Elsevier

NeuroImage

Volume 19, Issue 2, June 2003, Pages 233-245
NeuroImage

Regular article
Subcortical, cerebellar, and magnetic resonance based consistent brain image registration

https://doi.org/10.1016/S1053-8119(03)00100-9Get rights and content

Abstract

A new landmark-initialized segmentation and intensity-based (LI-SI) inverse-consistent linear elastic image registration algorithm is presented. This method uses manually identified landmarks, segmented volumetric (anatomical) structures, and normalized image signal intensity information to coregister datasets. The features used for image registration and evaluation include 35 cortical, cerebellar, and commissure landmarks manually identified by experts, subcortical and cerebellar regions defined semi-automatically by an artificial neural network and manually trimmed for validity by experts, and tissue classified images that were generated using a discriminant analysis of three magnetic resonance image sets representing T1, T2, and PD modalities. Four groups of results were computed for coregistering 16 datasets with the following registration techniques: rigid registration, extended Talairach registration, intensity-only inverse-consistent linear elastic registration, and the new LI-SI registration. Results are presented showing that relative overlap measurements increased as the dimensionality of the registration algorithm and amount of anatomical information increased. The average relative overlap improved from 0.53 for the rigid registration to 0.55 for the Talairach registration to 0.74 for the intensity-only and to 0.85 for the LI-SI algorithm. We showed a statistically significant improvement for all but one structure using the intensity-only algorithm compared to the Talairach registration. Furthermore, statistically significant improvements for all structures were achieved using the LI-SI algorithm compared to the intensity-only algorithm.

Introduction

The human brain mapping community has extensively used the generation of an average space for voxel-wise statistical tests to determine differences between groups. An average space is generated to minimize the effect of structural variability that exists within groups of data. This paper compares four methods of image registration to determine their effects on residual anatomic variability by looking at the resulting overlap of subcortical and cerebellar structures.

In functional imaging, voxel-wise comparison has been termed function of interest (FOI) analysis. The advantage of using FOI-based approaches is that the entire brain can be studied. This allows for exploratory studies of brain circuitry that is not well understood (Arndt et al., 1995). A disadvantage of many FOI approaches is that a large spatial filter is often required to smooth the functional imaging data to remove both structural and functional variability. In early positron emission tomography (PET) studies, a filter of 18–20 mm was applied to the data Worsley et al 1992, Andreasen et al 1995, Friston et al 1991, and many recent FOI-based studies also use similar-size spatial filters. The large filter size was justifiable with the early PET scanners that had an inherent spatial resolution of 7–10 mm. With the development of three-dimensional PET cameras and functional magnetic resonance (fMR) imaging, images can be acquired with a spatial resolution on the order of 1–3 mm in plane with a 3- to 7-mm slice thickness. The application of large smoothing filters to data obtained at high resolution will reduce the spatial specificity that exists in the original data White et al 2001, Low and Sorenson 1997, Poline and Mazoyer 1994.

An alternate analysis approach to the FOI method is the region-of-interest (ROI) method that defines a volumetric subregion based either on a high-resolution structural MR image or a template Bohm et al 1983, Fox et al 1985, Friston et al 1989, Evans et al 1991. Measurements of the functional activation within these regions can be compared statistically. This approach is better suited to studies where the specific function of brain regions is well understood for the task being performed Fox 1991, Andreasen et al 1992b. However, if the activated region is small compared to the region of interest, averaging over a large region may wash out differences that exist. The template method, on the other hand, has all of the drawbacks of FOI-based methods due to anatomic variability.

When generating an average space for FOI analysis, the structural variability in the coordinate system should be minimized. After minimization of structural variability, filtering will only be necessary to remove functional variability. This should result in more focal regions defined as areas of significant differences. If all anatomic variability were eliminated, then the FOI method would essentially provide for a very large number of ROIs where each voxel in the brain is considered a separate ROI. Since each voxel has spatial autocorrelation and works in concert with its neighborhood via a highly sophisticated neural network, this limit is too high, but it provides insight into the goal of having a good average space for conducting FOI-based analysis.

One of the first methods for spatial averaging was the Talairach atlas coordinate system Talairach and Tournoux 1988, Talairach et al 1967. Evident in even the earliest study of the Talairach atlas coordinate system, Talairach and Tournoux found that significant anatomic variability remained after spatial registration. In their original publication, substantial variability in the location of the Rolandic fissure was found after registering 20 brains into the Talairach coordinate system. These findings were originally confirmed by Vannier et al. (Vanier et al., 1985) and have since been replicated several times Steinmetz et al 1990, White et al 2001.

