Information Technology Reference
In-Depth Information
G
might be most time-
demanding operation in the registration algorithm. In order to improve speed of
the algorithm the author used a method proposed in [38]. In [38] the authors used
approximation of Fourier transform that enables to solve the problem with six
multiplication and additions (MADDS) per dimension independent to value
⊗
For large Gaussian kernels computation of
σ
in
Gaussian kernel. The maximum error between approximated and real value is
%
σ
<
0
.
68
.
4.4 Comparison of Registration Algorithms
In order to find best registration model for 2-dimensional CT image matching au-
thor introduce measure based on correlation coefficient:
(
)
(
)
1
−
CC
(
A
,
B
)
−
1
−
CC
(
A
,
B
)
CC
(
A
,
B
)
=
BR
AR
(4.4)
(
)
error
1
−
CC
(
A
,
B
)
BR
CC
BR
(
A
,
B
)
Where
is correlation coefficient of images A and B before
CC
AR
(
A
,
B
)
registration and
is correlation coefficient of images A and B after
registration.
If
CC
(
A
,
B
)
>
0
CC
(
A
,
B
)
>
CC
(
A
,
B
)
it means that
and if
error
AR
BR
CC
(
A
,
B
)
<
0
CC
(
A
,
B
)
<
CC
(
A
,
B
)
it means that
. The registra-
error
AR
BR
CC
(
A
,
B
)
tion algorithm that has highest value of
will be chosen as the al-
error
gorithm for the considered problem.
The author brain atlas constructed for purpose of labeling two dimensional CT
images was consisted of 11 brain templates and 11 different brain tissues labels
(that complexity showed to be sufficient). Labels for those templates were created
by adaptation of Talairach Atlas labels [23] that is popular and broadly used stan-
dard brain atlas template. The comparison of
,(
of three previously
presented algorithm with different parameters (Table 3) was performed on set of
93 CT images. The average
CC
A
B
)
error
CC
(
A
,
B
)
for considered algorithms is
error
presented in Fig 4.
Because
CC
distribution appeared to be normal the minimally important
error
CC
difference (MIR) criterion was used in order to compare
between
error
algorithms (Fig 4).
Search WWH ::
Custom Search