Biomedical Engineering Reference
In-Depth Information
directly from this type of image data. These measures are necessary to accurately
assess developmental and/or pathological changes in gross brain structures and
pathways.
In order to test the efficacy of Hyperelastic Warping in the registration of
normal mouse brain anatomy, normal T
1
-weighted micro-MRI images were ob-
tained from two different, intact, excised mouse brains. The image datasets
were 256
3
voxels, FOV
=
1
.
54
×
1
.
54
×
1
.
5 cm, and had 60
µ
m isotropic reso-
lution. A 40
×
40
×
49 rectilinear FE mesh was created for the 3D problems
(73,008 elements). The deforming template was modeled using a neo-Hookean
hyperelastic material with a shear modulus of 450 Pa and a bulk modulus of
400 Pa [22].
These 3-D results demonstrate the efficacy of Hyperelastic Warping when
used on relatively large datasets. Volume-rendered images (Figs. 12.6) show
that excellent external registration was achieved between the deformed tem-
plate and target image a datasets. The 3-D model was rezoned three times to
achieve this registration. It is interesting to note that rezoning allowed a dis-
section artifact in the target image dataset (Fig. 12.6C), that was not present in
the template image data (Fig. 12.6A), to be extruded from the relatively smooth
template to generate the same structure in the deformed template (Fig. 12.6B).
Without the use of rezoning, this excellent alignment would have been impossi-
ble due to extreme mesh distortion resulting in element inversion. Examination
of representative transverse and longitudinal image planes illustrated that very
good internal registration was also achieved, as demonstrated by the correspon-
dence of anatomical regions and sulci between the deformed template and target
(Fig. 12.7. panels A-D).
Computational requirements for this problem were determined primarily by
the size of the finite element mesh used to discretize the template and, to a
lesser extent, by the size of the image datasets. The analysis required 3.38 GB
of memory. Because the main computational expense in the algorithm is the
inversion of a large system of linear equation resulting from the nodal degrees
of freedom in the FE mesh, CPU requirements grew as the square of the size
of the FE mesh. For this analysis, the mesh resulted in a linear system with
165,148 degrees of freedom. Total wall clock time for the 3-D neuroanatomical
registration analysis was 14 hours, with the three mesh rezones accounting
for 18% of the analysis time and the sequential spatial filtering accounting for
5% of the analysis time. The vast majority of the remaining analysis time is