Recently, high-dimensional registration algorithms based on features such as landmarks Bookstein 1978, Bookstein 1991, Evans et al 1988, Rohr et al 1999 and lines Davatzikos et al 1996, Thirion 1998a, Subsol 1999, signal intensity of the images Miller et al 1993, Christensen et al 1996, Johnson and Christensen 2001a, Johnson and Christensen 2001b, Ashburner et al 1999, Kochunov et al 1999, Thirion 1998b, Rueckert et al 1999, Woods et al 1988b, or the surface of the brain Pelizzari et al 1989, Thompson and Toga 1996, Fischl et al 1999, Essen et al 1998, Joshi 1997, Davatzikos 1997 have been proposed and developed. Combining landmark and intensity information has been shown Christensen et al 1997, Kybic 2001, Hellier and Barillot 2001 to give better unidirectional image registration results than using either landmark or intensity information alone. Likewise, combining segmented subvolumes with intensity information (Christensen et al., 2001) has been shown to produce better registration results than using intensity information alone. These methods have begun to make their way into the analysis of functional imaging data in order to remove some of the anatomical variability that exists in the data. There have been relatively few systematic studies of the effects of these methods on structure overlaps and determination of the amount of anatomic variability that remains after the application of these methods Kochunov et al 2000, Davatzikos 1997, Dawant et al 1999, Woods et al 1988b.

Even though the number of parameters used to characterize a registration can vary from approximately 10 to over a million, it is unclear whether substructures within the brain actually overlap significantly better after registration compared to before registration. This is especially important in high-dimensional transforms because of their ability to make two images appear very “similar.” It is difficult to quantify what this means in terms of their ability to eliminate structural variability.

This paper presents a new 3D high-dimensional registration algorithm (referred to as LI-SI) that uses a landmark initialization to a linear elastic registration that is driven by both regional anatomic segmentations and image intensity to define consistent mappings between two image sets. The ability of this algorithm to remove structural variability from the resulting registered images was investigated and compared to three other registration methods. The three additional registration methods were rigid registration, Talairach piecewise linear registration, and a high-dimensional, intensity-only, inverse-consistent registration.

Section snippets

Acquisition

This study utilized a sample of subjects comprising 16 males with a mean age of 25.88 years (range 20–41 years, SD 5.70 years). The mean brain volume for the sample was 1543.99 cc (range 1333.17–1792.76, standard deviation 113.99). In the sample, 13 of the subjects were right-handed, 2 were left-handed, and 1 was ambidextrous. The registration template subject was chosen as a right-handed 20-year-old male with a brain volume of 1548.30 cc. Subjects were enrolled voluntarily into an MR imaging

Results

As expected, the ability of the registration algorithms to remove anatomical variability was improved as the dimensionality of the registration algorithm increased and as more information was included in the registration algorithm. The rigid registration is considered a preprocessing step and may be used as a reference for the remaining three registration algorithms.

The registration methods were compared using the relative overlap measure to quantify how well the images were registered with

Summary and conclusions

The results of this study suggest that significant improvements in structure overlap can be obtained by including additional anatomical information. This is due in part to some gray matter structures that abut one another like the nucleus accumbens, putamen, and caudate. A registration algorithm with a limited amount of anatomical information, like signal intensity-only, may not correctly determine the border between these two structures because of overall brain shape differences. Including

Acknowledgements

The authors thank the following people for their work on defining the landmarks and editing the neural network definitions: Ruth Spinks (cortical landmarks), Ronald Pierson (cerebellar landmarks and cerebellar lobes), Karen Albright (thalamus), Jamee Lipcamon (putamen), Barb Malina and Lisa Groff (caudate). Finally, we thank Shirley Harland for proofreading the manuscript. This work was supported in part by the following NIH Grants: NS35368, DC03590, CA75731, MH31593, MH40856, MHCRC43271, and

References (62)

  • R. Pierson et al.

    Measurement of cerebellar subregions on MR images

    NeuroImage

    (2002)
  • D. Shattuck et al.

    Magnetic resonance image tissue classification using a partial volume model

    NeuroImage 13

    (2001)
  • G. Subsol

    Crest lines for curve-based warping

  • J. Thirion

    Image matching as a diffusion processan analogy with maxwell’s demons

    Medical Image Analysis

    (1998)
  • J. Thirion

    Image matching as a diffusion processan analogy with maxwell’s demons

    Medical Image Analysis

    (1998)
  • T. White et al.

