Biomedical Engineering Reference
In-Depth Information
The problem is that we do not know what the correct segmentation for
an unsegmented image is. Therefore, we can only hope to find a more or less
successful heuristic for selecting the best atlas for a given image. There are at
least two easily accessible characteristic numbers that describe the similarity
between an image and an atlas. One is the final value of the registration criterion,
or image similarity measure, after either affine or non-rigid registration. The
other is the magnitude of the deformation (i.e., non-rigid transformation) that is
required to map the coordinates of the image onto that of the atlas.
Based on these two concepts, we have compared four different criteria for
selecting the single atlas that is most likely to produce the best segmentation of
a given raw image. These criteria are:
NMI affine: Image similarity after affine registration.
The atlas image
with the highest NMI similarity to the raw image after
affine
registration
is selected and used for its segmentation. This criterion requires only an
affine registration to be computed between the raw image and each of the
atlases. It is therefore considerably less computationally expensive than
the remaining three criteria described below.
NMI non-rigid: Image similarity after non-rigid registration.
The
atlas with the highest NMI value after
non-rigid
registration is selected
and used for segmentation.
DEF avg: Average deformation of the atlas over all voxels.
After
non-rigid registration, the magnitude of the deformation between the raw
image and each individual atlas is computed and averaged over all voxels.
The atlas with the smallest average deformation is selected and used for
segmentation. Whereas the above criteria are based on intensity similarity,
this criterion is based on geometric (i.e., shape) similarity.
DEF max: Maximum deformation of the atlas over all voxels.
This
criterion is identical to the previous one, except that it uses the maximum
deformation over all voxels rather than the average. This criterion pays
more attention to outliers. The idea is that atlases that match well overall
may be substantially inaccurate in some regions.
Segmentations were generated for each of the 20 bee brains, with the remaining
19 brains as possible atlas candidates in each case. For each raw image, one of
