Elsevier

NeuroImage

Volume 34, Issue 1, 1 January 2007, Pages 204-211
NeuroImage

Connectivity-based parcellation of human cortex using diffusion MRI: Establishing reproducibility, validity and observer independence in BA 44/45 and SMA/pre-SMA

https://doi.org/10.1016/j.neuroimage.2006.08.022Get rights and content

Abstract

The identification of specialized, functional regions of the human cortex is a vital precondition for neuroscience and clinical neurosurgery. Functional imaging modalities are used for their delineation in living subjects, but these methods rely on subject cooperation, and many regions of the human brain cannot be activated specifically.

Diffusion tractography is a novel tool to identify such areas in the human brain, utilizing underlying white matter pathways to separate regions of differing specialization. We explore the reproducibility, generalizability and validity of diffusion tractography-based localization in four functional areas across subjects, timepoints and scanners, and validate findings against fMRI and post-mortem cytoarchitectonic data. With reproducibility across modalities, clustering methods, scanners, timepoints, and subjects in the order of 80–90%, we conclude that diffusion tractography represents a useful and objective tool for parcellation of the human cortex into functional regions, enabling studies into individual functional anatomy even when there are no specific activation paradigms available.

Introduction

The identification of functionally disparate regions of the human cortex in individual subjects is an important aspect of neurosurgery and a prerequisite to many modern methods in neuroscience. To gain insight into functional relationships and networks, techniques such as transcranial magnetic stimulation (TMS) rely on the precise definition of valid targets (Paus et al., 1997). Furthermore, functional regions are taken into account when planning for neurosurgery (Thiel et al., 1998) to minimize post-operative functional deficits and maximize patient benefit in tumor surgery (Proescholdt et al., 2005).

The current gold standard for delineation of specialized regions in the human brain is post-mortem analysis of cyto- or myeloarchitecture (Brodmann, 1909, Economo et al., 1925), which enables parcellation of the human cortex on an objective basis (Schleicher et al., 1999) at microscopic resolution.

For in vivo studies, this kind of data has not been available so far. Consequently, surrogate markers serve to define these regions: for some regions, such as primary sensory and motor cortex, targets can be reliably identified by segmenting anatomical MR images using gross landmarks such as sulci (Geyer et al., 1996, Geyer et al., 1999, Yousry et al., 1997), but this approach cannot be reliably used in all regions of cortex: considerable disagreement between anatomical landmarks and functional borders has been observed in other regions (Amunts et al., 1999).

Traditionally, the in vivo identification of functional regions in the human cortex relies upon functional imaging methods such as functional magnetic resonance imaging (fMRI) or positron emission tomography (PET), where a subject performs a task to activate the region under study while being scanned. Such tasks, however, are not available for many areas of interest. Also, functional mapping experiments rely on the subject’s ability to understand and perform the task correctly. This task compliance can be impaired in the diseased, in children and in elderly subjects, or where language barriers cannot easily be overcome.

Naturally, areas within a functional network need to interchange information, and underlying white matter projections for these networks have been characterized in post-mortem studies (Flechsig, 1876, Flechsig, 1883).

In recent years, diffusion tensor imaging (DTI) has been introduced as a means to measure cerebral white matter fiber distributions in vivo, bridging the gap between functional imaging methods and anatomical ex vivo approaches. Connectivity between regions of the brain can now be measured in vivo with probabilistic diffusion tractography (Behrens et al., 2003b, Parker et al., 2003, Rushworth et al., 2005). Relationships between preferred connectivity to other regions and regional brain function have been demonstrated. Various different approaches were undertaken to analyze the connection patterns measured, creating parcellations of brain regions solely on the basis of their interconnections with other portions of the brain (Behrens et al., 2003a, Behrens and Johansen-Berg, 2005, Johansen-Berg et al., 2004, Johansen-Berg et al., 2005). This new enabling technology offers insight into connectional architecture, but there is only sparse validation of its functional implications and reliability.

In an initial approach, a priori target areas were defined comprising regions of the human cortex that different nuclei in the thalamus were expected to connect to (Behrens et al., 2003a, Johansen-Berg et al., 2005). They served as a model driving segmentation, where each voxel in thalamus was assigned to the target region it connected to most strongly. Then, a semi-automated approach using spectral reordering of connectivity matrices followed to parcellate cortical areas without using a priori information about postulated target regions (Johansen-Berg et al., 2004). Connectivity profiles for every voxel in a seed mask were generated, where connection probabilities to every other voxel in the brain are quantified. Then, these connectivity profiles are reordered such that seed voxels with similar connectivity patterns are positioned together in a cross-correlation matrix. A human observer finally selects cut-off points in the connectivity matrix to separate regions with different profiles. An alternative approach, using k-means clustering to identify regions of distinct connectivity objectively has also been shown to subdivide lateral premotor (Anwander et al., 2005) and inferior frontal areas (Anwander et al., 2006).

Although these approaches offer substantial promise, there are a number of outstanding questions regarding their reproducibility, generalizability, and validity. Aside from the aforementioned pilot studies, there are no data yet on inter-session and inter-subject reliability of DTI-based brain segmentation. Here, we test an automated approach to parcellation of human cortical regions that does not rely on manual selection of cluster borders, removing observer bias from the analysis process.

