Elsevier

NeuroImage

Volume 31, Issue 3, 1 July 2006, Pages 968-980
NeuroImage

An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest

https://doi.org/10.1016/j.neuroimage.2006.01.021Get rights and content

Abstract

In this study, we have assessed the validity and reliability of an automated labeling system that we have developed for subdividing the human cerebral cortex on magnetic resonance images into gyral based regions of interest (ROIs). Using a dataset of 40 MRI scans we manually identified 34 cortical ROIs in each of the individual hemispheres. This information was then encoded in the form of an atlas that was utilized to automatically label ROIs. To examine the validity, as well as the intra- and inter-rater reliability of the automated system, we used both intraclass correlation coefficients (ICC), and a new method known as mean distance maps, to assess the degree of mismatch between the manual and the automated sets of ROIs. When compared with the manual ROIs, the automated ROIs were highly accurate, with an average ICC of 0.835 across all of the ROIs, and a mean distance error of less than 1 mm. Intra- and inter-rater comparisons yielded little to no difference between the sets of ROIs. These findings suggest that the automated method we have developed for subdividing the human cerebral cortex into standard gyral-based neuroanatomical regions is both anatomically valid and reliable. This method may be useful for both morphometric and functional studies of the cerebral cortex as well as for clinical investigations aimed at tracking the evolution of disease-induced changes over time, including clinical trials in which MRI-based measures are used to examine response to treatment.

Introduction

Structural magnetic resonance imaging (MRI) provides extensive detail about the anatomical structure of the brain. It is becoming increasingly important for characterizing cortical changes associated with the normal aging process (see Raz, 2004 for review) and further differentiating these from the degenerative changes associated with dementing illnesses such as Alzheimer's disease (AD). Furthermore, structural MRI has now become an essential tool for the clinical care of patients with brain disease and is of increasing use in clinical trials to identify response to treatment. For example, MRI is now a secondary endpoint in clinical trials of patients with multiple sclerosis (MS), as MS lesions can now be quantified quickly and reliably (see Bakshi et al., 2005). Research and clinical investigations of patients with AD are beginning to incorporate MRI measurements, but these have been primarily restricted to assessments of whole brain atrophy (Freeborough and Fox, 1998a, Freeborough and Fox, 1998b, Fox et al., 1999, Fox et al., 2001, Fox et al., 2005) or manual measures of the hippocampus (Jack et al., 1995, Jack et al., 1997, Jack et al., 2003).

The application of MRI to research and clinical studies has been limited by the ability to quantify the critical dimensions of interest. Methods have been developed to automatically quantify regions of interest (ROI), but these have not as of yet been incorporated into clinical trials. Some of these methods have focused on a single ROI, such as the hippocampus (Hsu et al., 2002, Csernansky et al., 2000, Csernansky et al., 2005, Wang et al., 2003), and the cingulate gyrus (Miller et al., 2003) or sets of subcortical ROIs (Fischl et al., 2002). Developing semi-automated procedures for quantifying cortical ROIs has been more challenging due to the substantial inter-individual variability of the topographic features of the cortex (Zilles et al., 1988, Ono et al., 1990, Kennedy et al., 1998).

Initial efforts at measuring cortical ROIs on MRI scans required substantial operator involvement (e.g., Damasio and Damasio, 1989, Rademacher et al., 1992). More recent and more automated methods employ a variety of approaches to the problem of labeling cortical features, including template-driven warping approaches, where a local correspondence is established between a manually labeled atlas brain and individual subject's brain images (Thompson et al., 1996, Sandor and Leahy, 1997, Hammers et al., 2003, Buckner et al., 2004, Mega et al., 2005), watershed-based approaches to extract cortical sulci (Lohmann, 1998, Rettmann et al., 2002), and graph-based techniques which represent sulci as vertices on a graph (Mangin et al., 1995, Le Goualher et al., 1999).

We recently reported a probabilistic labeling algorithm (Fischl et al., 2004) that was applied to two different systems for defining cortical regions of interest (Rademacher et al., 1992, Destrieux et al., 1998). The strength of this algorithm is that it is not tied to a specific neuroanatomical template, but instead incorporates not only the probable location of a region of interest, but also the potential inter-subject variance of the location of the region, derived from whatever training set is employed.

In the present study, we have expanded on this work in several ways. First, we have developed the definitions of the regions of interest using curvature based information (i.e., sulcal representations) available on images of the cortex that are ‘inflated’ (Dale and Sereno, 1993, Dale et al., 1999, Fischl et al., 1999a, Fischl et al., 1999b, Fischl et al., 2001, Fischl and Dale, 2000); anatomic curvature is visually represented well on inflated images as they provide a view of the brain in which the entire cortical surface is exposed, including the tissue deep in the sulci. Moreover, since the same type of anatomic curvature information is utilized by the probabilistic labeling algorithm, we hypothesized that defining labels on the inflated surface using curvature information would improve the accuracy of the manual definitions of the cortical regions of interest. Second, we employed a training set consisting of 40 MRI scans that included young, middle-aged and elderly controls, as well as patients with Alzheimer's disease (AD). These 40 scans were manually labelled for 34 cortical regions of interest, and an atlas was generated.

With regard to brain atlases, two approaches can be used towards developing atlases. The first approach involves the identification and selection of a highly selected group of individuals (such as cases of AD patients with CDR 1.0) and building an atlas that is optimized for that group alone. Though it is hypothetically possible to construct an atlas from such a selected set of homogenous cases, apply this atlas onto a larger cohort of similar cases and thus achieve high accuracy, it has limited practical value for large morphometric studies aiming to assess anatomic changes across different population types. The second approach, such as the one presented in this manuscript, involves the development of a more generalized atlas that incorporates a wide range of anatomic and atrophic variance. When applied, the result produces an atlas that is likely to be slightly less accurate within a selected group (i.e., when applied only to cases of AD patients with CDR 1.0) but is applicable across several different groups and ultimately, is more accurate across these groups. This more generalized approach to atlas building is important to utilize when the underlying variable is continuous making group distinctions somewhat arbitrary.