    Anatomic and functional variabilitythe effects of filter size in group fMRI data analysis

    NeuroImage

    (2001)
  • N. Andreasen et al.

    Voxel processing techniques for the antemortem study of neuroanatomy and neuropathology using magnetic resonance imaging

    J. Neuropsychiatry Clin.Neurosci.

    (1993)
  • N. Andreasen et al.

    Image processing for the study of brain structure and functionproblems and programs

    J. Neuropsychiatry Clin. Neurosci.

    (1992)
  • N. Andreasen et al.

    Hypofrontality in neuroleptic-naive patients and in patients with chronic schizophrenia. assessment with xenon 133 single-photon emission computed tomography and the tower of London

    Arch. Gen. Psychiatry

    (1992)
  • S. Arndt et al.

    A comparison of approaches to the statistical analysis of [15O]H2O PET cognitive activation studies

    J. Neuropsychiatry

    (1995)
  • C. Bohm et al.

    Adjustable computerized stereotaxic brain atlas for transmission and emission tomography

    Am. J. Neuroradiol.

    (1983)
  • F. Bookstein
    (1978)
  • F. Bookstein

    Morphometric Tools for Landmark Data

    (1991)
  • G. Christensen

    Consistent linear-elastic transformations for image matching

  • G. Christensen et al.

    Consistent image registration

    IEEETrans. Med. Imaging

    (2001)
  • G. Christensen et al.

    Volumetric transformation of brain anatomy

    IEEE Trans. Med. Imaging

    (1997)
  • G. Christensen et al.

    Deformable templates using large deformation kinematics

    IEEE Trans. Image Process.

    (1996)
  • C. Davatzikos et al.

    Image registration based on boundary mapping

    IEEE Trans. Med. Imaging

    (1996)
  • B. Dawant et al.

    Automatic 3-D segmentation of internal structures of the head in MR images using a combination of similarity and free-form transformationsPart i, methodology and validation on normal subjects

    IEEE Trans. Med. Imaging

    (1999)
  • L. Dice

    Measures of the amount of ecologic association between species

    Ecology

    (1945)
  • D.V. Essen et al.

    Functional and structural mapping of human cerebral cortexsolutions are in the surfaces

    Proc. Natl. Acad. Sci. USA

    (1998)
  • Cited by (25)

    • Quantifying deep grey matter atrophy using automated segmentation approaches: A systematic review of structural MRI studies

      2019, NeuroImage
      Citation Excerpt :

      Ensuring consistency between the forward and reverse transformations between atlas and target can be achieved with a symmetric consistent diffeomorphic registration, which was shown to outperform B-spline based free form deformation and ITK Demon’s registration for the segmentation of the DGM (Tao et al., 2009). Subcortical alignment has been further improved in several studies using several approaches, including combining volumetric and surface registration (Postelnicu et al., 2009), performing two-way free-form deformations to extract a regional measure of registration accuracy (Heckemann et al., 2009) and initialising an inverse-consistent registration with landmarks and utilising intensity information from T1, T2 and PD sequences (Magnotta et al., 2003). Unlike landmark-driven correspondences, which provide anatomical relevance but may be irregularly distributed, and intensity-driven correspondences, which provide distributed warps but may not be anatomically feasible, one approach iterates between segmentation of the target from the atlas, anatomical correspondence from these segmentations, and atlas-to-target deformation to determine an optimal correspondence that is robust to brain injury (Meier and Fisher, 2005).

    • Robust automated constellation-based landmark detection in human brain imaging

      2018, NeuroImage
      Citation Excerpt :

      This model sequentially takes advantage of each additional landmark for a better prediction of the next landmark location, and it allows the localization of an arbitrary number of secondary landmarks in different regions of the human brain. Identification of additional landmarks directly benefit registration algorithms that use several manually identified cortical, cerebellar, and commissure landmarks for feature detection (Magnotta et al., 2003). Additional landmarks are also desirable in co-registration approaches.

    • A fast algorithm to estimate inverse consistent image transformation based on corresponding landmarks

      2015, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      Intensity-based algorithms work better for small deformation problem, but lead to mis-registration for large deformation in images. The other one is landmark-based method, which attempts to find the correspondence of landmarks between images [12,24–27]. Commonly used point landmarks are the unique points in intensity map, such as tip-like, saddle-like and sphere-line structure.

    View all citing articles on Scopus
    View full text