transformations are completely blind to these individual sulcal and gyral folding patterns and thus are only able to perform a ‘global normalisation’ or ‘registration’ of the data to a common space, yet are unable to actually “align” macro-anatomical features. Whilst improvements of intensity-based volumetric alignment schemes have been proposed (Amunts and Zilles, 2001, Ashburner, 2007), approaches that explicitly align the folding pattern of the cortical surface (Fischl et al., 1999b, Goebel et al., 2006) or integrate volume and surface registration (Joshi et al., 2007, Joshi et al., 2009) appear more promising to obtain optimal macro-anatomical alignment.

The advantages of not limiting data analysis to volume space (3 dimensional) coordinate systems but to include surface (2 dimensional) representations have been known for over a decade. Pioneering work from a number of labs demonstrated the visualisation and analytic advantages of cortical surface reconstructions (Dale et al., 1999, DeYoe et al., 1996, Fischl et al., 1999a, Goebel et al., 1998, Sereno et al., 1995, Thompson and Toga, 1996, Van Essen, 2005, Wandell et al., 2000). The creation of flat maps and inflated hemispheres from surface reconstructions of the cortical sheet meant that activity in the entire cortex, including hidden sulcal regions, could be visualised on single, canonical representations reflecting the intrinsic topology of the cortex.

The shift to 2D surface representations of fMRI data also lead to advanced alignment procedures. Early approaches of macro-anatomical surface-based alignment were based on manual selection of landmarks (Van Essen, 2004, Van Essen, 2005, Van Essen et al., 1998). Fischl et al. showed that manual landmarks were not necessary for macro-anatomical alignment, instead, a local curvature-based optimization criterion was used on spherical cortex representations (Fischl et al., 1999b). The curvature-based alignment method demonstrated much improved macro-anatomical correspondence across brains and provided first evidence of improved concomitant alignment of histiologically defined Brodman areas (Fischl et al., 2008). The use of such cortex based alignment (CBA) schemes has continued to grow and has been extended by transforming functional time course data directly in cortex space allowing one to carry out multi-subject general linear model (GLM) analysis and Independent Component Analysis (ICA) directly in cortex space (Formisano et al., 2004, Goebel et al., 2006) and to register brains to existing atlases (Yeo et al., 2010).

An important examination of the effectiveness and accuracy of surface based alignment techniques was carried out by Pantazis et al. (2010). The authors compared the performance of landmark based registration, where a set of manually defined macro-anatomical landmarks act as targets of alignment, and curvature based registration, which automatically extracts curvature information from each individual cortical mesh and uses local curvature deviations to drive alignment.

Landmark based alignment techniques offer more flexibility than automatic curvature-based techniques which is important in some circumstances such as when carrying out lesion studies. Even amongst brains of healthy subjects, large variation in size, shape and orientation of gyri and sulci is difficult to reconcile using automatic curvature driven alignment. Furthermore the added flexibility of landmark based registration is very useful when one considers the fact that there is not always a one-to-one relationship between the number of sulci and gyri across subjects. For example, it is well known that there is variation in the number of transverse temporal gyri across subjects, with some subjects possessing a doubling of this structure (Leonard et al., 1998).

Landmark based alignment schemes also have disadvantages however. Firstly it is very time consuming to manually label landmarks. Secondly, to do so accurately requires a significant amount of anatomical knowledge and lastly, the personal interaction in the alignment introduces the possibility of subjective labelling influencing the resulting registration. In comparison, curvature based alignment is relatively fast; as long as one starts with high quality cortical surface mesh reconstructions an automatic alignment can be completed in a matter of minutes. Another clear benefit is that it does not require detailed anatomical knowledge and that it operates largely automatically. Furthermore curvature based alignment creates reproducible results with respect to a defined objective function leaving no room for experimenter bias and error to influence the alignment.

In order to investigate the relationship between structure and function in the brain, the current project evaluates the degree of spatial correspondence of functionally homologue specialised brain areas across subjects with respect to different anatomical alignment schemes. Spatial correspondence is quantified through the creation of probabilistic maps and assessment of the geodesic distance between homologue areas. Due to the fact that volume based registration techniques are unable to account for individual differences in cortical folding and assuming there is an inherent relationship between brain anatomy and brain function, we expect a higher degree of variation in the location of functional areas when brains are registered in volume space than when the same brains are registered with advanced surface driven techniques. More specifically, if advanced curvature/landmark-based alignment largely removes macro-anatomical variability (tested below) then the calculated probabilistic maps and distance measurements should largely reflect the true variability of the location of functionally defined homologue brain areas.

