Biomedical Engineering Reference
In-Depth Information
Manual selection of the brain and head parts on the MRI images is preclu-
ded, since it would require several hours of human work for each subject. It
is therefore necessary to develop an automatic algorithm to analyze these
images, with minimum intervention from the user. We present here a method
that we have devised to extract the brain and head volume [61,62] and that
we currently use to set prior knowledge, and to display the solutions.
The initial data are T1-weighted MRI images. Each data set contains
from 122 to 125 axial slices of 256
×
0
.
9375 mm, and out-of-plane resolution is 1.5 mm. Since MRI suffers from
numerous artifacts, a direct extraction of 3D objects from only gray level
information is not possible. Therefore, our algorithm proceeds in two steps.
The first step is the automatic extraction of estimates of the brain and head
objects from gray level information. In the second step, these estimates are
filtered and errors are corrected. For the first step, the gray level histogram is
computed for all of the MRI volume, and is smoothed by using the Silverman
bump-hunting method [63]. For that method, the histogram is considered as
a probability density function and is estimated by
×
256 voxels. In-plane resolution is 0
.
9375
f
x
,
N
1
−
g
(
i, j, k
)
h
p
h
(
x
)=
(3.102)
h
·
N
i,j,k
where
N
is the number of voxels,
f
is a Gaussian function N(0,1),
g
(
i, j, k
)
is the gray level of voxel (
i, j, k
), and
h
is a parameter controlling the degree
of smoothing. A large value of
h
will produce a unimodal histogram, while
smaller values of
h
will give histograms with more and more modes. Our
algorithm begins with a large value of
h
and makes a dichotomic search for
the smallest value of
h
that will produce a distribution with three modes
Fig. 3.49.
An example of a vector map. The tangential component of the field is
projected onto a plane
