Biomedical Engineering Reference
In-Depth Information
in IDL (Interactive Data Language, Research System Inc., USA) and created in
the authors' laboratory for visualizing and analyzing registered image volumes.
We navigate transverse, coronal, and sagittal MR images slice-by-slice to search
the landmarks. The same unique features such as corners and intersections are
identified with a cursor on magnified images. A single person repeats this sev-
eral times over a few weeks, and results are averaged to give a 3D location
for each landmark. A radiologist confirms the landmark selection. Following
registration, we calculate the root-mean-squared (RMS) distance over the six
landmarks [21].
Although this method provides an independent means for evaluating skeletal
registration accuracy, there is error in localizing the bony landmarks. To deter-
mine the effect of localization error, we perform least-squares point-to-point reg-
istration [22] and compare results to MI registration. The rationale is that if we
could identify point landmarks without error on the bony pelvis, point-to-point
registration would be perfect. Hence, any displacement left after registration
is introduced by localization error. We determine the optimal transformation
for matching the six corresponding landmarks. Points are transformed, and dis-
tances between corresponding points are determined. RMS values are computed
and compared to the RMS values from MI registration.
3.2.3.2
Registration Consistency
We use the registration consistency as proposed by [29] for registration eval-
uation. For each of the three volunteers, we use three volumes: reference,
diagnosis, and empty bladder, all of which are obtained with the subject in
the similar position. We call these three volumes A, B, and C, respectively.
They give three pairs of registrations (A-B, B-C, and C-A) and three sets of
transformation parameters (
T
ab
,
T
bc
,
T
ca
). Using the transformation parameters,
we transform voxel positions in A to B, and then to C, and then back to A. The
distance between the original location and the final position is calculated. Since
this is introduced by three transformations, we estimate the error for a single
transformation, by multiplying by 3
−
1
/
2
[29].
3.2.3.3
Voxel Displacements
To test the dependency of registration on algorithmic features such as image
cropping, one can compare transformation parameters. However, we choose a