The common approach to solve the problem of poor overlap of functional regions when carrying out a whole-brain group study is to spatially smooth data, typically with a Gaussian kernel of 8 mm full width at half maximum. However as fMRI imaging techniques advance and data is acquired at an increasingly finer scale, smoothing in effect discards a vast amount of potentially useful data (Kriegeskorte et al., 2006). With an expected increase of spatial correspondence between functional areas across subjects, advanced surface based alignment techniques promise more powerful group statistics without the need to extensively smooth functional data. In a similar way that Fischl investigated the relationship between cortical folding and underlying cytoarchitecture (Fischl et al., 2008) the present study investigates the relationship between cortical folding and the location of functionally specialised brain areas.

In light of the potential issues of curvature-based cortical alignment (Pantazis et al., 2010) we first did a thorough investigation into our cortex based curvature alignment scheme to ensure that we significantly increased correspondence of macro-anatomical landmarks across subjects and that no major sulci were misaligned. Then functional probabilistic maps were calculated in volume (Talairach) space and aligned cortex (surface) space based on 10 subjects using 8 localisation experiments to label 13 functional areas. The obtained intriguing results motivate an extension of the project to create a freely available functional probabilistic atlas covering the whole cortex.

Section snippets

Materials and methods

Functional areas were localised in participants by using standard localising paradigms (see below for details). After defining specific functional regions in each subject we compared the spatial overlap of functional regions across subjects and geodesic distance of peak vertices between subjects, with respect to different normalisation/alignment schemes. All data analysis was carried out using the BrainVoyager QX v2.3 software package (Brain Innovation, Maastricht, The Netherlands).

Macro-anatomical alignment results

CBA was successful in aligning most macro-anatomical landmarks. In the left hemisphere 22 of the 26 sulcal landmarks were brought closer together across subjects after CBA. In the right hemisphere 21 of the landmarks showed less variability across subjects after alignment. Landmarks with many interruptions (e.g. occipito-temporal sulcus) or which are very short (e.g. paracentral sulcus and lateral orbital sulcus) were those least well aligned (see Fig. 3, Fig. 4). Note that all major landmarks

Discussion

By substantially reducing macro-anatomical variability of the cortex through curvature driven surface alignment we have been able to reveal the underlying relationship between structure and function for a number of specialised functional brain regions.

In order to investigate structural–functional correspondence through the creation of probabilistic functional maps, it is crucial to achieve whole brain macro-anatomical correspondence across subjects. Our careful examination of the efficacy of

Conclusions

Since our analysis suggests that there is no general level of structural–functional correspondence, it is important to map more functional brain regions. We will continue to not only add more functional areas to our map but also more subjects. With the resulting extensive database we will create a probabilistic functional atlas across the whole cortex that will be made freely available to the neuroimaging community1

Acknowledgments

We wish to thank Niclas Kilian-Hütten and Aude Collioud for their helpful comments.

References (47)

  • T. Paus

    Location and function of the human frontal eye-field: a selective review

    Neuropsychologia

    (1996)
  • N. Picard et al.

    Imaging the premotor areas

    Curr. Opin. Neurobiol.

    (2001)
  • B. Sorger et al.

    Understanding the functional neuroanatomy of acquired prosopagnosia

    Neuroimage

    (2007)
  • B. Thirion et al.

    Analysis of a large fMRI cohort: statistical and methodological issues for group analyses

    Neuroimage

    (2007)
  • D.C. Van Essen

    Surface-based approaches to spatial localization and registration in primate cerebral cortex

    Neuroimage

    (2004)
  • D.C. Van Essen

    A population-average, landmark- and surface-based (PALS) atlas of human cerebral cortex

    Neuroimage

    (2005)
  • K.S. Weiner et al.

    Sparsely-distributed organization of face and limb activations in human ventral temporal cortex

    Neuroimage

    (2010)
  • K. Amunts et al.

    Advances in cytoarchitectonic mapping of the human cerebral cortex

    Neuroimaging Clin. N. Am.

    (2001)
  • C.J. Bruce et al.

    Primate frontal eye fields. I. Single neurons discharging before saccades

    J. Neurophysiol.

    (1985)
  • T.H. Cormen et al.

    Introduction to Algorithms

    (2001)
  • E.A. DeYoe et al.

    Mapping striate and extrastriate visual areas in human cerebral cortex

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

    (1996)
  • S. Eickhoff et al.

    High-resolution MRI reflects myeloarchitecture and cytoarchitecture of human cerebral cortex

    Hum. Brain Mapp.

    (2005)
  • A.C. Evans et al.

    3D statistical neuroanatomical mofels from 305 MRI volumes

  • Cited by (236)

    View all citing articles on Scopus
    View full text