Biomedical Engineering Reference
In-Depth Information
Fig. 3.50.
Segmention of MR images. Histogram segmentation (
left
) gives an ap-
proximate segmentation between brain, head, and background, but many details
are still erroneously resolved. Morphological filtering (
right
) corrects them and gives
an accurate segmentation of the head and of the brain
(background, head, and brain). The smoothed histogram is then automati-
cally thresholded into background, head, and brain by searching for points
with a derivative smaller than a given threshold, and corresponding to the
beginning or end of the peaks.
Since the MRI data suffer from various sources of noise, and so that there
is no unambiguous correspondence between gray level and anatomical struc-
tures, a simple segmentation based on gray values does not give adequate
results, so it must be refined by appropriate post-filtering. We make exten-
sive use of two functions of mathematical morphology [64]: erosion (which
corresponds to removing the outer voxels of an object), and the opposite
operation, which is called dilation. The combination of these two operations
with logical operations on binary images is sucient to correctly extract the
head and brain objects from the segmentation results previously obtained.
For this, the whole head object is first filled and separated from the backgro-
und. The brain is then correctly extracted by a series of erosions to remove
spurious brain areas, and dilations to reconstruct the whole brain object.
This segmentation algorithm has been tested on 15 data sets from nor-
mal subjects, and four data sets from patients suffering from brain tumors.
All the data sets from normal patients were correctly segmented in a fully
automated way. For the four other cases, manual adjustments of the para-
meters were necessary to obtain adequate results for the normal part of the
brain. Due to an extremely variable gray level, dimension, and shape, tumors
were generally not satisfactorily extracted and would need some amount of
manual contouring to be accurately segmented. On a standard PC (400 MHz
Pentium II processor), the whole segmentation process took less than 20 s,
thus allowing its use for routine MEG measurements.
The segmentation information was directly used for prior computation.
For 3D visualization, the next step is the computation of the boundaries of
the brain and of the head. This is done with the marching cube algorithm
