Pattern Classification Methods for Analysis and Visualization of Brain Perfusion CT Maps - Computational Intelligence Paradigms in Advanced Pattern Classification - page 151

In order to perform a perfusion image analysis it is necessary to find line that
separate the left brain hemisphere from the right one (in the following text this line
is called a “symmetry axis”). In many CT scans the head of the patient is slightly
rotated on small angle towards the bed of tomograph (the symmetry axis is not or-
thogonal towards to OX axis). The symmetry axis of the image is not always the
same line that separates the brain hemispheres. This makes the problem more com-
plicated. In commercial software the symmetry axis is often selected manually [41].
The algorithm of symmetry axis derivation proposed by the author find the
symmetry axis by computing centers of masses of image horizontal “slices”. The
symmetry axis is approximated as the straight line that minimize the least square
error between all centers of mass coordinates.
After computing the symmetry axis, values of corresponding pixels in the left
and right sides of the image are divided, that generates the asymmetry map. The
asymmetry map is thresholded (in order to find difference of sufficient value). Af-
ter that operation image morphological algorithms are used in order to eliminate
small asymmetries that can not be classified as the lesions. The same algorithm
for detection of regions of potential lesions can be use for both CBF and CBV
maps.
The algorithm works in the following steps:
1. Detect the symmetry axis in the perfusion map using the algorithm described
above.
2. Reduce the number of gray levels in the perfusion map. This operation yields
homogenous areas with sharp borders in both sides of the perfusion map.
3. Perform median filtration. The resulting image lacks some details (the presence
of them would make further analysis of the image harder, because they would
generate a large number of asymmetric regions). If there are too many small
asymmetric regions, the correct classification may turn out to be impossible.
4. Compare the left and the right side (hemisphere) of the image pixel by pixel.
For each pair of pixels, the higher gray level value is divided by the smaller
value (0 by 0 is omitted, x by 0 is replaced with x by 1).
5. Detect a potential asymmetry. Binarize with a bottom threshold T (only pixels
with a value greater than T are left in the resultant image). T is the adaptation
parameter of the algorithm and depends on the perfusion map type (for CBF
and CBV T = 60). The A binary map of asymmetry is generated.
6. The symmetry axis obtained in step 1 does not necessarily separate brain into
symmetric regions. In the asymmetry map A, the border of the brain region may
be detected as a potential asymmetry. In order to eliminate this effect, the bi-
nary mask of “potentially the biggest brain region” P is created. The mask is of
the same size as the A mask and is composed of pixels for which the pixel or its
symmetrical pixel (across to the detected symmetry axis) belongs to brain tis-
sues. All the “holes” in the interior region of the map are removed. The border
of the mask is also removed by morphological opening with a semi-circular
structural element having the diameter of 5.