Biomedical Engineering Reference
In-Depth Information
•
Provides a framework for statistical analysis with well-established
random field models
35
•
Allows the rapid reanalysis of ensembles of data using different
processing criteria
The uncertainties associated with the atlas, the advantages of stereotaxic
space, and the availability at the MNI of a large database of MRI volumes obtai-
ned from young, normal subjects have lead our group to construct a 3D probabi-
listic atlas of gross neuroanatomy. This atlas is described in the next section.
14.2.1.2
MNI Stereotaxic Space
36,37
Rather than
using a single brain, we have established a model from more than 300 MRI data
sets from young normals.
Significant morphometric variability exists between individuals.
38,39
Our model was defined in a coordinate system
similar to that proposed by Talairach;
2
however, we use a single linear trans-
formation instead of 12 piecewise linear transformations to map a brain into
stereotaxic space.
In order to build the stereotaxic space model, we proceeded with a two-
stage procedure. In the first stage, each MRI volume was manually regis-
tered with the stereotaxic coordinate system using the method described by
Evans et al.
36
The line defined by the AC-PC line was manually estimated
by identifying the following points defined on the midsagittal plane: the
occipital pole, the superior aspect of the cerebellum, the inferior aspect of the
splenium of the corpus callosum, the posterior commissure, the inferior
aspect of the thalamus, the anterior commissure, and the inferior aspect of the
genu of the corpus callosum. A line was fitted through these points and the
most anterior and most posterior aspects of the brain identified. A vertical
perpendicular bisector of the AC-PC line was drawn to identify the most
superior aspect of the cortex. Lateral perpendicular bisectors were drawn to
identify the most lateral points of the cerebral volume. These points were
used to define a
linear transformation required to bring each data set
into stereotaxic space. Note that this is a significant difference from the
method originally described by Talairach. Each volume was resampled onto
a 1 mm voxel grid according to this transformation and was subsequently
normalized for mean image intensity. The entire ensemble of MRI data sets
was averaged to create the first-pass mean MRI brain, which was then avail-
able as a target for registration.
This first-pass mean MRI brain was degraded by random errors introduced
in the alignment of each subject's AC-PC line by manual identification. A sec-
ond stage was then initiated, using the first-pass mean average as the target
for an automated 3D image-matching algorithm.
single
28
Each individual brain was
again transformed from its native space to the stereotaxic space by mapping it,
again with a single linear transformation (rigid body plus three scaling para-
meters), to the target volume. This process reduced the effect of random align-
ment errors and increased the contrast of the averaged result. The entire