We use this approach to differentiate between regions previously shown to be characterized by differing connectivity, namely SMA (Supplementary Motor Area) and pre-SMA (Johansen-Berg et al., 2004), and Brodmann’s areas 44 and 45 (Anwander et al., 2006).

SMA is an area that is important in the temporal organization of movements and contains a somatotopic representation of the body, just like the primary motor cortex (Goldberg, 1985, Tanji, 2001), while pre-SMA plays a more abstract role (Picard and Strick, 1996): it is active in the planning and preparation of movement, initiation of movement on cues, acquisition of new motor skills, and higher order aspects of speech, such as self-ordered number generation (Petrides et al., 1993). Brodmann’s areas 44 and 45 are part of Broca’s area on the left side of the brain, a part of the brain that deals with the understanding and generation of speech. They are particularly active in semantic processing (area 45) and integrating sensory information with motor patterns (area 44) (Gough et al., 2005), and play a key role in the implementation of natural language grammar (Friederici, 2004) and phonological processing.

For SMA vs. pre-SMA, where functional localizer tasks exist, we validate results against findings from functional magnetic resonance imaging (fMRI). For BA44/45, where it is more challenging to identify reliable functional localizers (Amunts et al., 1999, Cabeza and Nyberg, 2000), we validate segmentations against population maps based on cytoarchitectonic data as well as a previously published semi-automated approach with manual division of a spectrally reordered connectivity matrix. Using data acquired in the same subjects on 3 different days, and on two different scanners, we quantify the reproducibility of these approaches.

Section snippets

Pre-SMA vs. SMA

A detailed description of data acquisition for pre-SMA vs. SMA is available in Johansen-Berg et al. (2004). In brief, nine healthy volunteers (age 24–35, five male) underwent DTI scanning on a 1.5 T Sonata MR scanner (Siemens, Erlangen, Germany) using the standard quadrature head coil supplied with the system. Diffusion was measured in 60 isotropically distributed directions using echo-planar imaging (SE-EPI, TE 97 ms, TR 10.1 s, 72 axial slices, voxel size 2 mm × 2 mm × 2 mm) using a b-value of 1000 s mm

Validation of medial frontal segmentation

Fig. 2 displays a sagittal slice through a typical result, comparing fMRI and both DTI approaches.

Internal validation, expressed as bilateral conditional probability of concordant classification, demonstrated an agreement between both DTI-based clustering approaches of 92%. External validation of DTI against fMRI results, expressed as likelihood of correct classification with DTI-based approaches given the fMRI result, established an average likelihood of concordant classification of 74–87%

Discussion

Probabilistic diffusion tractography is a novel, non-invasive tool in the analysis of structural and functional brain anatomy, and parcellation of the human brain into functional areas is one of its newest applications (Anwander et al., 2006, Johansen-Berg et al., 2004). However, extending the use of such parcellations requires validation of this approach by demonstrating its reproducibility over time as well as across subjects and scanners. We have now demonstrated feasibility, accuracy,

Acknowledgments

We are grateful for financial support from the UK BBSRC (J.K.), UK MRC (H.J.B., T.B.) and Wellcome Trust (H.J.B.).

References (40)

  • K. Amunts et al.

    Broca’s region revisited: cytoarchitecture and intersubject variability

    Journal of Comparative Neurology

    (1999)
  • A. Anwander et al.

    DTI-Tractography Based Parcellation of Human Precentral Gyrus

    (2005)
  • A. Anwander et al.

    Connectivity-based parcellation of Broca’s area

    Cereb.Cortex

    (2006)
  • T.E.J. Behrens et al.

    Relating connectional architecture to grey matter function using diffusion imaging

    Philosophical Transactions of the Royal Society B-Biological Sciences

    (2005)
  • T.E.J. Behrens et al.

    Non-invasive mapping of connections between human thalamus and cortex using diffusion imaging

    Nature Neuroscience

    (2003)
  • T.E.J. Behrens et al.

    Characterization and propagation of uncertainty in diffusion-weighted MR imaging

    Magnetic Resonance in Medicine

    (2003)
  • J. Bodurka et al.

    Scalable multichannel MRI data acquisition system

    Magnetic Resonance in Medicine

    (2004)
  • K. Brodmann

    Vergleichende Lokalisationslehre der Grosshirnrinde in ihren Prinzipien dargestellt auf Grund des Zellenbaues

    (1909)
  • R. Cabeza et al.

    Imaging cognition II: an empirical review of 275 PET and fMRI studies

    Journal of Cognitive Neuroscience

    (2000)
  • C. Economo et al.

    Die Cytoarchitektonik der Hirnrinde des erwachsenen Menschen

    (1925)
  • Cited by (168)

    • Functional clustering of whole brain white matter fibers

      2020, Journal of Neuroscience Methods
      Citation Excerpt :

      Meanwhile, Hausdorff distance has been applied for fiber clustering as well (Maddah et al., 2008b). Quite a few studies resorted to statistical bundle modelling for representation of fiber features, which included a 9-D feature vector (Brun et al., 2004), B-Splines (Maddah et al., 2005), polynomials (Klein et al., 2007), among others. Recent work is focused on group clustering of large populations to benefit the neuroscience community (Guevara et al., 2012; Ros et al., 2013; Garyfallidis et al., 2017; Zhang et al., 2018; Siless et al., 2018).

    View all citing articles on Scopus
    View full text