Automatic volume estimation of gross cerebral structures $^\dagger$

DL Collins, NJ Kabani and AC Evans

McConnell Brain Imaging Centre, Montréal Neurological Institute,
McGill University, Montréal, Canada

Abstract:

A new automatic method is used to segment and compute the volumes of gross anatomical structures within the human brain in a collection of 150 MRI volumes of young normal subjects. The method merges the output of a tissue classification process with automatic region identification in order to accurately segment anatomical regions completely automatically.

Introduction

The volume of a number of specific cerebral structures have been correlated with different pathologies such as epilepsy, schizophrenia, Alzheimer's disease, multiple infarct dementia and hydrocephalus. To determine structural abnormality, comparisons to a normative data-base are required. However, estimation of normal population parameters is a daunting task since manual structure segmentation is time-consuming, error-prone and inter- and intra-observer variabilities may confound the estimation of true structure variability. To address these problems, we have developed a fully automatic hybrid segmentation scheme based on [1] (non-linear registration) and [7] (tissue classification) that can be applied to large ensembles of MRI volumes and thus establish the required data base.

Concepts

• Stereotaxic Space: A brain-based coordinate system used to facilitate comparisons among different brains by accounting for position, orientation and scale. At the MNI, a Talairach-like transformation defined by only nine parameters (3 translations, 3 rotations and 3 scales) is used instead of the more complex piece-wise linear transformation described by Talairach [6].

• Classification: Is the term used to assign labels from different tissue classes such as grey-matter, white-matter or cerebro-spinal fluid (CSF) to each voxel within a volume.

• Segmentation: Is the term used to describe the parcelation of an image volume into anatomically meaning structures such as the superior temporal gyrus or a head of the caudate nucleus.

• $\parbox{3in}{ {\bf Average MRI:} The ICBM150 MRI average is a voxel-... ...is region's higher anatomical variability when compared to the basal ganglia.}$ 

• $\parbox{3in}{ {\bf Structure SPAM:} is created by computing a voxel-... ...ic space. The collection of all structure SPAMs forms the {\bf SPAM Atlas}. }$ 

• $\parbox{3in}{ {\bf Max-Proba Atlas:} is based on the SPAM Atlas. At ... ...cture identification for any normal brain transformed into stereotaxic space.}$  $\parbox{3in}{ \begin{center} \leavevmode \centerline{\epsfxsize=2.5i... ...fbox{spam150.eps}} \vspace{0.5em} \centerline{Max-Proba Atlas. } \end{center}}$

Methods

Automatic segmentation is achieved by estimating the non-linear spatial transformation required to register all voxels from a subject's MRI volume with an average MRI brain (Fig. 2) that is co-registered with a SPAM (Statistical Probability Anatomy Map) Atlas (Fig. 2) in a Talairach-like stereotaxic space [4]. The 90 average gross anatomical structures of the Max-Proba Atlas are mapped through the inverse transform to effectively define customized masks on the subject's MRI for the most-likely region for each structure [2]. Tissue classes identified by such as grey-matter, white-matter and CSF are masked by these regions to complete the segmentation. This methodology was applied to the MRI volumes of 152 normal subjects (86 male, 66 female, age $24.6\pm4.8$ years) as part of the ICBM project [5].

Results

In the following figure and table, all volumes are in cm³. Table 1 shows volumes for structures not visible in the surface renderings of Fig. 4. Significant differences (p<0.01, two-tailed Student's t-test with Bonferoni correction) were found for the precentral gyrus (left smaller than right), postcentral gyrus (left larger than right), middle temporal gyrus (left smaller than right), angular gyrus (left smaller than right), inferior occipital gyrus (left larger than right). The left-right volume difference for temporal (left larger than right) and parietal lobes (left smaller than right) are significant at the p=0.05 level.

Structure Volumes.  

$\parbox{5in}{ These images show differen... ...$ '. The isosurface of each SPAM was extracted at the 50\% probability level.}$

Table 1. Structure Volumes.  

	left	right
structure	mean(sd)	mean(sd)
insula	8.8(1.1)	8.8(1.1)
caudate	5.4(0.7)	5.1(0.6)
putamen	5.4(0.7)	5.5(0.7)
globus pallidus	0.9(0.2)	1.4(0.2)
thalamus	8.3(0.9)	9.0(1.0)
nucleus accumbens	0.3(0.1)	0.4(0.1)
subthalamic nucleus	0.1(0.0)	0.1(0.0)
anterior internal capsule	3.3(0.5)	3.0(0.6)
posterior internal capsule	1.6(0.3)	1.6(0.3)
lateral ventricle	8.9(3.6)	8.0(3.4)
corpus collosum	10.9(1.5)
fornix	0.3(0.1)

Table 2. Lobe Volumes.  

	left	right
lobe	mean(sd)	mean(sd)
frontal	175.0(25.3)	174.0(25.0)
temporal	119.3(18.1)	109.8(16.4)
parietal	95.0(13.7)	99.3(14.3)
occipital	44.8(7.9)	45.9(8.2)

$\parbox{5in}{ Lobe volumes in cc$^3$\space (mean$\pm$ sd). Only tem... ...left smaller than right) lobe volume differences are significant ($p<0.05$ ).}$

Conclusion

The method presented here is completely automatic, fully objective, and has been applied to a large ensemble of brain volumes. The resulting volume statistics will prove useful as a normative data base for comparisons in future studies of normal or pathological brains.

Bibliography

1

D. Collins, C. Holmes, T. Peters, and A. Evans.
Automatic 3D model-based neuroanatomical segmentation.
Human Brain Mapping, 3(3):190-208, 1995.

2

D. Collins, A. Zijdenbos, and A. Evans.
Improved automatic gross cerebral structure segmentation.
In A. Evans, editor, 4th International Conference on Functional Mapping of the Human Brain, Montreal, June 1998. Organization for Human Brain Mapping.
abstract no. 707.

3

D. L. Collins, P. Neelin, T. M. Peters, and A. C. Evans.
Automatic 3D inter-subject registration of MR volumetric data in standardized talairach space.
Journal of Computer Assisted Tomography, 18(2):192-205, March/April 1994.

4

A. Evans, D. Collins, and C. Holmes.
Computational approaches to quantifying human neuroanatomical variability.
In J. Mazziotta and A. Toga, editors, Brain Mapping: The Methods, pages 343-361. Academic Press, 1996.

5

J. Mazziotta, A. Toga, A. Evans, P. Fox, and J. Lancaster.
A probabilistic atlas of the human brain: theory and rationale for its development. the international consortium for brain mapping.
NeuroImage, 2(2):89-101, 1995.

6

J. Talairach, G. Szikla, and P. Tournoux.
Atlas d'anatomie stereotaxique du telencephale.
Masson, Paris, 1967.

7

A. P. Zijdenbos, A. C. Evans, F. Riahi, J. Sled, J. Chui, and V. Kollokian.
Automatic quantification of multiple sclerosis lesion volume using stereotaxic space.
In Proceedings of the 4th International Conference on Visualization in Biomedical Computing, VBC `96:, pages 439-448, Hamburg, September 1996.

Acknowledgements

The authors would like to express their appreciation for support from the Human Frontier Science Project Organization, the Canadian Medical Research Council (SP-30), the McDonnell-Pew Cognitive Neuroscience Center Program, the U.S. Human Brain Map Project (HBMP), NIMH and NIDA. This work forms part of a continuing project of the HBMP-funded International Consortium for Brain Mapping (ICBM) to develop a probabilistic atlas of human neuroanatomy. The authors also wish to thank W.F.C. Baaré for manual segmentation examples.