We used intraclass correlation coefficients (ICC) to assess validity and reliability. Since we were particularly interested in the anatomical accuracy of the regions of interest, we developed a method that uses the mean distance of “mislabeling” on the cortical surface (known as mean distance maps) to detect the geographical mismatch between manual and automated regions or between sets of regions generated by the same (intra-rater) or different (inter-rater) operators applying the automated process. Finally, since we were interested in assessing the applicability of our automated atlas, we employed a jackknife/leave-one-out technique (a statistical re-sampling method) to test the reliability of our atlas on novel datasets.

Section snippets

Subjects

The participants in the study were enrolled by the Washington University Alzheimer's Disease Research Center (ADRC) in St. Louis. As such, all were screened for neurological impairment, depression and psychoactive medications use (see Fotenos et al., 2005 for details of this sample). As a part of this assessment, all subjects were screened for the presence of major vascular risk factors (e.g., atrial fibrillation, diabetes). Subjects found to have clinically relevant abnormalities on MRI (e.g.,

Manual versus automated analysis (validity)

Table 1 lists the intraclass correlation coefficients for 32 of the 34 labels in each hemisphere. One region, the corpus callosum, was excluded a priori, since it is a white matter structure and, as noted above, was only included in order to better define the regions around it. A second region, the frontal pole, was excluded after examination of the preliminary analyses. This region proved to be unreliable (average ICC 0.26), in all likelihood because it was defined as that region in the most

Discussion

Automated systems for labeling cortical structures provide an efficient way to undertake a complex and otherwise labor-intensive process, provided the accuracy of these methods are sufficient. In the work presented here our motivation was to develop an automated anatomic labeling system from a range of subjects to better account for cortical inter-subject variability that could then be applied to a variety of subjects in studies involving the cerebral cortex. The current findings suggest that

Acknowledgments

This work was supported by National Institute of Aging Grant P01-AG04953, the National Center for Research Resources (P41-RR14075, R01 RR 16594-01A1 and the NCRR BIRN Morphometric Project BIRN002, U24 RR021382), the National Institute for Biomedical Imaging and Bioengineering (R01 EB001550), the Mental Illness and Neuroscience Discovery (MIND) Institute, and Pfizer Incorporated. The authors would like to thank Drs. Thomas Kemper and Douglas Rosene for their insightful discussions concerning the

References (63)

  • G.M. Halliday et al.

    Identifying severely atrophic cortical subregions in Alzheimer's disease

    Neurobiol. Aging

    (2003)
  • C.R.J. Jack et al.

    MRI-based hippocampal volumetrics: data acquisition, normal ranges, and optimal protocol

    Magn. Reson. Imaging

    (1995)
  • M.S. Mega et al.

    Automated brain tissue assessment in the elderly and demented population: construction and validation of a sub-volume probabilistic brain atlas

    NeuroImage

    (2005)
  • M.E. Rettmann et al.

    Automated sulcal segmentation using watersheds on the cortical surface

    NeuroImage

    (2002)
  • F. Ségonne et al.

    A hybrid approach to the skull stripping problem in MRI

    NeuroImage

    (2004)
  • P.M. Thompson et al.

    High-resolution random mesh algorithms for creating a probabilistic 3D surface atlas of the human brain

    NeuroImage

    (1996)
  • N. Tzourio-Mazoyer et al.

    Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain

    NeuroImage

    (2002)
  • D.C. Van Essen

    A Population—Average, Landmark- and Surface-based (PALS) atlas of human cerebral cortex

    NeuroImage

    (2005)
  • L. Wang et al.

    Changes in hippocampal volume and shape across time distinguish dementia of the Alzheimer type from healthy aging

    NeuroImage

    (2003)
  • C.G. Wible et al.

    Parcellation of the human prefrontal cortex using MRI

    Psychiatry Res.

    (1997)
  • H. Yamasue et al.

    Localized volume reduction in prefrontal, temporolimbic, and paralimbic regions in schizophrenia: an MRI parcellation study

    Psychiatry Res.

    (2004)
  • M. Ballmaier et al.

    Localizing gray matter deficits in late-onset depression using computational cortical pattern matching methods

    Am. J. Psychiatry

    (2004)
  • J.G. Csernansky et al.

    Early DAT is distinguished from aging by high-dimensional mapping of the hippocampus. Dementia of the Alzheimer type

    Neurology

    (2000)
  • A.M. Dale et al.

    Improved localization of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: a linear approach

    J. Cogn. Neurosci.

    (1993)
  • H. Damasio et al.

    Lesion Analysis in Neuropsychology

    (1989)
  • C. Destrieux et al.

    Variability of the human brain studied on the flattened cortical surface

    Abstr.-Soc. Neurosci.

    (1998)
  • H. Duvernoy

    The Human Brain

    (1991)
  • B. Efron

    The Jackknife, The Bootstrap and Other Resampling Plans

    (1982)
  • B. Efron et al.

    An Introduction to The Bootstrap

    (1993)
  • B. Fischl et al.

    Measuring the thickness of the human cerebral cortex from magnetic resonance images

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

    (2000)
  • B. Fischl et al.

    High-resolution intersubject averaging and a coordinate system for the cortical surface

    Hum. Brain Mapp.

    (1999)
  • View full text